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_send_data *sd, struct lnet_peer *peer,
1436                            __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(sd->sd_best_ni, sd->sd_dst_nid,
1454                                    peer, peer_net);
1455 }
1456
1457 static int
1458 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1459                     struct lnet_peer_ni **best_lpni)
1460 {
1461         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1462         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1463         struct lnet_peer *lp1 = r1->lr_gateway;
1464         struct lnet_peer *lp2 = r2->lr_gateway;
1465         struct lnet_peer_ni *lpni1;
1466         struct lnet_peer_ni *lpni2;
1467         struct lnet_send_data sd;
1468         int rc;
1469
1470         sd.sd_best_ni = NULL;
1471         sd.sd_dst_nid = LNET_NID_ANY;
1472         lpni1 = lnet_find_best_lpni_on_net(&sd, lp1, r1->lr_lnet);
1473         lpni2 = lnet_find_best_lpni_on_net(&sd, lp2, r2->lr_lnet);
1474         LASSERT(lpni1 && lpni2);
1475
1476         if (r1->lr_priority < r2->lr_priority) {
1477                 *best_lpni = lpni1;
1478                 return 1;
1479         }
1480
1481         if (r1->lr_priority > r2->lr_priority) {
1482                 *best_lpni = lpni2;
1483                 return -1;
1484         }
1485
1486         if (r1_hops < r2_hops) {
1487                 *best_lpni = lpni1;
1488                 return 1;
1489         }
1490
1491         if (r1_hops > r2_hops) {
1492                 *best_lpni = lpni2;
1493                 return -1;
1494         }
1495
1496         rc = lnet_compare_peers(lpni1, lpni2);
1497         if (rc == 1) {
1498                 *best_lpni = lpni1;
1499                 return rc;
1500         } else if (rc == -1) {
1501                 *best_lpni = lpni2;
1502                 return rc;
1503         }
1504
1505         if (r1->lr_seq - r2->lr_seq <= 0) {
1506                 *best_lpni = lpni1;
1507                 return 1;
1508         }
1509
1510         *best_lpni = lpni2;
1511         return -1;
1512 }
1513
1514 static struct lnet_route *
1515 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1516                        struct lnet_route **prev_route,
1517                        struct lnet_peer_ni **gwni)
1518 {
1519         struct lnet_peer_ni *best_gw_ni = NULL;
1520         struct lnet_route *best_route;
1521         struct lnet_route *last_route;
1522         struct lnet_remotenet *rnet;
1523         struct lnet_peer *lp_best;
1524         struct lnet_route *route;
1525         struct lnet_peer *lp;
1526         int rc;
1527
1528         rnet = lnet_find_rnet_locked(remote_net);
1529         if (rnet == NULL)
1530                 return NULL;
1531
1532         lp_best = NULL;
1533         best_route = last_route = NULL;
1534         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1535                 lp = route->lr_gateway;
1536
1537                 if (!lnet_is_route_alive(route))
1538                         continue;
1539
1540                 if (lp_best == NULL) {
1541                         best_route = last_route = route;
1542                         lp_best = lp;
1543                 }
1544
1545                 /* no protection on below fields, but it's harmless */
1546                 if (last_route->lr_seq - route->lr_seq < 0)
1547                         last_route = route;
1548
1549                 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1550                 if (rc < 0)
1551                         continue;
1552
1553                 best_route = route;
1554                 lp_best = lp;
1555         }
1556
1557         *prev_route = last_route;
1558         *gwni = best_gw_ni;
1559
1560         return best_route;
1561 }
1562
1563 static struct lnet_ni *
1564 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1565                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1566                  int md_cpt)
1567 {
1568         struct lnet_ni *ni = NULL;
1569         unsigned int shortest_distance;
1570         int best_credits;
1571         int best_healthv;
1572
1573         /*
1574          * If there is no peer_ni that we can send to on this network,
1575          * then there is no point in looking for a new best_ni here.
1576         */
1577         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1578                 return best_ni;
1579
1580         if (best_ni == NULL) {
1581                 shortest_distance = UINT_MAX;
1582                 best_credits = INT_MIN;
1583                 best_healthv = 0;
1584         } else {
1585                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1586                                                      best_ni->ni_dev_cpt);
1587                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1588                 best_healthv = atomic_read(&best_ni->ni_healthv);
1589         }
1590
1591         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1592                 unsigned int distance;
1593                 int ni_credits;
1594                 int ni_healthv;
1595                 int ni_fatal;
1596
1597                 ni_credits = atomic_read(&ni->ni_tx_credits);
1598                 ni_healthv = atomic_read(&ni->ni_healthv);
1599                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1600
1601                 /*
1602                  * calculate the distance from the CPT on which
1603                  * the message memory is allocated to the CPT of
1604                  * the NI's physical device
1605                  */
1606                 distance = cfs_cpt_distance(lnet_cpt_table(),
1607                                             md_cpt,
1608                                             ni->ni_dev_cpt);
1609
1610                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1611                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1612                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1613                         : "not seleced", best_credits, shortest_distance,
1614                         (best_ni) ? best_ni->ni_seq : 0);
1615
1616                 /*
1617                  * All distances smaller than the NUMA range
1618                  * are treated equally.
1619                  */
1620                 if (distance < lnet_numa_range)
1621                         distance = lnet_numa_range;
1622
1623                 /*
1624                  * Select on health, shorter distance, available
1625                  * credits, then round-robin.
1626                  */
1627                 if (ni_fatal) {
1628                         continue;
1629                 } else if (ni_healthv < best_healthv) {
1630                         continue;
1631                 } else if (ni_healthv > best_healthv) {
1632                         best_healthv = ni_healthv;
1633                         /*
1634                          * If we're going to prefer this ni because it's
1635                          * the healthiest, then we should set the
1636                          * shortest_distance in the algorithm in case
1637                          * there are multiple NIs with the same health but
1638                          * different distances.
1639                          */
1640                         if (distance < shortest_distance)
1641                                 shortest_distance = distance;
1642                 } else if (distance > shortest_distance) {
1643                         continue;
1644                 } else if (distance < shortest_distance) {
1645                         shortest_distance = distance;
1646                 } else if (ni_credits < best_credits) {
1647                         continue;
1648                 } else if (ni_credits == best_credits) {
1649                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1650                                 continue;
1651                 }
1652                 best_ni = ni;
1653                 best_credits = ni_credits;
1654         }
1655
1656         CDEBUG(D_NET, "selected best_ni %s\n",
1657                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1658
1659         return best_ni;
1660 }
1661
1662 /*
1663  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1664  * because such traffic is required to perform discovery. We therefore
1665  * exclude all GET and PUT on that portal. We also exclude all ACK and
1666  * REPLY traffic, but that is because the portal is not tracked in the
1667  * message structure for these message types. We could restrict this
1668  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1669  */
1670 static bool
1671 lnet_msg_discovery(struct lnet_msg *msg)
1672 {
1673         if (msg->msg_type == LNET_MSG_PUT) {
1674                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1675                         return true;
1676         } else if (msg->msg_type == LNET_MSG_GET) {
1677                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1678                         return true;
1679         }
1680         return false;
1681 }
1682
1683 #define SRC_SPEC        0x0001
1684 #define SRC_ANY         0x0002
1685 #define LOCAL_DST       0x0004
1686 #define REMOTE_DST      0x0008
1687 #define MR_DST          0x0010
1688 #define NMR_DST         0x0020
1689 #define SND_RESP        0x0040
1690
1691 /* The following to defines are used for return codes */
1692 #define REPEAT_SEND     0x1000
1693 #define PASS_THROUGH    0x2000
1694
1695 /* The different cases lnet_select pathway needs to handle */
1696 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1697 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1698 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1699 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1700 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1701 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1702 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1703 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1704
1705 static int
1706 lnet_handle_lo_send(struct lnet_send_data *sd)
1707 {
1708         struct lnet_msg *msg = sd->sd_msg;
1709         int cpt = sd->sd_cpt;
1710
1711         /* No send credit hassles with LOLND */
1712         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1713         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1714         if (!msg->msg_routing)
1715                 msg->msg_hdr.src_nid =
1716                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1717         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1718         lnet_msg_commit(msg, cpt);
1719         msg->msg_txni = the_lnet.ln_loni;
1720
1721         return LNET_CREDIT_OK;
1722 }
1723
1724 static int
1725 lnet_handle_send(struct lnet_send_data *sd)
1726 {
1727         struct lnet_ni *best_ni = sd->sd_best_ni;
1728         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1729         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1730         struct lnet_msg *msg = sd->sd_msg;
1731         int cpt2;
1732         __u32 send_case = sd->sd_send_case;
1733         int rc;
1734         __u32 routing = send_case & REMOTE_DST;
1735          struct lnet_rsp_tracker *rspt;
1736
1737         /*
1738          * Increment sequence number of the selected peer so that we
1739          * pick the next one in Round Robin.
1740          */
1741         best_lpni->lpni_seq++;
1742
1743         /*
1744          * grab a reference on the peer_ni so it sticks around even if
1745          * we need to drop and relock the lnet_net_lock below.
1746          */
1747         lnet_peer_ni_addref_locked(best_lpni);
1748
1749         /*
1750          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1751          * message. This ensures that we get a CPT that is correct for
1752          * the NI when the NI has been restricted to a subset of all CPTs.
1753          * If the selected CPT differs from the one currently locked, we
1754          * must unlock and relock the lnet_net_lock(), and then check whether
1755          * the configuration has changed. We don't have a hold on the best_ni
1756          * yet, and it may have vanished.
1757          */
1758         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1759         if (sd->sd_cpt != cpt2) {
1760                 __u32 seq = lnet_get_dlc_seq_locked();
1761                 lnet_net_unlock(sd->sd_cpt);
1762                 sd->sd_cpt = cpt2;
1763                 lnet_net_lock(sd->sd_cpt);
1764                 if (seq != lnet_get_dlc_seq_locked()) {
1765                         lnet_peer_ni_decref_locked(best_lpni);
1766                         return REPEAT_SEND;
1767                 }
1768         }
1769
1770         /*
1771          * store the best_lpni in the message right away to avoid having
1772          * to do the same operation under different conditions
1773          */
1774         msg->msg_txpeer = best_lpni;
1775         msg->msg_txni = best_ni;
1776
1777         /*
1778          * grab a reference for the best_ni since now it's in use in this
1779          * send. The reference will be dropped in lnet_finalize()
1780          */
1781         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1782
1783         /*
1784          * Always set the target.nid to the best peer picked. Either the
1785          * NID will be one of the peer NIDs selected, or the same NID as
1786          * what was originally set in the target or it will be the NID of
1787          * a router if this message should be routed
1788          */
1789         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1790
1791         /*
1792          * lnet_msg_commit assigns the correct cpt to the message, which
1793          * is used to decrement the correct refcount on the ni when it's
1794          * time to return the credits
1795          */
1796         lnet_msg_commit(msg, sd->sd_cpt);
1797
1798         /*
1799          * If we are routing the message then we keep the src_nid that was
1800          * set by the originator. If we are not routing then we are the
1801          * originator and set it here.
1802          */
1803         if (!msg->msg_routing)
1804                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1805
1806         if (routing) {
1807                 msg->msg_target_is_router = 1;
1808                 msg->msg_target.pid = LNET_PID_LUSTRE;
1809                 /*
1810                  * since we're routing we want to ensure that the
1811                  * msg_hdr.dest_nid is set to the final destination. When
1812                  * the router receives this message it knows how to route
1813                  * it.
1814                  *
1815                  * final_dst_lpni is set at the beginning of the
1816                  * lnet_select_pathway() function and is never changed.
1817                  * It's safe to use it here.
1818                  */
1819                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1820         } else {
1821                 /*
1822                  * if we're not routing set the dest_nid to the best peer
1823                  * ni NID that we picked earlier in the algorithm.
1824                  */
1825                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1826         }
1827
1828         /*
1829          * if we have response tracker block update it with the next hop
1830          * nid
1831          */
1832         if (msg->msg_md) {
1833                 rspt = msg->msg_md->md_rspt_ptr;
1834                 if (rspt) {
1835                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1836                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1837                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1838                 }
1839         }
1840
1841         rc = lnet_post_send_locked(msg, 0);
1842
1843         if (!rc)
1844                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1845                        libcfs_nid2str(msg->msg_hdr.src_nid),
1846                        libcfs_nid2str(msg->msg_txni->ni_nid),
1847                        libcfs_nid2str(sd->sd_src_nid),
1848                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1849                        libcfs_nid2str(sd->sd_dst_nid),
1850                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1851                        libcfs_nid2str(sd->sd_rtr_nid),
1852                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1853
1854         return rc;
1855 }
1856
1857 static inline void
1858 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1859 {
1860         if (sd->sd_send_case & NMR_DST &&
1861             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1862             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1863             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1864                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1865                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1866                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1867                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1868                                                  sd->sd_best_ni->ni_nid);
1869         }
1870 }
1871
1872 /*
1873  * Source Specified
1874  * Local Destination
1875  * non-mr peer
1876  *
1877  * use the source and destination NIDs as the pathway
1878  */
1879 static int
1880 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1881 {
1882         /* the destination lpni is set before we get here. */
1883
1884         /* find local NI */
1885         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1886         if (!sd->sd_best_ni) {
1887                 CERROR("Can't send to %s: src %s is not a "
1888                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1889                                 libcfs_nid2str(sd->sd_src_nid));
1890                 return -EINVAL;
1891         }
1892
1893         /*
1894          * the preferred NID will only be set for NMR peers
1895          */
1896         lnet_set_non_mr_pref_nid(sd);
1897
1898         return lnet_handle_send(sd);
1899 }
1900
1901 /*
1902  * Source Specified
1903  * Local Destination
1904  * MR Peer
1905  *
1906  * Don't run the selection algorithm on the peer NIs. By specifying the
1907  * local NID, we're also saying that we should always use the destination NID
1908  * provided. This handles the case where we should be using the same
1909  * destination NID for the all the messages which belong to the same RPC
1910  * request.
1911  */
1912 static int
1913 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1914 {
1915         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1916         if (!sd->sd_best_ni) {
1917                 CERROR("Can't send to %s: src %s is not a "
1918                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1919                                 libcfs_nid2str(sd->sd_src_nid));
1920                 return -EINVAL;
1921         }
1922
1923         if (sd->sd_best_lpni &&
1924             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1925                 return lnet_handle_lo_send(sd);
1926         else if (sd->sd_best_lpni)
1927                 return lnet_handle_send(sd);
1928
1929         CERROR("can't send to %s. no NI on %s\n",
1930                libcfs_nid2str(sd->sd_dst_nid),
1931                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1932
1933         return -EHOSTUNREACH;
1934 }
1935
1936 struct lnet_ni *
1937 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1938                               struct lnet_peer *peer,
1939                               struct lnet_peer_net *peer_net,
1940                               int cpt,
1941                               bool incr_seq)
1942 {
1943         struct lnet_net *local_net;
1944         struct lnet_ni *best_ni;
1945
1946         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1947         if (!local_net)
1948                 return NULL;
1949
1950         /*
1951          * Iterate through the NIs in this local Net and select
1952          * the NI to send from. The selection is determined by
1953          * these 3 criterion in the following priority:
1954          *      1. NUMA
1955          *      2. NI available credits
1956          *      3. Round Robin
1957          */
1958         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1959                                    peer, peer_net, cpt);
1960
1961         if (incr_seq && best_ni)
1962                 best_ni->ni_seq++;
1963
1964         return best_ni;
1965 }
1966
1967 static int
1968 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1969                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1970                              int cpt)
1971 {
1972         struct lnet_peer *peer;
1973         lnet_nid_t primary_nid;
1974         int rc;
1975
1976         lnet_peer_ni_addref_locked(lpni);
1977
1978         peer = lpni->lpni_peer_net->lpn_peer;
1979
1980         if (lnet_peer_gw_discovery(peer)) {
1981                 lnet_peer_ni_decref_locked(lpni);
1982                 return 0;
1983         }
1984
1985         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1986                 lnet_peer_ni_decref_locked(lpni);
1987                 return 0;
1988         }
1989
1990         rc = lnet_discover_peer_locked(lpni, cpt, false);
1991         if (rc) {
1992                 lnet_peer_ni_decref_locked(lpni);
1993                 return rc;
1994         }
1995         /* The peer may have changed. */
1996         peer = lpni->lpni_peer_net->lpn_peer;
1997         spin_lock(&peer->lp_lock);
1998         if (lnet_peer_is_uptodate_locked(peer)) {
1999                 spin_unlock(&peer->lp_lock);
2000                 lnet_peer_ni_decref_locked(lpni);
2001                 return 0;
2002         }
2003         /* queue message and return */
2004         msg->msg_rtr_nid_param = rtr_nid;
2005         msg->msg_sending = 0;
2006         msg->msg_txpeer = NULL;
2007         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2008         primary_nid = peer->lp_primary_nid;
2009         spin_unlock(&peer->lp_lock);
2010
2011         lnet_peer_ni_decref_locked(lpni);
2012
2013         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2014                 msg, libcfs_nid2str(primary_nid));
2015
2016         return LNET_DC_WAIT;
2017 }
2018
2019 static int
2020 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2021                              lnet_nid_t dst_nid,
2022                              struct lnet_peer_ni **gw_lpni,
2023                              struct lnet_peer **gw_peer)
2024 {
2025         int rc;
2026         __u32 local_lnet;
2027         struct lnet_peer *gw;
2028         struct lnet_peer *lp;
2029         struct lnet_peer_net *lpn;
2030         struct lnet_peer_net *best_lpn = NULL;
2031         struct lnet_remotenet *rnet;
2032         struct lnet_route *best_route = NULL;
2033         struct lnet_route *last_route = NULL;
2034         struct lnet_peer_ni *lpni = NULL;
2035         struct lnet_peer_ni *gwni = NULL;
2036         lnet_nid_t src_nid = sd->sd_src_nid;
2037
2038         /* If a router nid was specified then we are replying to a GET or
2039          * sending an ACK. In this case we use the gateway associated with the
2040          * specified router nid.
2041          */
2042         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2043                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2044                 if (!gwni) {
2045                         CERROR("No peer NI for gateway %s\n",
2046                                libcfs_nid2str(sd->sd_rtr_nid));
2047                         return -EHOSTUNREACH;
2048                 }
2049                 gw = gwni->lpni_peer_net->lpn_peer;
2050                 lnet_peer_ni_decref_locked(gwni);
2051                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2052         } else {
2053                 /* we've already looked up the initial lpni using dst_nid */
2054                 lpni = sd->sd_best_lpni;
2055                 /* the peer tree must be in existence */
2056                 LASSERT(lpni && lpni->lpni_peer_net &&
2057                         lpni->lpni_peer_net->lpn_peer);
2058                 lp = lpni->lpni_peer_net->lpn_peer;
2059
2060                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2061                         /* is this remote network reachable?  */
2062                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2063                         if (!rnet)
2064                                 continue;
2065
2066                         if (!best_lpn)
2067                                 best_lpn = lpn;
2068
2069                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2070                                 continue;
2071
2072                         best_lpn = lpn;
2073                 }
2074
2075                 if (!best_lpn) {
2076                         CERROR("peer %s has no available nets\n",
2077                                libcfs_nid2str(sd->sd_dst_nid));
2078                         return -EHOSTUNREACH;
2079                 }
2080
2081                 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd, lp, best_lpn->lpn_net_id);
2082                 if (!sd->sd_best_lpni) {
2083                         CERROR("peer %s down\n",
2084                                libcfs_nid2str(sd->sd_dst_nid));
2085                         return -EHOSTUNREACH;
2086                 }
2087
2088                 best_route = lnet_find_route_locked(NULL, best_lpn->lpn_net_id,
2089                                                     &last_route, &gwni);
2090                 if (!best_route) {
2091                         CERROR("no route to %s from %s\n",
2092                                libcfs_nid2str(dst_nid),
2093                                libcfs_nid2str(src_nid));
2094                         return -EHOSTUNREACH;
2095                 }
2096
2097                 if (!gwni) {
2098                         CERROR("Internal Error. Route expected to %s from %s\n",
2099                                libcfs_nid2str(dst_nid),
2100                                libcfs_nid2str(src_nid));
2101                         return -EFAULT;
2102                 }
2103
2104                 gw = best_route->lr_gateway;
2105                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2106                 local_lnet = best_route->lr_lnet;
2107
2108         }
2109
2110         /*
2111          * Discover this gateway if it hasn't already been discovered.
2112          * This means we might delay the message until discovery has
2113          * completed
2114          */
2115         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2116         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2117                                           sd->sd_cpt);
2118         if (rc)
2119                 return rc;
2120
2121         if (!sd->sd_best_ni)
2122                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2123                                         lnet_peer_get_net_locked(gw,
2124                                                                  local_lnet),
2125                                         sd->sd_md_cpt,
2126                                         true);
2127
2128         if (!sd->sd_best_ni) {
2129                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2130                        libcfs_net2str(local_lnet),
2131                        libcfs_nid2str(sd->sd_src_nid));
2132                 return -EFAULT;
2133         }
2134
2135         *gw_lpni = gwni;
2136         *gw_peer = gw;
2137
2138         /*
2139          * increment the sequence numbers since now we're sure we're
2140          * going to use this path
2141          */
2142         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2143                 LASSERT(best_route && last_route);
2144                 best_route->lr_seq = last_route->lr_seq + 1;
2145                 best_lpn->lpn_seq++;
2146         }
2147
2148         return 0;
2149 }
2150
2151 /*
2152  * Handle two cases:
2153  *
2154  * Case 1:
2155  *  Source specified
2156  *  Remote destination
2157  *  Non-MR destination
2158  *
2159  * Case 2:
2160  *  Source specified
2161  *  Remote destination
2162  *  MR destination
2163  *
2164  * The handling of these two cases is similar. Even though the destination
2165  * can be MR or non-MR, we'll deal directly with the router.
2166  */
2167 static int
2168 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2169 {
2170         int rc;
2171         struct lnet_peer_ni *gw_lpni = NULL;
2172         struct lnet_peer *gw_peer = NULL;
2173
2174         /* find local NI */
2175         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2176         if (!sd->sd_best_ni) {
2177                 CERROR("Can't send to %s: src %s is not a "
2178                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2179                                 libcfs_nid2str(sd->sd_src_nid));
2180                 return -EINVAL;
2181         }
2182
2183         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2184                                      &gw_peer);
2185         if (rc)
2186                 return rc;
2187
2188         if (sd->sd_send_case & NMR_DST)
2189                 /*
2190                 * since the final destination is non-MR let's set its preferred
2191                 * NID before we send
2192                 */
2193                 lnet_set_non_mr_pref_nid(sd);
2194
2195         /*
2196          * We're going to send to the gw found so let's set its
2197          * info
2198          */
2199         sd->sd_peer = gw_peer;
2200         sd->sd_best_lpni = gw_lpni;
2201
2202         return lnet_handle_send(sd);
2203 }
2204
2205 struct lnet_ni *
2206 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2207                                bool discovery)
2208 {
2209         struct lnet_peer_net *peer_net = NULL;
2210         struct lnet_ni *best_ni = NULL;
2211
2212         /*
2213          * The peer can have multiple interfaces, some of them can be on
2214          * the local network and others on a routed network. We should
2215          * prefer the local network. However if the local network is not
2216          * available then we need to try the routed network
2217          */
2218
2219         /* go through all the peer nets and find the best_ni */
2220         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2221                 /*
2222                  * The peer's list of nets can contain non-local nets. We
2223                  * want to only examine the local ones.
2224                  */
2225                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2226                         continue;
2227                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2228                                                    peer_net, md_cpt, false);
2229
2230                 /*
2231                  * if this is a discovery message and lp_disc_net_id is
2232                  * specified then use that net to send the discovery on.
2233                  */
2234                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2235                     discovery)
2236                         break;
2237         }
2238
2239         if (best_ni)
2240                 /* increment sequence number so we can round robin */
2241                 best_ni->ni_seq++;
2242
2243         return best_ni;
2244 }
2245
2246 static struct lnet_ni *
2247 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2248 {
2249         struct lnet_ni *best_ni = NULL;
2250         struct lnet_peer_net *peer_net;
2251         struct lnet_peer *peer = sd->sd_peer;
2252         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2253         struct lnet_peer_ni *lpni;
2254         int cpt = sd->sd_cpt;
2255
2256         /*
2257          * We must use a consistent source address when sending to a
2258          * non-MR peer. However, a non-MR peer can have multiple NIDs
2259          * on multiple networks, and we may even need to talk to this
2260          * peer on multiple networks -- certain types of
2261          * load-balancing configuration do this.
2262          *
2263          * So we need to pick the NI the peer prefers for this
2264          * particular network.
2265          */
2266
2267         /* Get the target peer_ni */
2268         peer_net = lnet_peer_get_net_locked(peer,
2269                         LNET_NIDNET(best_lpni->lpni_nid));
2270         LASSERT(peer_net != NULL);
2271         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2272                                 lpni_peer_nis) {
2273                 if (lpni->lpni_pref_nnids == 0)
2274                         continue;
2275                 LASSERT(lpni->lpni_pref_nnids == 1);
2276                 best_ni = lnet_nid2ni_locked(
2277                                 lpni->lpni_pref.nid, cpt);
2278                 break;
2279         }
2280
2281         return best_ni;
2282 }
2283
2284 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2285 static int
2286 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2287 {
2288         struct lnet_ni *best_ni = NULL;
2289         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2290
2291         /*
2292          * We must use a consistent source address when sending to a
2293          * non-MR peer. However, a non-MR peer can have multiple NIDs
2294          * on multiple networks, and we may even need to talk to this
2295          * peer on multiple networks -- certain types of
2296          * load-balancing configuration do this.
2297          *
2298          * So we need to pick the NI the peer prefers for this
2299          * particular network.
2300          */
2301
2302         best_ni = lnet_find_existing_preferred_best_ni(sd);
2303
2304         /* if best_ni is still not set just pick one */
2305         if (!best_ni) {
2306                 best_ni =
2307                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2308                                                 sd->sd_best_lpni->lpni_peer_net,
2309                                                 sd->sd_md_cpt, true);
2310                 /* If there is no best_ni we don't have a route */
2311                 if (!best_ni) {
2312                         CERROR("no path to %s from net %s\n",
2313                                 libcfs_nid2str(best_lpni->lpni_nid),
2314                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2315                         return -EHOSTUNREACH;
2316                 }
2317         }
2318
2319         sd->sd_best_ni = best_ni;
2320
2321         /* Set preferred NI if necessary. */
2322         lnet_set_non_mr_pref_nid(sd);
2323
2324         return 0;
2325 }
2326
2327
2328 /*
2329  * Source not specified
2330  * Local destination
2331  * Non-MR Peer
2332  *
2333  * always use the same source NID for NMR peers
2334  * If we've talked to that peer before then we already have a preferred
2335  * source NI associated with it. Otherwise, we select a preferred local NI
2336  * and store it in the peer
2337  */
2338 static int
2339 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2340 {
2341         int rc;
2342
2343         /* sd->sd_best_lpni is already set to the final destination */
2344
2345         /*
2346          * At this point we should've created the peer ni and peer. If we
2347          * can't find it, then something went wrong. Instead of assert
2348          * output a relevant message and fail the send
2349          */
2350         if (!sd->sd_best_lpni) {
2351                 CERROR("Internal fault. Unable to send msg %s to %s. "
2352                        "NID not known\n",
2353                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2354                        libcfs_nid2str(sd->sd_dst_nid));
2355                 return -EFAULT;
2356         }
2357
2358         rc = lnet_select_preferred_best_ni(sd);
2359         if (!rc)
2360                 rc = lnet_handle_send(sd);
2361
2362         return rc;
2363 }
2364
2365 static int
2366 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2367 {
2368         /*
2369          * NOTE we've already handled the remote peer case. So we only
2370          * need to worry about the local case here.
2371          *
2372          * if we're sending a response, ACK or reply, we need to send it
2373          * to the destination NID given to us. At this point we already
2374          * have the peer_ni we're suppose to send to, so just find the
2375          * best_ni on the peer net and use that. Since we're sending to an
2376          * MR peer then we can just run the selection algorithm on our
2377          * local NIs and pick the best one.
2378          */
2379         if (sd->sd_send_case & SND_RESP) {
2380                 sd->sd_best_ni =
2381                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2382                                                 sd->sd_best_lpni->lpni_peer_net,
2383                                                 sd->sd_md_cpt, true);
2384
2385                 if (!sd->sd_best_ni) {
2386                         /*
2387                          * We're not going to deal with not able to send
2388                          * a response to the provided final destination
2389                          */
2390                         CERROR("Can't send response to %s. "
2391                                "No local NI available\n",
2392                                 libcfs_nid2str(sd->sd_dst_nid));
2393                         return -EHOSTUNREACH;
2394                 }
2395
2396                 return lnet_handle_send(sd);
2397         }
2398
2399         /*
2400          * If we get here that means we're sending a fresh request, PUT or
2401          * GET, so we need to run our standard selection algorithm.
2402          * First find the best local interface that's on any of the peer's
2403          * networks.
2404          */
2405         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2406                                         sd->sd_md_cpt,
2407                                         lnet_msg_discovery(sd->sd_msg));
2408         if (sd->sd_best_ni) {
2409                 sd->sd_best_lpni =
2410                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2411                                              sd->sd_best_ni->ni_net->net_id);
2412
2413                 /*
2414                  * if we're successful in selecting a peer_ni on the local
2415                  * network, then send to it. Otherwise fall through and
2416                  * try and see if we can reach it over another routed
2417                  * network
2418                  */
2419                 if (sd->sd_best_lpni &&
2420                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2421                         /*
2422                          * in case we initially started with a routed
2423                          * destination, let's reset to local
2424                          */
2425                         sd->sd_send_case &= ~REMOTE_DST;
2426                         sd->sd_send_case |= LOCAL_DST;
2427                         return lnet_handle_lo_send(sd);
2428                 } else if (sd->sd_best_lpni) {
2429                         /*
2430                          * in case we initially started with a routed
2431                          * destination, let's reset to local
2432                          */
2433                         sd->sd_send_case &= ~REMOTE_DST;
2434                         sd->sd_send_case |= LOCAL_DST;
2435                         return lnet_handle_send(sd);
2436                 }
2437
2438                 CERROR("Internal Error. Expected to have a best_lpni: "
2439                        "%s -> %s\n",
2440                        libcfs_nid2str(sd->sd_src_nid),
2441                        libcfs_nid2str(sd->sd_dst_nid));
2442
2443                 return -EFAULT;
2444         }
2445
2446         /*
2447          * Peer doesn't have a local network. Let's see if there is
2448          * a remote network we can reach it on.
2449          */
2450         return PASS_THROUGH;
2451 }
2452
2453 /*
2454  * Case 1:
2455  *      Source NID not specified
2456  *      Local destination
2457  *      MR peer
2458  *
2459  * Case 2:
2460  *      Source NID not speified
2461  *      Remote destination
2462  *      MR peer
2463  *
2464  * In both of these cases if we're sending a response, ACK or REPLY, then
2465  * we need to send to the destination NID provided.
2466  *
2467  * In the remote case let's deal with MR routers.
2468  *
2469  */
2470
2471 static int
2472 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2473 {
2474         int rc = 0;
2475         struct lnet_peer *gw_peer = NULL;
2476         struct lnet_peer_ni *gw_lpni = NULL;
2477
2478         /*
2479          * handle sending a response to a remote peer here so we don't
2480          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2481          */
2482         if (sd->sd_send_case & REMOTE_DST &&
2483             sd->sd_send_case & SND_RESP) {
2484                 struct lnet_peer_ni *gw;
2485                 struct lnet_peer *gw_peer;
2486
2487                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2488                                                   &gw_peer);
2489                 if (rc < 0) {
2490                         CERROR("Can't send response to %s. "
2491                                "No route available\n",
2492                                 libcfs_nid2str(sd->sd_dst_nid));
2493                         return -EHOSTUNREACH;
2494                 } else if (rc > 0) {
2495                         return rc;
2496                 }
2497
2498                 sd->sd_best_lpni = gw;
2499                 sd->sd_peer = gw_peer;
2500
2501                 return lnet_handle_send(sd);
2502         }
2503
2504         /*
2505          * Even though the NID for the peer might not be on a local network,
2506          * since the peer is MR there could be other interfaces on the
2507          * local network. In that case we'd still like to prefer the local
2508          * network over the routed network. If we're unable to do that
2509          * then we select the best router among the different routed networks,
2510          * and if the router is MR then we can deal with it as such.
2511          */
2512         rc = lnet_handle_any_mr_dsta(sd);
2513         if (rc != PASS_THROUGH)
2514                 return rc;
2515
2516         /*
2517          * Now that we must route to the destination, we must consider the
2518          * MR case, where the destination has multiple interfaces, some of
2519          * which we can route to and others we do not. For this reason we
2520          * need to select the destination which we can route to and if
2521          * there are multiple, we need to round robin.
2522          */
2523         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2524                                           &gw_peer);
2525         if (rc)
2526                 return rc;
2527
2528         sd->sd_send_case &= ~LOCAL_DST;
2529         sd->sd_send_case |= REMOTE_DST;
2530
2531         sd->sd_peer = gw_peer;
2532         sd->sd_best_lpni = gw_lpni;
2533
2534         return lnet_handle_send(sd);
2535 }
2536
2537 /*
2538  * Source not specified
2539  * Remote destination
2540  * Non-MR peer
2541  *
2542  * Must send to the specified peer NID using the same source NID that
2543  * we've used before. If it's the first time to talk to that peer then
2544  * find the source NI and assign it as preferred to that peer
2545  */
2546 static int
2547 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2548 {
2549         int rc;
2550         struct lnet_peer_ni *gw_lpni = NULL;
2551         struct lnet_peer *gw_peer = NULL;
2552
2553         /*
2554          * Let's set if we have a preferred NI to talk to this NMR peer
2555          */
2556         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2557
2558         /*
2559          * find the router and that'll find the best NI if we didn't find
2560          * it already.
2561          */
2562         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2563                                           &gw_peer);
2564         if (rc)
2565                 return rc;
2566
2567         /*
2568          * set the best_ni we've chosen as the preferred one for
2569          * this peer
2570          */
2571         lnet_set_non_mr_pref_nid(sd);
2572
2573         /* we'll be sending to the gw */
2574         sd->sd_best_lpni = gw_lpni;
2575         sd->sd_peer = gw_peer;
2576
2577         return lnet_handle_send(sd);
2578 }
2579
2580 static int
2581 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2582 {
2583         /*
2584          * turn off the SND_RESP bit.
2585          * It will be checked in the case handling
2586          */
2587         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2588
2589         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2590                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2591                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2592                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2593                 libcfs_nid2str(sd->sd_dst_nid),
2594                 (send_case & LOCAL_DST) ? "local" : "routed");
2595
2596         switch (send_case) {
2597         /*
2598          * For all cases where the source is specified, we should always
2599          * use the destination NID, whether it's an MR destination or not,
2600          * since we're continuing a series of related messages for the
2601          * same RPC
2602          */
2603         case SRC_SPEC_LOCAL_NMR_DST:
2604                 return lnet_handle_spec_local_nmr_dst(sd);
2605         case SRC_SPEC_LOCAL_MR_DST:
2606                 return lnet_handle_spec_local_mr_dst(sd);
2607         case SRC_SPEC_ROUTER_NMR_DST:
2608         case SRC_SPEC_ROUTER_MR_DST:
2609                 return lnet_handle_spec_router_dst(sd);
2610         case SRC_ANY_LOCAL_NMR_DST:
2611                 return lnet_handle_any_local_nmr_dst(sd);
2612         case SRC_ANY_LOCAL_MR_DST:
2613         case SRC_ANY_ROUTER_MR_DST:
2614                 return lnet_handle_any_mr_dst(sd);
2615         case SRC_ANY_ROUTER_NMR_DST:
2616                 return lnet_handle_any_router_nmr_dst(sd);
2617         default:
2618                 CERROR("Unknown send case\n");
2619                 return -1;
2620         }
2621 }
2622
2623 static int
2624 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2625                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2626 {
2627         struct lnet_peer_ni     *lpni;
2628         struct lnet_peer        *peer;
2629         struct lnet_send_data   send_data;
2630         int                     cpt, rc;
2631         int                     md_cpt;
2632         __u32                   send_case = 0;
2633
2634         memset(&send_data, 0, sizeof(send_data));
2635
2636         /*
2637          * get an initial CPT to use for locking. The idea here is not to
2638          * serialize the calls to select_pathway, so that as many
2639          * operations can run concurrently as possible. To do that we use
2640          * the CPT where this call is being executed. Later on when we
2641          * determine the CPT to use in lnet_message_commit, we switch the
2642          * lock and check if there was any configuration change.  If none,
2643          * then we proceed, if there is, then we restart the operation.
2644          */
2645         cpt = lnet_net_lock_current();
2646
2647         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2648         if (md_cpt == CFS_CPT_ANY)
2649                 md_cpt = cpt;
2650
2651 again:
2652
2653         /*
2654          * If we're sending to ourselves then there is no need to go through
2655          * any selection. We can shortcut the entire process and send over
2656          * lolnd.
2657          *
2658          * However, we make two exceptions to this rule:
2659          * 1. If the src_nid is specified then our API defines that we must send
2660          *    via that interface.
2661          * 2. Recovery messages must be sent to the lnet_ni that is being
2662          *    recovered.
2663          */
2664         send_data.sd_msg = msg;
2665         send_data.sd_cpt = cpt;
2666         if (src_nid == LNET_NID_ANY && !msg->msg_recovery &&
2667             lnet_nid2ni_locked(dst_nid, cpt)) {
2668                 rc = lnet_handle_lo_send(&send_data);
2669                 lnet_net_unlock(cpt);
2670                 return rc;
2671         }
2672
2673         /*
2674          * find an existing peer_ni, or create one and mark it as having been
2675          * created due to network traffic. This call will create the
2676          * peer->peer_net->peer_ni tree.
2677          */
2678         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2679         if (IS_ERR(lpni)) {
2680                 lnet_net_unlock(cpt);
2681                 return PTR_ERR(lpni);
2682         }
2683
2684         /*
2685          * Cache the original src_nid. If we need to resend the message
2686          * then we'll need to know whether the src_nid was originally
2687          * specified for this message. If it was originally specified,
2688          * then we need to keep using the same src_nid since it's
2689          * continuing the same sequence of messages.
2690          */
2691         msg->msg_src_nid_param = src_nid;
2692
2693         /*
2694          * If necessary, perform discovery on the peer that owns this peer_ni.
2695          * Note, this can result in the ownership of this peer_ni changing
2696          * to another peer object.
2697          */
2698         rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2699         if (rc) {
2700                 lnet_peer_ni_decref_locked(lpni);
2701                 lnet_net_unlock(cpt);
2702                 return rc;
2703         }
2704         lnet_peer_ni_decref_locked(lpni);
2705
2706         peer = lpni->lpni_peer_net->lpn_peer;
2707
2708         /*
2709          * Identify the different send cases
2710          */
2711         if (src_nid == LNET_NID_ANY)
2712                 send_case |= SRC_ANY;
2713         else
2714                 send_case |= SRC_SPEC;
2715
2716         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2717                 send_case |= LOCAL_DST;
2718         else
2719                 send_case |= REMOTE_DST;
2720
2721         /*
2722          * if this is a non-MR peer or if we're recovering a peer ni then
2723          * let's consider this an NMR case so we can hit the destination
2724          * NID.
2725          */
2726         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2727                 send_case |= NMR_DST;
2728         else
2729                 send_case |= MR_DST;
2730
2731         if (msg->msg_type == LNET_MSG_REPLY ||
2732             msg->msg_type == LNET_MSG_ACK)
2733                 send_case |= SND_RESP;
2734
2735         /* assign parameters to the send_data */
2736         send_data.sd_rtr_nid = rtr_nid;
2737         send_data.sd_src_nid = src_nid;
2738         send_data.sd_dst_nid = dst_nid;
2739         send_data.sd_best_lpni = lpni;
2740         /*
2741          * keep a pointer to the final destination in case we're going to
2742          * route, so we'll need to access it later
2743          */
2744         send_data.sd_final_dst_lpni = lpni;
2745         send_data.sd_peer = peer;
2746         send_data.sd_md_cpt = md_cpt;
2747         send_data.sd_send_case = send_case;
2748
2749         rc = lnet_handle_send_case_locked(&send_data);
2750
2751         /*
2752          * Update the local cpt since send_data.sd_cpt might've been
2753          * updated as a result of calling lnet_handle_send_case_locked().
2754          */
2755         cpt = send_data.sd_cpt;
2756
2757         if (rc == REPEAT_SEND)
2758                 goto again;
2759
2760         lnet_net_unlock(cpt);
2761
2762         return rc;
2763 }
2764
2765 int
2766 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2767 {
2768         lnet_nid_t              dst_nid = msg->msg_target.nid;
2769         int                     rc;
2770
2771         /*
2772          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2773          * but we might want to use pre-determined router for ACK/REPLY
2774          * in the future
2775          */
2776         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2777         LASSERT(msg->msg_txpeer == NULL);
2778         LASSERT(msg->msg_txni == NULL);
2779         LASSERT(!msg->msg_sending);
2780         LASSERT(!msg->msg_target_is_router);
2781         LASSERT(!msg->msg_receiving);
2782
2783         msg->msg_sending = 1;
2784
2785         LASSERT(!msg->msg_tx_committed);
2786
2787         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2788         if (rc < 0) {
2789                 if (rc == -EHOSTUNREACH)
2790                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2791                 else
2792                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2793                 return rc;
2794         }
2795
2796         if (rc == LNET_CREDIT_OK)
2797                 lnet_ni_send(msg->msg_txni, msg);
2798
2799         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2800         return 0;
2801 }
2802
2803 enum lnet_mt_event_type {
2804         MT_TYPE_LOCAL_NI = 0,
2805         MT_TYPE_PEER_NI
2806 };
2807
2808 struct lnet_mt_event_info {
2809         enum lnet_mt_event_type mt_type;
2810         lnet_nid_t mt_nid;
2811 };
2812
2813 /* called with res_lock held */
2814 void
2815 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2816 {
2817         struct lnet_rsp_tracker *rspt;
2818
2819         /*
2820          * msg has a refcount on the MD so the MD is not going away.
2821          * The rspt queue for the cpt is protected by
2822          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2823          */
2824         if (!md->md_rspt_ptr)
2825                 return;
2826
2827         rspt = md->md_rspt_ptr;
2828
2829         /* debug code */
2830         LASSERT(rspt->rspt_cpt == cpt);
2831
2832         md->md_rspt_ptr = NULL;
2833
2834         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2835                 /*
2836                  * The monitor thread has invalidated this handle because the
2837                  * response timed out, but it failed to lookup the MD. That
2838                  * means this response tracker is on the zombie list. We can
2839                  * safely remove it under the resource lock (held by caller) and
2840                  * free the response tracker block.
2841                  */
2842                 list_del(&rspt->rspt_on_list);
2843                 lnet_rspt_free(rspt, cpt);
2844         } else {
2845                 /*
2846                  * invalidate the handle to indicate that a response has been
2847                  * received, which will then lead the monitor thread to clean up
2848                  * the rspt block.
2849                  */
2850                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2851         }
2852 }
2853
2854 void
2855 lnet_clean_zombie_rstqs(void)
2856 {
2857         struct lnet_rsp_tracker *rspt, *tmp;
2858         int i;
2859
2860         cfs_cpt_for_each(i, lnet_cpt_table()) {
2861                 list_for_each_entry_safe(rspt, tmp,
2862                                          the_lnet.ln_mt_zombie_rstqs[i],
2863                                          rspt_on_list) {
2864                         list_del(&rspt->rspt_on_list);
2865                         lnet_rspt_free(rspt, i);
2866                 }
2867         }
2868
2869         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2870 }
2871
2872 static void
2873 lnet_finalize_expired_responses(void)
2874 {
2875         struct lnet_libmd *md;
2876         struct list_head local_queue;
2877         struct lnet_rsp_tracker *rspt, *tmp;
2878         ktime_t now;
2879         int i;
2880
2881         if (the_lnet.ln_mt_rstq == NULL)
2882                 return;
2883
2884         cfs_cpt_for_each(i, lnet_cpt_table()) {
2885                 INIT_LIST_HEAD(&local_queue);
2886
2887                 lnet_net_lock(i);
2888                 if (!the_lnet.ln_mt_rstq[i]) {
2889                         lnet_net_unlock(i);
2890                         continue;
2891                 }
2892                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2893                 lnet_net_unlock(i);
2894
2895                 now = ktime_get();
2896
2897                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2898                         /*
2899                          * The rspt mdh will be invalidated when a response
2900                          * is received or whenever we want to discard the
2901                          * block the monitor thread will walk the queue
2902                          * and clean up any rsts with an invalid mdh.
2903                          * The monitor thread will walk the queue until
2904                          * the first unexpired rspt block. This means that
2905                          * some rspt blocks which received their
2906                          * corresponding responses will linger in the
2907                          * queue until they are cleaned up eventually.
2908                          */
2909                         lnet_res_lock(i);
2910                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2911                                 lnet_res_unlock(i);
2912                                 list_del(&rspt->rspt_on_list);
2913                                 lnet_rspt_free(rspt, i);
2914                                 continue;
2915                         }
2916
2917                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2918                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2919                                 struct lnet_peer_ni *lpni;
2920                                 lnet_nid_t nid;
2921
2922                                 md = lnet_handle2md(&rspt->rspt_mdh);
2923                                 if (!md) {
2924                                         /* MD has been queued for unlink, but
2925                                          * rspt hasn't been detached (Note we've
2926                                          * checked above that the rspt_mdh is
2927                                          * valid). Since we cannot lookup the MD
2928                                          * we're unable to detach the rspt
2929                                          * ourselves. Thus, move the rspt to the
2930                                          * zombie list where we'll wait for
2931                                          * either:
2932                                          *   1. The remaining operations on the
2933                                          *   MD to complete. In this case the
2934                                          *   final operation will result in
2935                                          *   lnet_msg_detach_md()->
2936                                          *   lnet_detach_rsp_tracker() where
2937                                          *   we will clean up this response
2938                                          *   tracker.
2939                                          *   2. LNet to shutdown. In this case
2940                                          *   we'll wait until after all LND Nets
2941                                          *   have shutdown and then we can
2942                                          *   safely free any remaining response
2943                                          *   tracker blocks on the zombie list.
2944                                          * Note: We need to hold the resource
2945                                          * lock when adding to the zombie list
2946                                          * because we may have concurrent access
2947                                          * with lnet_detach_rsp_tracker().
2948                                          */
2949                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2950                                         list_move(&rspt->rspt_on_list,
2951                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2952                                         lnet_res_unlock(i);
2953                                         continue;
2954                                 }
2955                                 LASSERT(md->md_rspt_ptr == rspt);
2956                                 md->md_rspt_ptr = NULL;
2957                                 lnet_res_unlock(i);
2958
2959                                 LNetMDUnlink(rspt->rspt_mdh);
2960
2961                                 nid = rspt->rspt_next_hop_nid;
2962
2963                                 list_del(&rspt->rspt_on_list);
2964                                 lnet_rspt_free(rspt, i);
2965
2966                                 /* If we're shutting down we just want to clean
2967                                  * up the rspt blocks
2968                                  */
2969                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2970                                         continue;
2971
2972                                 lnet_net_lock(i);
2973                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2974                                 lnet_net_unlock(i);
2975
2976                                 CDEBUG(D_NET,
2977                                        "Response timeout: md = %p: nid = %s\n",
2978                                        md, libcfs_nid2str(nid));
2979
2980                                 /*
2981                                  * If there is a timeout on the response
2982                                  * from the next hop decrement its health
2983                                  * value so that we don't use it
2984                                  */
2985                                 lnet_net_lock(0);
2986                                 lpni = lnet_find_peer_ni_locked(nid);
2987                                 if (lpni) {
2988                                         lnet_handle_remote_failure_locked(lpni);
2989                                         lnet_peer_ni_decref_locked(lpni);
2990                                 }
2991                                 lnet_net_unlock(0);
2992                         } else {
2993                                 lnet_res_unlock(i);
2994                                 break;
2995                         }
2996                 }
2997
2998                 if (!list_empty(&local_queue)) {
2999                         lnet_net_lock(i);
3000                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3001                         lnet_net_unlock(i);
3002                 }
3003         }
3004 }
3005
3006 static void
3007 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3008 {
3009         struct lnet_msg *msg;
3010
3011         while (!list_empty(resendq)) {
3012                 struct lnet_peer_ni *lpni;
3013
3014                 msg = list_entry(resendq->next, struct lnet_msg,
3015                                  msg_list);
3016
3017                 list_del_init(&msg->msg_list);
3018
3019                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3020                 if (!lpni) {
3021                         lnet_net_unlock(cpt);
3022                         CERROR("Expected that a peer is already created for %s\n",
3023                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3024                         msg->msg_no_resend = true;
3025                         lnet_finalize(msg, -EFAULT);
3026                         lnet_net_lock(cpt);
3027                 } else {
3028                         struct lnet_peer *peer;
3029                         int rc;
3030                         lnet_nid_t src_nid = LNET_NID_ANY;
3031
3032                         /*
3033                          * if this message is not being routed and the
3034                          * peer is non-MR then we must use the same
3035                          * src_nid that was used in the original send.
3036                          * Otherwise if we're routing the message (IE
3037                          * we're a router) then we can use any of our
3038                          * local interfaces. It doesn't matter to the
3039                          * final destination.
3040                          */
3041                         peer = lpni->lpni_peer_net->lpn_peer;
3042                         if (!msg->msg_routing &&
3043                             !lnet_peer_is_multi_rail(peer))
3044                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3045
3046                         /*
3047                          * If we originally specified a src NID, then we
3048                          * must attempt to reuse it in the resend as well.
3049                          */
3050                         if (msg->msg_src_nid_param != LNET_NID_ANY)
3051                                 src_nid = msg->msg_src_nid_param;
3052                         lnet_peer_ni_decref_locked(lpni);
3053
3054                         lnet_net_unlock(cpt);
3055                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3056                                libcfs_nid2str(src_nid),
3057                                libcfs_id2str(msg->msg_target),
3058                                lnet_msgtyp2str(msg->msg_type),
3059                                msg->msg_recovery,
3060                                msg->msg_retry_count);
3061                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3062                         if (rc) {
3063                                 CERROR("Error sending %s to %s: %d\n",
3064                                        lnet_msgtyp2str(msg->msg_type),
3065                                        libcfs_id2str(msg->msg_target), rc);
3066                                 msg->msg_no_resend = true;
3067                                 lnet_finalize(msg, rc);
3068                         }
3069                         lnet_net_lock(cpt);
3070                         if (!rc)
3071                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3072                 }
3073         }
3074 }
3075
3076 static void
3077 lnet_resend_pending_msgs(void)
3078 {
3079         int i;
3080
3081         cfs_cpt_for_each(i, lnet_cpt_table()) {
3082                 lnet_net_lock(i);
3083                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3084                 lnet_net_unlock(i);
3085         }
3086 }
3087
3088 /* called with cpt and ni_lock held */
3089 static void
3090 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3091 {
3092         struct lnet_handle_md recovery_mdh;
3093
3094         LNetInvalidateMDHandle(&recovery_mdh);
3095
3096         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3097             force) {
3098                 recovery_mdh = ni->ni_ping_mdh;
3099                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3100         }
3101         lnet_ni_unlock(ni);
3102         lnet_net_unlock(cpt);
3103         if (!LNetMDHandleIsInvalid(recovery_mdh))
3104                 LNetMDUnlink(recovery_mdh);
3105         lnet_net_lock(cpt);
3106         lnet_ni_lock(ni);
3107 }
3108
3109 static void
3110 lnet_recover_local_nis(void)
3111 {
3112         struct lnet_mt_event_info *ev_info;
3113         struct list_head processed_list;
3114         struct list_head local_queue;
3115         struct lnet_handle_md mdh;
3116         struct lnet_ni *tmp;
3117         struct lnet_ni *ni;
3118         lnet_nid_t nid;
3119         int healthv;
3120         int rc;
3121
3122         INIT_LIST_HEAD(&local_queue);
3123         INIT_LIST_HEAD(&processed_list);
3124
3125         /*
3126          * splice the recovery queue on a local queue. We will iterate
3127          * through the local queue and update it as needed. Once we're
3128          * done with the traversal, we'll splice the local queue back on
3129          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3130          * will be traversed in the next iteration.
3131          */
3132         lnet_net_lock(0);
3133         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3134                          &local_queue);
3135         lnet_net_unlock(0);
3136
3137         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3138                 /*
3139                  * if an NI is being deleted or it is now healthy, there
3140                  * is no need to keep it around in the recovery queue.
3141                  * The monitor thread is the only thread responsible for
3142                  * removing the NI from the recovery queue.
3143                  * Multiple threads can be adding NIs to the recovery
3144                  * queue.
3145                  */
3146                 healthv = atomic_read(&ni->ni_healthv);
3147
3148                 lnet_net_lock(0);
3149                 lnet_ni_lock(ni);
3150                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3151                     healthv == LNET_MAX_HEALTH_VALUE) {
3152                         list_del_init(&ni->ni_recovery);
3153                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3154                         lnet_ni_unlock(ni);
3155                         lnet_ni_decref_locked(ni, 0);
3156                         lnet_net_unlock(0);
3157                         continue;
3158                 }
3159
3160                 /*
3161                  * if the local NI failed recovery we must unlink the md.
3162                  * But we want to keep the local_ni on the recovery queue
3163                  * so we can continue the attempts to recover it.
3164                  */
3165                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3166                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3167                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3168                 }
3169
3170                 lnet_ni_unlock(ni);
3171                 lnet_net_unlock(0);
3172
3173
3174                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3175                        libcfs_nid2str(ni->ni_nid));
3176
3177                 lnet_ni_lock(ni);
3178                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3179                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3180                         lnet_ni_unlock(ni);
3181
3182                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3183                         if (!ev_info) {
3184                                 CERROR("out of memory. Can't recover %s\n",
3185                                        libcfs_nid2str(ni->ni_nid));
3186                                 lnet_ni_lock(ni);
3187                                 ni->ni_recovery_state &=
3188                                   ~LNET_NI_RECOVERY_PENDING;
3189                                 lnet_ni_unlock(ni);
3190                                 continue;
3191                         }
3192
3193                         mdh = ni->ni_ping_mdh;
3194                         /*
3195                          * Invalidate the ni mdh in case it's deleted.
3196                          * We'll unlink the mdh in this case below.
3197                          */
3198                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3199                         nid = ni->ni_nid;
3200
3201                         /*
3202                          * remove the NI from the local queue and drop the
3203                          * reference count to it while we're recovering
3204                          * it. The reason for that, is that the NI could
3205                          * be deleted, and the way the code is structured
3206                          * is if we don't drop the NI, then the deletion
3207                          * code will enter a loop waiting for the
3208                          * reference count to be removed while holding the
3209                          * ln_mutex_lock(). When we look up the peer to
3210                          * send to in lnet_select_pathway() we will try to
3211                          * lock the ln_mutex_lock() as well, leading to
3212                          * a deadlock. By dropping the refcount and
3213                          * removing it from the list, we allow for the NI
3214                          * to be removed, then we use the cached NID to
3215                          * look it up again. If it's gone, then we just
3216                          * continue examining the rest of the queue.
3217                          */
3218                         lnet_net_lock(0);
3219                         list_del_init(&ni->ni_recovery);
3220                         lnet_ni_decref_locked(ni, 0);
3221                         lnet_net_unlock(0);
3222
3223                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3224                         ev_info->mt_nid = nid;
3225                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3226                                             ev_info, the_lnet.ln_mt_eqh, true);
3227                         /* lookup the nid again */
3228                         lnet_net_lock(0);
3229                         ni = lnet_nid2ni_locked(nid, 0);
3230                         if (!ni) {
3231                                 /*
3232                                  * the NI has been deleted when we dropped
3233                                  * the ref count
3234                                  */
3235                                 lnet_net_unlock(0);
3236                                 LNetMDUnlink(mdh);
3237                                 continue;
3238                         }
3239                         /*
3240                          * Same note as in lnet_recover_peer_nis(). When
3241                          * we're sending the ping, the NI is free to be
3242                          * deleted or manipulated. By this point it
3243                          * could've been added back on the recovery queue,
3244                          * and a refcount taken on it.
3245                          * So we can't just add it blindly again or we'll
3246                          * corrupt the queue. We must check under lock if
3247                          * it's not on any list and if not then add it
3248                          * to the processed list, which will eventually be
3249                          * spliced back on to the recovery queue.
3250                          */
3251                         ni->ni_ping_mdh = mdh;
3252                         if (list_empty(&ni->ni_recovery)) {
3253                                 list_add_tail(&ni->ni_recovery, &processed_list);
3254                                 lnet_ni_addref_locked(ni, 0);
3255                         }
3256                         lnet_net_unlock(0);
3257
3258                         lnet_ni_lock(ni);
3259                         if (rc)
3260                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3261                 }
3262                 lnet_ni_unlock(ni);
3263         }
3264
3265         /*
3266          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3267          * reexamined in the next iteration.
3268          */
3269         list_splice_init(&processed_list, &local_queue);
3270         lnet_net_lock(0);
3271         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3272         lnet_net_unlock(0);
3273 }
3274
3275 static int
3276 lnet_resendqs_create(void)
3277 {
3278         struct list_head **resendqs;
3279         resendqs = lnet_create_array_of_queues();
3280
3281         if (!resendqs)
3282                 return -ENOMEM;
3283
3284         lnet_net_lock(LNET_LOCK_EX);
3285         the_lnet.ln_mt_resendqs = resendqs;
3286         lnet_net_unlock(LNET_LOCK_EX);
3287
3288         return 0;
3289 }
3290
3291 static void
3292 lnet_clean_local_ni_recoveryq(void)
3293 {
3294         struct lnet_ni *ni;
3295
3296         /* This is only called when the monitor thread has stopped */
3297         lnet_net_lock(0);
3298
3299         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3300                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3301                                 struct lnet_ni, ni_recovery);
3302                 list_del_init(&ni->ni_recovery);
3303                 lnet_ni_lock(ni);
3304                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3305                 lnet_ni_unlock(ni);
3306                 lnet_ni_decref_locked(ni, 0);
3307         }
3308
3309         lnet_net_unlock(0);
3310 }
3311
3312 static void
3313 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3314                                      bool force)
3315 {
3316         struct lnet_handle_md recovery_mdh;
3317
3318         LNetInvalidateMDHandle(&recovery_mdh);
3319
3320         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3321                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3322                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3323         }
3324         spin_unlock(&lpni->lpni_lock);
3325         lnet_net_unlock(cpt);
3326         if (!LNetMDHandleIsInvalid(recovery_mdh))
3327                 LNetMDUnlink(recovery_mdh);
3328         lnet_net_lock(cpt);
3329         spin_lock(&lpni->lpni_lock);
3330 }
3331
3332 static void
3333 lnet_clean_peer_ni_recoveryq(void)
3334 {
3335         struct lnet_peer_ni *lpni, *tmp;
3336
3337         lnet_net_lock(LNET_LOCK_EX);
3338
3339         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3340                                  lpni_recovery) {
3341                 list_del_init(&lpni->lpni_recovery);
3342                 spin_lock(&lpni->lpni_lock);
3343                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3344                 spin_unlock(&lpni->lpni_lock);
3345                 lnet_peer_ni_decref_locked(lpni);
3346         }
3347
3348         lnet_net_unlock(LNET_LOCK_EX);
3349 }
3350
3351 static void
3352 lnet_clean_resendqs(void)
3353 {
3354         struct lnet_msg *msg, *tmp;
3355         struct list_head msgs;
3356         int i;
3357
3358         INIT_LIST_HEAD(&msgs);
3359
3360         cfs_cpt_for_each(i, lnet_cpt_table()) {
3361                 lnet_net_lock(i);
3362                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3363                 lnet_net_unlock(i);
3364                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3365                         list_del_init(&msg->msg_list);
3366                         msg->msg_no_resend = true;
3367                         lnet_finalize(msg, -ESHUTDOWN);
3368                 }
3369         }
3370
3371         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3372 }
3373
3374 static void
3375 lnet_recover_peer_nis(void)
3376 {
3377         struct lnet_mt_event_info *ev_info;
3378         struct list_head processed_list;
3379         struct list_head local_queue;
3380         struct lnet_handle_md mdh;
3381         struct lnet_peer_ni *lpni;
3382         struct lnet_peer_ni *tmp;
3383         lnet_nid_t nid;
3384         int healthv;
3385         int rc;
3386
3387         INIT_LIST_HEAD(&local_queue);
3388         INIT_LIST_HEAD(&processed_list);
3389
3390         /*
3391          * Always use cpt 0 for locking across all interactions with
3392          * ln_mt_peerNIRecovq
3393          */
3394         lnet_net_lock(0);
3395         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3396                          &local_queue);
3397         lnet_net_unlock(0);
3398
3399         list_for_each_entry_safe(lpni, tmp, &local_queue,
3400                                  lpni_recovery) {
3401                 /*
3402                  * The same protection strategy is used here as is in the
3403                  * local recovery case.
3404                  */
3405                 lnet_net_lock(0);
3406                 healthv = atomic_read(&lpni->lpni_healthv);
3407                 spin_lock(&lpni->lpni_lock);
3408                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3409                     healthv == LNET_MAX_HEALTH_VALUE) {
3410                         list_del_init(&lpni->lpni_recovery);
3411                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3412                         spin_unlock(&lpni->lpni_lock);
3413                         lnet_peer_ni_decref_locked(lpni);
3414                         lnet_net_unlock(0);
3415                         continue;
3416                 }
3417
3418                 /*
3419                  * If the peer NI has failed recovery we must unlink the
3420                  * md. But we want to keep the peer ni on the recovery
3421                  * queue so we can try to continue recovering it
3422                  */
3423                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3424                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3425                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3426                 }
3427
3428                 spin_unlock(&lpni->lpni_lock);
3429                 lnet_net_unlock(0);
3430
3431                 /*
3432                  * NOTE: we're racing with peer deletion from user space.
3433                  * It's possible that a peer is deleted after we check its
3434                  * state. In this case the recovery can create a new peer
3435                  */
3436                 spin_lock(&lpni->lpni_lock);
3437                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3438                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3439                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3440                         spin_unlock(&lpni->lpni_lock);
3441
3442                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3443                         if (!ev_info) {
3444                                 CERROR("out of memory. Can't recover %s\n",
3445                                        libcfs_nid2str(lpni->lpni_nid));
3446                                 spin_lock(&lpni->lpni_lock);
3447                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3448                                 spin_unlock(&lpni->lpni_lock);
3449                                 continue;
3450                         }
3451
3452                         /* look at the comments in lnet_recover_local_nis() */
3453                         mdh = lpni->lpni_recovery_ping_mdh;
3454                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3455                         nid = lpni->lpni_nid;
3456                         lnet_net_lock(0);
3457                         list_del_init(&lpni->lpni_recovery);
3458                         lnet_peer_ni_decref_locked(lpni);
3459                         lnet_net_unlock(0);
3460
3461                         ev_info->mt_type = MT_TYPE_PEER_NI;
3462                         ev_info->mt_nid = nid;
3463                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3464                                             ev_info, the_lnet.ln_mt_eqh, true);
3465                         lnet_net_lock(0);
3466                         /*
3467                          * lnet_find_peer_ni_locked() grabs a refcount for
3468                          * us. No need to take it explicitly.
3469                          */
3470                         lpni = lnet_find_peer_ni_locked(nid);
3471                         if (!lpni) {
3472                                 lnet_net_unlock(0);
3473                                 LNetMDUnlink(mdh);
3474                                 continue;
3475                         }
3476
3477                         lpni->lpni_recovery_ping_mdh = mdh;
3478                         /*
3479                          * While we're unlocked the lpni could've been
3480                          * readded on the recovery queue. In this case we
3481                          * don't need to add it to the local queue, since
3482                          * it's already on there and the thread that added
3483                          * it would've incremented the refcount on the
3484                          * peer, which means we need to decref the refcount
3485                          * that was implicitly grabbed by find_peer_ni_locked.
3486                          * Otherwise, if the lpni is still not on
3487                          * the recovery queue, then we'll add it to the
3488                          * processed list.
3489                          */
3490                         if (list_empty(&lpni->lpni_recovery))
3491                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3492                         else
3493                                 lnet_peer_ni_decref_locked(lpni);
3494                         lnet_net_unlock(0);
3495
3496                         spin_lock(&lpni->lpni_lock);
3497                         if (rc)
3498                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3499                 }
3500                 spin_unlock(&lpni->lpni_lock);
3501         }
3502
3503         list_splice_init(&processed_list, &local_queue);
3504         lnet_net_lock(0);
3505         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3506         lnet_net_unlock(0);
3507 }
3508
3509 static int
3510 lnet_monitor_thread(void *arg)
3511 {
3512         time64_t recovery_timeout = 0;
3513         time64_t rsp_timeout = 0;
3514         int interval;
3515         time64_t now;
3516
3517         wait_for_completion(&the_lnet.ln_started);
3518         /*
3519          * The monitor thread takes care of the following:
3520          *  1. Checks the aliveness of routers
3521          *  2. Checks if there are messages on the resend queue to resend
3522          *     them.
3523          *  3. Check if there are any NIs on the local recovery queue and
3524          *     pings them
3525          *  4. Checks if there are any NIs on the remote recovery queue
3526          *     and pings them.
3527          */
3528         cfs_block_allsigs();
3529
3530         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3531                 now = ktime_get_real_seconds();
3532
3533                 if (lnet_router_checker_active())
3534                         lnet_check_routers();
3535
3536                 lnet_resend_pending_msgs();
3537
3538                 if (now >= rsp_timeout) {
3539                         lnet_finalize_expired_responses();
3540                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3541                 }
3542
3543                 if (now >= recovery_timeout) {
3544                         lnet_recover_local_nis();
3545                         lnet_recover_peer_nis();
3546                         recovery_timeout = now + lnet_recovery_interval;
3547                 }
3548
3549                 /*
3550                  * TODO do we need to check if we should sleep without
3551                  * timeout?  Technically, an active system will always
3552                  * have messages in flight so this check will always
3553                  * evaluate to false. And on an idle system do we care
3554                  * if we wake up every 1 second? Although, we've seen
3555                  * cases where we get a complaint that an idle thread
3556                  * is waking up unnecessarily.
3557                  *
3558                  * Take into account the current net_count when you wake
3559                  * up for alive router checking, since we need to check
3560                  * possibly as many networks as we have configured.
3561                  */
3562                 interval = min(lnet_recovery_interval,
3563                                min((unsigned int) alive_router_check_interval /
3564                                         lnet_current_net_count,
3565                                    lnet_transaction_timeout / 2));
3566                 wait_for_completion_interruptible_timeout(
3567                         &the_lnet.ln_mt_wait_complete,
3568                         cfs_time_seconds(interval));
3569                 /* Must re-init the completion before testing anything,
3570                  * including ln_mt_state.
3571                  */
3572                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3573         }
3574
3575         /* Shutting down */
3576         lnet_net_lock(LNET_LOCK_EX);
3577         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3578         lnet_net_unlock(LNET_LOCK_EX);
3579
3580         /* signal that the monitor thread is exiting */
3581         up(&the_lnet.ln_mt_signal);
3582
3583         return 0;
3584 }
3585
3586 /*
3587  * lnet_send_ping
3588  * Sends a ping.
3589  * Returns == 0 if success
3590  * Returns > 0 if LNetMDBind or prior fails
3591  * Returns < 0 if LNetGet fails
3592  */
3593 int
3594 lnet_send_ping(lnet_nid_t dest_nid,
3595                struct lnet_handle_md *mdh, int nnis,
3596                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3597 {
3598         struct lnet_md md = { NULL };
3599         struct lnet_process_id id;
3600         struct lnet_ping_buffer *pbuf;
3601         int rc;
3602
3603         if (dest_nid == LNET_NID_ANY) {
3604                 rc = -EHOSTUNREACH;
3605                 goto fail_error;
3606         }
3607
3608         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3609         if (!pbuf) {
3610                 rc = ENOMEM;
3611                 goto fail_error;
3612         }
3613
3614         /* initialize md content */
3615         md.start     = &pbuf->pb_info;
3616         md.length    = LNET_PING_INFO_SIZE(nnis);
3617         md.threshold = 2; /* GET/REPLY */
3618         md.max_size  = 0;
3619         md.options   = LNET_MD_TRUNCATE;
3620         md.user_ptr  = user_data;
3621         md.eq_handle = eqh;
3622
3623         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3624         if (rc) {
3625                 lnet_ping_buffer_decref(pbuf);
3626                 CERROR("Can't bind MD: %d\n", rc);
3627                 rc = -rc; /* change the rc to positive */
3628                 goto fail_error;
3629         }
3630         id.pid = LNET_PID_LUSTRE;
3631         id.nid = dest_nid;
3632
3633         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3634                      LNET_RESERVED_PORTAL,
3635                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3636
3637         if (rc)
3638                 goto fail_unlink_md;
3639
3640         return 0;
3641
3642 fail_unlink_md:
3643         LNetMDUnlink(*mdh);
3644         LNetInvalidateMDHandle(mdh);
3645 fail_error:
3646         return rc;
3647 }
3648
3649 static void
3650 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3651                            int status, bool unlink_event)
3652 {
3653         lnet_nid_t nid = ev_info->mt_nid;
3654
3655         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3656                 struct lnet_ni *ni;
3657
3658                 lnet_net_lock(0);
3659                 ni = lnet_nid2ni_locked(nid, 0);
3660                 if (!ni) {
3661                         lnet_net_unlock(0);
3662                         return;
3663                 }
3664                 lnet_ni_lock(ni);
3665                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3666                 if (status)
3667                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3668                 lnet_ni_unlock(ni);
3669                 lnet_net_unlock(0);
3670
3671                 if (status != 0) {
3672                         CERROR("local NI (%s) recovery failed with %d\n",
3673                                libcfs_nid2str(nid), status);
3674                         return;
3675                 }
3676                 /*
3677                  * need to increment healthv for the ni here, because in
3678                  * the lnet_finalize() path we don't have access to this
3679                  * NI. And in order to get access to it, we'll need to
3680                  * carry forward too much information.
3681                  * In the peer case, it'll naturally be incremented
3682                  */
3683                 if (!unlink_event)
3684                         lnet_inc_healthv(&ni->ni_healthv);
3685         } else {
3686                 struct lnet_peer_ni *lpni;
3687                 int cpt;
3688
3689                 cpt = lnet_net_lock_current();
3690                 lpni = lnet_find_peer_ni_locked(nid);
3691                 if (!lpni) {
3692                         lnet_net_unlock(cpt);
3693                         return;
3694                 }
3695                 spin_lock(&lpni->lpni_lock);
3696                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3697                 if (status)
3698                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3699                 spin_unlock(&lpni->lpni_lock);
3700                 lnet_peer_ni_decref_locked(lpni);
3701                 lnet_net_unlock(cpt);
3702
3703                 if (status != 0)
3704                         CERROR("peer NI (%s) recovery failed with %d\n",
3705                                libcfs_nid2str(nid), status);
3706         }
3707 }
3708
3709 void
3710 lnet_mt_event_handler(struct lnet_event *event)
3711 {
3712         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3713         struct lnet_ping_buffer *pbuf;
3714
3715         /* TODO: remove assert */
3716         LASSERT(event->type == LNET_EVENT_REPLY ||
3717                 event->type == LNET_EVENT_SEND ||
3718                 event->type == LNET_EVENT_UNLINK);
3719
3720         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3721                event->status);
3722
3723         switch (event->type) {
3724         case LNET_EVENT_UNLINK:
3725                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3726                        libcfs_nid2str(ev_info->mt_nid));
3727                 /* fallthrough */
3728         case LNET_EVENT_REPLY:
3729                 lnet_handle_recovery_reply(ev_info, event->status,
3730                                            event->type == LNET_EVENT_UNLINK);
3731                 break;
3732         case LNET_EVENT_SEND:
3733                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3734                                libcfs_nid2str(ev_info->mt_nid),
3735                                (event->status) ? "unsuccessfully" :
3736                                "successfully", event->status);
3737                 break;
3738         default:
3739                 CERROR("Unexpected event: %d\n", event->type);
3740                 break;
3741         }
3742         if (event->unlinked) {
3743                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3744                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3745                 lnet_ping_buffer_decref(pbuf);
3746         }
3747 }
3748
3749 static int
3750 lnet_rsp_tracker_create(void)
3751 {
3752         struct list_head **rstqs;
3753         rstqs = lnet_create_array_of_queues();
3754
3755         if (!rstqs)
3756                 return -ENOMEM;
3757
3758         the_lnet.ln_mt_rstq = rstqs;
3759
3760         return 0;
3761 }
3762
3763 static void
3764 lnet_rsp_tracker_clean(void)
3765 {
3766         lnet_finalize_expired_responses();
3767
3768         cfs_percpt_free(the_lnet.ln_mt_rstq);
3769         the_lnet.ln_mt_rstq = NULL;
3770 }
3771
3772 int lnet_monitor_thr_start(void)
3773 {
3774         int rc = 0;
3775         struct task_struct *task;
3776
3777         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3778                 return -EALREADY;
3779
3780         rc = lnet_resendqs_create();
3781         if (rc)
3782                 return rc;
3783
3784         rc = lnet_rsp_tracker_create();
3785         if (rc)
3786                 goto clean_queues;
3787
3788         sema_init(&the_lnet.ln_mt_signal, 0);
3789
3790         lnet_net_lock(LNET_LOCK_EX);
3791         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3792         lnet_net_unlock(LNET_LOCK_EX);
3793         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3794         if (IS_ERR(task)) {
3795                 rc = PTR_ERR(task);
3796                 CERROR("Can't start monitor thread: %d\n", rc);
3797                 goto clean_thread;
3798         }
3799
3800         return 0;
3801
3802 clean_thread:
3803         lnet_net_lock(LNET_LOCK_EX);
3804         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3805         lnet_net_unlock(LNET_LOCK_EX);
3806         /* block until event callback signals exit */
3807         down(&the_lnet.ln_mt_signal);
3808         /* clean up */
3809         lnet_net_lock(LNET_LOCK_EX);
3810         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3811         lnet_net_unlock(LNET_LOCK_EX);
3812         lnet_rsp_tracker_clean();
3813         lnet_clean_local_ni_recoveryq();
3814         lnet_clean_peer_ni_recoveryq();
3815         lnet_clean_resendqs();
3816         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3817         return rc;
3818 clean_queues:
3819         lnet_rsp_tracker_clean();
3820         lnet_clean_local_ni_recoveryq();
3821         lnet_clean_peer_ni_recoveryq();
3822         lnet_clean_resendqs();
3823         return rc;
3824 }
3825
3826 void lnet_monitor_thr_stop(void)
3827 {
3828         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3829                 return;
3830
3831         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3832         lnet_net_lock(LNET_LOCK_EX);
3833         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3834         lnet_net_unlock(LNET_LOCK_EX);
3835
3836         /* tell the monitor thread that we're shutting down */
3837         complete(&the_lnet.ln_mt_wait_complete);
3838
3839         /* block until monitor thread signals that it's done */
3840         down(&the_lnet.ln_mt_signal);
3841         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3842
3843         /* perform cleanup tasks */
3844         lnet_rsp_tracker_clean();
3845         lnet_clean_local_ni_recoveryq();
3846         lnet_clean_peer_ni_recoveryq();
3847         lnet_clean_resendqs();
3848 }
3849
3850 void
3851 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3852                   __u32 msg_type)
3853 {
3854         lnet_net_lock(cpt);
3855         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3856         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3857         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3858         lnet_net_unlock(cpt);
3859
3860         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3861 }
3862
3863 static void
3864 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3865 {
3866         struct lnet_hdr *hdr = &msg->msg_hdr;
3867
3868         if (msg->msg_wanted != 0)
3869                 lnet_setpayloadbuffer(msg);
3870
3871         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3872
3873         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3874          * it back into the ACK during lnet_finalize() */
3875         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3876                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3877
3878         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3879                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3880 }
3881
3882 static int
3883 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3884 {
3885         struct lnet_hdr         *hdr = &msg->msg_hdr;
3886         struct lnet_match_info  info;
3887         int                     rc;
3888         bool                    ready_delay;
3889
3890         /* Convert put fields to host byte order */
3891         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3892         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3893         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3894
3895         /* Primary peer NID. */
3896         info.mi_id.nid  = msg->msg_initiator;
3897         info.mi_id.pid  = hdr->src_pid;
3898         info.mi_opc     = LNET_MD_OP_PUT;
3899         info.mi_portal  = hdr->msg.put.ptl_index;
3900         info.mi_rlength = hdr->payload_length;
3901         info.mi_roffset = hdr->msg.put.offset;
3902         info.mi_mbits   = hdr->msg.put.match_bits;
3903         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3904
3905         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3906         ready_delay = msg->msg_rx_ready_delay;
3907
3908  again:
3909         rc = lnet_ptl_match_md(&info, msg);
3910         switch (rc) {
3911         default:
3912                 LBUG();
3913
3914         case LNET_MATCHMD_OK:
3915                 lnet_recv_put(ni, msg);
3916                 return 0;
3917
3918         case LNET_MATCHMD_NONE:
3919                 if (ready_delay)
3920                         /* no eager_recv or has already called it, should
3921                          * have been attached on delayed list */
3922                         return 0;
3923
3924                 rc = lnet_ni_eager_recv(ni, msg);
3925                 if (rc == 0) {
3926                         ready_delay = true;
3927                         goto again;
3928                 }
3929                 /* fall through */
3930
3931         case LNET_MATCHMD_DROP:
3932                 CNETERR("Dropping PUT from %s portal %d match %llu"
3933                         " offset %d length %d: %d\n",
3934                         libcfs_id2str(info.mi_id), info.mi_portal,
3935                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3936
3937                 return -ENOENT; /* -ve: OK but no match */
3938         }
3939 }
3940
3941 static int
3942 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3943 {
3944         struct lnet_match_info info;
3945         struct lnet_hdr *hdr = &msg->msg_hdr;
3946         struct lnet_process_id source_id;
3947         struct lnet_handle_wire reply_wmd;
3948         int rc;
3949
3950         /* Convert get fields to host byte order */
3951         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3952         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3953         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3954         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3955
3956         source_id.nid = hdr->src_nid;
3957         source_id.pid = hdr->src_pid;
3958         /* Primary peer NID */
3959         info.mi_id.nid  = msg->msg_initiator;
3960         info.mi_id.pid  = hdr->src_pid;
3961         info.mi_opc     = LNET_MD_OP_GET;
3962         info.mi_portal  = hdr->msg.get.ptl_index;
3963         info.mi_rlength = hdr->msg.get.sink_length;
3964         info.mi_roffset = hdr->msg.get.src_offset;
3965         info.mi_mbits   = hdr->msg.get.match_bits;
3966         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3967
3968         rc = lnet_ptl_match_md(&info, msg);
3969         if (rc == LNET_MATCHMD_DROP) {
3970                 CNETERR("Dropping GET from %s portal %d match %llu"
3971                         " offset %d length %d\n",
3972                         libcfs_id2str(info.mi_id), info.mi_portal,
3973                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3974                 return -ENOENT; /* -ve: OK but no match */
3975         }
3976
3977         LASSERT(rc == LNET_MATCHMD_OK);
3978
3979         lnet_build_msg_event(msg, LNET_EVENT_GET);
3980
3981         reply_wmd = hdr->msg.get.return_wmd;
3982
3983         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3984                        msg->msg_offset, msg->msg_wanted);
3985
3986         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3987
3988         if (rdma_get) {
3989                 /* The LND completes the REPLY from her recv procedure */
3990                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3991                              msg->msg_offset, msg->msg_len, msg->msg_len);
3992                 return 0;
3993         }
3994
3995         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3996         msg->msg_receiving = 0;
3997
3998         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3999         if (rc < 0) {
4000                 /* didn't get as far as lnet_ni_send() */
4001                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4002                        libcfs_nid2str(ni->ni_nid),
4003                        libcfs_id2str(info.mi_id), rc);
4004
4005                 lnet_finalize(msg, rc);
4006         }
4007
4008         return 0;
4009 }
4010
4011 static int
4012 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4013 {
4014         void *private = msg->msg_private;
4015         struct lnet_hdr *hdr = &msg->msg_hdr;
4016         struct lnet_process_id src = {0};
4017         struct lnet_libmd *md;
4018         int rlength;
4019         int mlength;
4020         int cpt;
4021
4022         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4023         lnet_res_lock(cpt);
4024
4025         src.nid = hdr->src_nid;
4026         src.pid = hdr->src_pid;
4027
4028         /* NB handles only looked up by creator (no flips) */
4029         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4030         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4031                 CNETERR("%s: Dropping REPLY from %s for %s "
4032                         "MD %#llx.%#llx\n",
4033                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4034                         (md == NULL) ? "invalid" : "inactive",
4035                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4036                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4037                 if (md != NULL && md->md_me != NULL)
4038                         CERROR("REPLY MD also attached to portal %d\n",
4039                                md->md_me->me_portal);
4040
4041                 lnet_res_unlock(cpt);
4042                 return -ENOENT; /* -ve: OK but no match */
4043         }
4044
4045         LASSERT(md->md_offset == 0);
4046
4047         rlength = hdr->payload_length;
4048         mlength = MIN(rlength, (int)md->md_length);
4049
4050         if (mlength < rlength &&
4051             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4052                 CNETERR("%s: Dropping REPLY from %s length %d "
4053                         "for MD %#llx would overflow (%d)\n",
4054                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4055                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4056                         mlength);
4057                 lnet_res_unlock(cpt);
4058                 return -ENOENT; /* -ve: OK but no match */
4059         }
4060
4061         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4062                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4063                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4064
4065         lnet_msg_attach_md(msg, md, 0, mlength);
4066
4067         if (mlength != 0)
4068                 lnet_setpayloadbuffer(msg);
4069
4070         lnet_res_unlock(cpt);
4071
4072         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4073
4074         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4075         return 0;
4076 }
4077
4078 static int
4079 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4080 {
4081         struct lnet_hdr *hdr = &msg->msg_hdr;
4082         struct lnet_process_id src = {0};
4083         struct lnet_libmd *md;
4084         int cpt;
4085
4086         src.nid = hdr->src_nid;
4087         src.pid = hdr->src_pid;
4088
4089         /* Convert ack fields to host byte order */
4090         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4091         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4092
4093         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4094         lnet_res_lock(cpt);
4095
4096         /* NB handles only looked up by creator (no flips) */
4097         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4098         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4099                 /* Don't moan; this is expected */
4100                 CDEBUG(D_NET,
4101                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4102                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4103                        (md == NULL) ? "invalid" : "inactive",
4104                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4105                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4106                 if (md != NULL && md->md_me != NULL)
4107                         CERROR("Source MD also attached to portal %d\n",
4108                                md->md_me->me_portal);
4109
4110                 lnet_res_unlock(cpt);
4111                 return -ENOENT;                  /* -ve! */
4112         }
4113
4114         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4115                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4116                hdr->msg.ack.dst_wmd.wh_object_cookie);
4117
4118         lnet_msg_attach_md(msg, md, 0, 0);
4119
4120         lnet_res_unlock(cpt);
4121
4122         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4123
4124         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4125         return 0;
4126 }
4127
4128 /**
4129  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4130  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4131  * \retval -ve                  error code
4132  */
4133 int
4134 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4135 {
4136         int     rc = 0;
4137
4138         if (!the_lnet.ln_routing)
4139                 return -ECANCELED;
4140
4141         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4142             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4143                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4144                         msg->msg_rx_ready_delay = 1;
4145                 } else {
4146                         lnet_net_unlock(msg->msg_rx_cpt);
4147                         rc = lnet_ni_eager_recv(ni, msg);
4148                         lnet_net_lock(msg->msg_rx_cpt);
4149                 }
4150         }
4151
4152         if (rc == 0)
4153                 rc = lnet_post_routed_recv_locked(msg, 0);
4154         return rc;
4155 }
4156
4157 int
4158 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4159 {
4160         int     rc;
4161
4162         switch (msg->msg_type) {
4163         case LNET_MSG_ACK:
4164                 rc = lnet_parse_ack(ni, msg);
4165                 break;
4166         case LNET_MSG_PUT:
4167                 rc = lnet_parse_put(ni, msg);
4168                 break;
4169         case LNET_MSG_GET:
4170                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4171                 break;
4172         case LNET_MSG_REPLY:
4173                 rc = lnet_parse_reply(ni, msg);
4174                 break;
4175         default: /* prevent an unused label if !kernel */
4176                 LASSERT(0);
4177                 return -EPROTO;
4178         }
4179
4180         LASSERT(rc == 0 || rc == -ENOENT);
4181         return rc;
4182 }
4183
4184 char *
4185 lnet_msgtyp2str (int type)
4186 {
4187         switch (type) {
4188         case LNET_MSG_ACK:
4189                 return ("ACK");
4190         case LNET_MSG_PUT:
4191                 return ("PUT");
4192         case LNET_MSG_GET:
4193                 return ("GET");
4194         case LNET_MSG_REPLY:
4195                 return ("REPLY");
4196         case LNET_MSG_HELLO:
4197                 return ("HELLO");
4198         default:
4199                 return ("<UNKNOWN>");
4200         }
4201 }
4202
4203 void
4204 lnet_print_hdr(struct lnet_hdr *hdr)
4205 {
4206         struct lnet_process_id src = {
4207                 .nid = hdr->src_nid,
4208                 .pid = hdr->src_pid,
4209         };
4210         struct lnet_process_id dst = {
4211                 .nid = hdr->dest_nid,
4212                 .pid = hdr->dest_pid,
4213         };
4214         char *type_str = lnet_msgtyp2str(hdr->type);
4215
4216         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4217         CWARN("    From %s\n", libcfs_id2str(src));
4218         CWARN("    To   %s\n", libcfs_id2str(dst));
4219
4220         switch (hdr->type) {
4221         default:
4222                 break;
4223
4224         case LNET_MSG_PUT:
4225                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4226                       "match bits %llu\n",
4227                       hdr->msg.put.ptl_index,
4228                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4229                       hdr->msg.put.ack_wmd.wh_object_cookie,
4230                       hdr->msg.put.match_bits);
4231                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4232                       hdr->payload_length, hdr->msg.put.offset,
4233                       hdr->msg.put.hdr_data);
4234                 break;
4235
4236         case LNET_MSG_GET:
4237                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4238                       "match bits %llu\n", hdr->msg.get.ptl_index,
4239                       hdr->msg.get.return_wmd.wh_interface_cookie,
4240                       hdr->msg.get.return_wmd.wh_object_cookie,
4241                       hdr->msg.get.match_bits);
4242                 CWARN("    Length %d, src offset %d\n",
4243                       hdr->msg.get.sink_length,
4244                       hdr->msg.get.src_offset);
4245                 break;
4246
4247         case LNET_MSG_ACK:
4248                 CWARN("    dst md %#llx.%#llx, "
4249                       "manipulated length %d\n",
4250                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4251                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4252                       hdr->msg.ack.mlength);
4253                 break;
4254
4255         case LNET_MSG_REPLY:
4256                 CWARN("    dst md %#llx.%#llx, "
4257                       "length %d\n",
4258                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4259                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4260                       hdr->payload_length);
4261         }
4262
4263 }
4264
4265 int
4266 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4267            void *private, int rdma_req)
4268 {
4269         struct lnet_peer_ni *lpni;
4270         struct lnet_msg *msg;
4271         __u32 payload_length;
4272         lnet_pid_t dest_pid;
4273         lnet_nid_t dest_nid;
4274         lnet_nid_t src_nid;
4275         bool push = false;
4276         int for_me;
4277         __u32 type;
4278         int rc = 0;
4279         int cpt;
4280
4281         LASSERT (!in_interrupt ());
4282
4283         type = le32_to_cpu(hdr->type);
4284         src_nid = le64_to_cpu(hdr->src_nid);
4285         dest_nid = le64_to_cpu(hdr->dest_nid);
4286         dest_pid = le32_to_cpu(hdr->dest_pid);
4287         payload_length = le32_to_cpu(hdr->payload_length);
4288
4289         for_me = (ni->ni_nid == dest_nid);
4290         cpt = lnet_cpt_of_nid(from_nid, ni);
4291
4292         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4293                 libcfs_nid2str(dest_nid),
4294                 libcfs_nid2str(ni->ni_nid),
4295                 libcfs_nid2str(src_nid),
4296                 lnet_msgtyp2str(type),
4297                 (for_me) ? "for me" : "routed");
4298
4299         switch (type) {
4300         case LNET_MSG_ACK:
4301         case LNET_MSG_GET:
4302                 if (payload_length > 0) {
4303                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4304                                libcfs_nid2str(from_nid),
4305                                libcfs_nid2str(src_nid),
4306                                lnet_msgtyp2str(type), payload_length);
4307                         return -EPROTO;
4308                 }
4309                 break;
4310
4311         case LNET_MSG_PUT:
4312         case LNET_MSG_REPLY:
4313                 if (payload_length >
4314                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4315                         CERROR("%s, src %s: bad %s payload %d "
4316                                "(%d max expected)\n",
4317                                libcfs_nid2str(from_nid),
4318                                libcfs_nid2str(src_nid),
4319                                lnet_msgtyp2str(type),
4320                                payload_length,
4321                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4322                         return -EPROTO;
4323                 }
4324                 break;
4325
4326         default:
4327                 CERROR("%s, src %s: Bad message type 0x%x\n",
4328                        libcfs_nid2str(from_nid),
4329                        libcfs_nid2str(src_nid), type);
4330                 return -EPROTO;
4331         }
4332
4333         if (the_lnet.ln_routing &&
4334             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4335                 lnet_ni_lock(ni);
4336                 spin_lock(&ni->ni_net->net_lock);
4337                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4338                 spin_unlock(&ni->ni_net->net_lock);
4339                 if (ni->ni_status != NULL &&
4340                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4341                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4342                         push = true;
4343                 }
4344                 lnet_ni_unlock(ni);
4345         }
4346
4347         if (push)
4348                 lnet_push_update_to_peers(1);
4349
4350         /* Regard a bad destination NID as a protocol error.  Senders should
4351          * know what they're doing; if they don't they're misconfigured, buggy
4352          * or malicious so we chop them off at the knees :) */
4353
4354         if (!for_me) {
4355                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4356                         /* should have gone direct */
4357                         CERROR("%s, src %s: Bad dest nid %s "
4358                                "(should have been sent direct)\n",
4359                                 libcfs_nid2str(from_nid),
4360                                 libcfs_nid2str(src_nid),
4361                                 libcfs_nid2str(dest_nid));
4362                         return -EPROTO;
4363                 }
4364
4365                 if (lnet_islocalnid(dest_nid)) {
4366                         /* dest is another local NI; sender should have used
4367                          * this node's NID on its own network */
4368                         CERROR("%s, src %s: Bad dest nid %s "
4369                                "(it's my nid but on a different network)\n",
4370                                 libcfs_nid2str(from_nid),
4371                                 libcfs_nid2str(src_nid),
4372                                 libcfs_nid2str(dest_nid));
4373                         return -EPROTO;
4374                 }
4375
4376                 if (rdma_req && type == LNET_MSG_GET) {
4377                         CERROR("%s, src %s: Bad optimized GET for %s "
4378                                "(final destination must be me)\n",
4379                                 libcfs_nid2str(from_nid),
4380                                 libcfs_nid2str(src_nid),
4381                                 libcfs_nid2str(dest_nid));
4382                         return -EPROTO;
4383                 }
4384
4385                 if (!the_lnet.ln_routing) {
4386                         CERROR("%s, src %s: Dropping message for %s "
4387                                "(routing not enabled)\n",
4388                                 libcfs_nid2str(from_nid),
4389                                 libcfs_nid2str(src_nid),
4390                                 libcfs_nid2str(dest_nid));
4391                         goto drop;
4392                 }
4393         }
4394
4395         /* Message looks OK; we're not going to return an error, so we MUST
4396          * call back lnd_recv() come what may... */
4397
4398         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4399             fail_peer(src_nid, 0)) {                    /* shall we now? */
4400                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4401                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4402                        lnet_msgtyp2str(type));
4403                 goto drop;
4404         }
4405
4406         if (!list_empty(&the_lnet.ln_drop_rules) &&
4407             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4408                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4409                               "silent message loss\n",
4410                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4411                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4412                 goto drop;
4413         }
4414
4415         if (lnet_drop_asym_route && for_me &&
4416             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4417                 struct lnet_net *net;
4418                 struct lnet_remotenet *rnet;
4419                 bool found = true;
4420
4421                 /* we are dealing with a routed message,
4422                  * so see if route to reach src_nid goes through from_nid
4423                  */
4424                 lnet_net_lock(cpt);
4425                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4426                 if (!net) {
4427                         lnet_net_unlock(cpt);
4428                         CERROR("net %s not found\n",
4429                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4430                         return -EPROTO;
4431                 }
4432
4433                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4434                 if (rnet) {
4435                         struct lnet_peer *gw = NULL;
4436                         struct lnet_peer_ni *lpni = NULL;
4437                         struct lnet_route *route;
4438
4439                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4440                                 found = false;
4441                                 gw = route->lr_gateway;
4442                                 if (route->lr_lnet != net->net_id)
4443                                         continue;
4444                                 /*
4445                                  * if the nid is one of the gateway's NIDs
4446                                  * then this is a valid gateway
4447                                  */
4448                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4449                                                 NULL, lpni)) != NULL) {
4450                                         if (lpni->lpni_nid == from_nid) {
4451                                                 found = true;
4452                                                 break;
4453                                         }
4454                                 }
4455                         }
4456                 }
4457                 lnet_net_unlock(cpt);
4458                 if (!found) {
4459                         /* we would not use from_nid to route a message to
4460                          * src_nid
4461                          * => asymmetric routing detected but forbidden
4462                          */
4463                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4464                                libcfs_nid2str(from_nid),
4465                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4466                         goto drop;
4467                 }
4468         }
4469
4470         msg = lnet_msg_alloc();
4471         if (msg == NULL) {
4472                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4473                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4474                        lnet_msgtyp2str(type));
4475                 goto drop;
4476         }
4477
4478         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4479          * pointers NULL etc */
4480
4481         msg->msg_type = type;
4482         msg->msg_private = private;
4483         msg->msg_receiving = 1;
4484         msg->msg_rdma_get = rdma_req;
4485         msg->msg_len = msg->msg_wanted = payload_length;
4486         msg->msg_offset = 0;
4487         msg->msg_hdr = *hdr;
4488         /* for building message event */
4489         msg->msg_from = from_nid;
4490         if (!for_me) {
4491                 msg->msg_target.pid     = dest_pid;
4492                 msg->msg_target.nid     = dest_nid;
4493                 msg->msg_routing        = 1;
4494
4495         } else {
4496                 /* convert common msg->hdr fields to host byteorder */
4497                 msg->msg_hdr.type       = type;
4498                 msg->msg_hdr.src_nid    = src_nid;
4499                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4500                 msg->msg_hdr.dest_nid   = dest_nid;
4501                 msg->msg_hdr.dest_pid   = dest_pid;
4502                 msg->msg_hdr.payload_length = payload_length;
4503         }
4504
4505         lnet_net_lock(cpt);
4506         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4507         if (IS_ERR(lpni)) {
4508                 lnet_net_unlock(cpt);
4509                 CERROR("%s, src %s: Dropping %s "
4510                        "(error %ld looking up sender)\n",
4511                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4512                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4513                 lnet_msg_free(msg);
4514                 if (rc == -ESHUTDOWN)
4515                         /* We are shutting down.  Don't do anything more */
4516                         return 0;
4517                 goto drop;
4518         }
4519
4520         if (the_lnet.ln_routing)
4521                 lpni->lpni_last_alive = ktime_get_seconds();
4522
4523         msg->msg_rxpeer = lpni;
4524         msg->msg_rxni = ni;
4525         lnet_ni_addref_locked(ni, cpt);
4526         /* Multi-Rail: Primary NID of source. */
4527         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4528
4529         /*
4530          * mark the status of this lpni as UP since we received a message
4531          * from it. The ping response reports back the ns_status which is
4532          * marked on the remote as up or down and we cache it here.
4533          */
4534         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4535
4536         lnet_msg_commit(msg, cpt);
4537
4538         /* message delay simulation */
4539         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4540                      lnet_delay_rule_match_locked(hdr, msg))) {
4541                 lnet_net_unlock(cpt);
4542                 return 0;
4543         }
4544
4545         if (!for_me) {
4546                 rc = lnet_parse_forward_locked(ni, msg);
4547                 lnet_net_unlock(cpt);
4548
4549                 if (rc < 0)
4550                         goto free_drop;
4551
4552                 if (rc == LNET_CREDIT_OK) {
4553                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4554                                      0, payload_length, payload_length);
4555                 }
4556                 return 0;
4557         }
4558
4559         lnet_net_unlock(cpt);
4560
4561         rc = lnet_parse_local(ni, msg);
4562         if (rc != 0)
4563                 goto free_drop;
4564         return 0;
4565
4566  free_drop:
4567         LASSERT(msg->msg_md == NULL);
4568         lnet_finalize(msg, rc);
4569
4570  drop:
4571         lnet_drop_message(ni, cpt, private, payload_length, type);
4572         return 0;
4573 }
4574 EXPORT_SYMBOL(lnet_parse);
4575
4576 void
4577 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4578 {
4579         while (!list_empty(head)) {
4580                 struct lnet_process_id id = {0};
4581                 struct lnet_msg *msg;
4582
4583                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4584                 list_del(&msg->msg_list);
4585
4586                 id.nid = msg->msg_hdr.src_nid;
4587                 id.pid = msg->msg_hdr.src_pid;
4588
4589                 LASSERT(msg->msg_md == NULL);
4590                 LASSERT(msg->msg_rx_delayed);
4591                 LASSERT(msg->msg_rxpeer != NULL);
4592                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4593
4594                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4595                       " offset %d length %d: %s\n",
4596                       libcfs_id2str(id),
4597                       msg->msg_hdr.msg.put.ptl_index,
4598                       msg->msg_hdr.msg.put.match_bits,
4599                       msg->msg_hdr.msg.put.offset,
4600                       msg->msg_hdr.payload_length, reason);
4601
4602                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4603                  * called lnet_drop_message(), so I just hang onto msg as well
4604                  * until that's done */
4605
4606                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4607                                   msg->msg_private, msg->msg_len,
4608                                   msg->msg_type);
4609
4610                 msg->msg_no_resend = true;
4611                 /*
4612                  * NB: message will not generate event because w/o attached MD,
4613                  * but we still should give error code so lnet_msg_decommit()
4614                  * can skip counters operations and other checks.
4615                  */
4616                 lnet_finalize(msg, -ENOENT);
4617         }
4618 }
4619
4620 void
4621 lnet_recv_delayed_msg_list(struct list_head *head)
4622 {
4623         while (!list_empty(head)) {
4624                 struct lnet_msg *msg;
4625                 struct lnet_process_id id;
4626
4627                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4628                 list_del(&msg->msg_list);
4629
4630                 /* md won't disappear under me, since each msg
4631                  * holds a ref on it */
4632
4633                 id.nid = msg->msg_hdr.src_nid;
4634                 id.pid = msg->msg_hdr.src_pid;
4635
4636                 LASSERT(msg->msg_rx_delayed);
4637                 LASSERT(msg->msg_md != NULL);
4638                 LASSERT(msg->msg_rxpeer != NULL);
4639                 LASSERT(msg->msg_rxni != NULL);
4640                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4641
4642                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4643                        "match %llu offset %d length %d.\n",
4644                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4645                         msg->msg_hdr.msg.put.match_bits,
4646                         msg->msg_hdr.msg.put.offset,
4647                         msg->msg_hdr.payload_length);
4648
4649                 lnet_recv_put(msg->msg_rxni, msg);
4650         }
4651 }
4652
4653 static void
4654 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4655                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4656 {
4657         s64 timeout_ns;
4658         bool new_entry = true;
4659         struct lnet_rsp_tracker *local_rspt;
4660
4661         /*
4662          * MD has a refcount taken by message so it's not going away.
4663          * The MD however can be looked up. We need to secure the access
4664          * to the md_rspt_ptr by taking the res_lock.
4665          * The rspt can be accessed without protection up to when it gets
4666          * added to the list.
4667          */
4668
4669         lnet_res_lock(cpt);
4670         local_rspt = md->md_rspt_ptr;
4671         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4672         if (local_rspt != NULL) {
4673                 /*
4674                  * we already have an rspt attached to the md, so we'll
4675                  * update the deadline on that one.
4676                  */
4677                 LIBCFS_FREE(rspt, sizeof(*rspt));
4678                 new_entry = false;
4679         } else {
4680                 /* new md */
4681                 rspt->rspt_mdh = mdh;
4682                 rspt->rspt_cpt = cpt;
4683                 /* store the rspt so we can access it when we get the REPLY */
4684                 md->md_rspt_ptr = rspt;
4685                 local_rspt = rspt;
4686         }
4687         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4688
4689         /*
4690          * add to the list of tracked responses. It's added to tail of the
4691          * list in order to expire all the older entries first.
4692          */
4693         lnet_net_lock(cpt);
4694         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4695                 list_del_init(&local_rspt->rspt_on_list);
4696         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4697         lnet_net_unlock(cpt);
4698         lnet_res_unlock(cpt);
4699 }
4700
4701 /**
4702  * Initiate an asynchronous PUT operation.
4703  *
4704  * There are several events associated with a PUT: completion of the send on
4705  * the initiator node (LNET_EVENT_SEND), and when the send completes
4706  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4707  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4708  * used at the target node to indicate the completion of incoming data
4709  * delivery.
4710  *
4711  * The local events will be logged in the EQ associated with the MD pointed to
4712  * by \a mdh handle. Using a MD without an associated EQ results in these
4713  * events being discarded. In this case, the caller must have another
4714  * mechanism (e.g., a higher level protocol) for determining when it is safe
4715  * to modify the memory region associated with the MD.
4716  *
4717  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4718  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4719  *
4720  * \param self Indicates the NID of a local interface through which to send
4721  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4722  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4723  * must be "free floating" (See LNetMDBind()).
4724  * \param ack Controls whether an acknowledgment is requested.
4725  * Acknowledgments are only sent when they are requested by the initiating
4726  * process and the target MD enables them.
4727  * \param target A process identifier for the target process.
4728  * \param portal The index in the \a target's portal table.
4729  * \param match_bits The match bits to use for MD selection at the target
4730  * process.
4731  * \param offset The offset into the target MD (only used when the target
4732  * MD has the LNET_MD_MANAGE_REMOTE option set).
4733  * \param hdr_data 64 bits of user data that can be included in the message
4734  * header. This data is written to an event queue entry at the target if an
4735  * EQ is present on the matching MD.
4736  *
4737  * \retval  0      Success, and only in this case events will be generated
4738  * and logged to EQ (if it exists).
4739  * \retval -EIO    Simulated failure.
4740  * \retval -ENOMEM Memory allocation failure.
4741  * \retval -ENOENT Invalid MD object.
4742  *
4743  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4744  */
4745 int
4746 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4747         struct lnet_process_id target, unsigned int portal,
4748         __u64 match_bits, unsigned int offset,
4749         __u64 hdr_data)
4750 {
4751         struct lnet_msg *msg;
4752         struct lnet_libmd *md;
4753         int cpt;
4754         int rc;
4755         struct lnet_rsp_tracker *rspt = NULL;
4756
4757         LASSERT(the_lnet.ln_refcount > 0);
4758
4759         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4760             fail_peer(target.nid, 1)) {                 /* shall we now? */
4761                 CERROR("Dropping PUT to %s: simulated failure\n",
4762                        libcfs_id2str(target));
4763                 return -EIO;
4764         }
4765
4766         msg = lnet_msg_alloc();
4767         if (msg == NULL) {
4768                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4769                        libcfs_id2str(target));
4770                 return -ENOMEM;
4771         }
4772         msg->msg_vmflush = !!memory_pressure_get();
4773
4774         cpt = lnet_cpt_of_cookie(mdh.cookie);
4775
4776         if (ack == LNET_ACK_REQ) {
4777                 rspt = lnet_rspt_alloc(cpt);
4778                 if (!rspt) {
4779                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4780                                 libcfs_id2str(target));
4781                         return -ENOMEM;
4782                 }
4783                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4784         }
4785
4786         lnet_res_lock(cpt);
4787
4788         md = lnet_handle2md(&mdh);
4789         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4790                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4791                        match_bits, portal, libcfs_id2str(target),
4792                        md == NULL ? -1 : md->md_threshold);
4793                 if (md != NULL && md->md_me != NULL)
4794                         CERROR("Source MD also attached to portal %d\n",
4795                                md->md_me->me_portal);
4796                 lnet_res_unlock(cpt);
4797
4798                 LIBCFS_FREE(rspt, sizeof(*rspt));
4799                 lnet_msg_free(msg);
4800                 return -ENOENT;
4801         }
4802
4803         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4804
4805         lnet_msg_attach_md(msg, md, 0, 0);
4806
4807         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4808
4809         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4810         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4811         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4812         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4813
4814         /* NB handles only looked up by creator (no flips) */
4815         if (ack == LNET_ACK_REQ) {
4816                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4817                         the_lnet.ln_interface_cookie;
4818                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4819                         md->md_lh.lh_cookie;
4820         } else {
4821                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4822                         LNET_WIRE_HANDLE_COOKIE_NONE;
4823                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4824                         LNET_WIRE_HANDLE_COOKIE_NONE;
4825         }
4826
4827         lnet_res_unlock(cpt);
4828
4829         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4830
4831         if (ack == LNET_ACK_REQ)
4832                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4833
4834         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4835                                  CFS_FAIL_ONCE))
4836                 rc = -EIO;
4837         else
4838                 rc = lnet_send(self, msg, LNET_NID_ANY);
4839
4840         if (rc != 0) {
4841                 CNETERR("Error sending PUT to %s: %d\n",
4842                         libcfs_id2str(target), rc);
4843                 msg->msg_no_resend = true;
4844                 lnet_finalize(msg, rc);
4845         }
4846
4847         /* completion will be signalled by an event */
4848         return 0;
4849 }
4850 EXPORT_SYMBOL(LNetPut);
4851
4852 /*
4853  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4854  * returns a msg for the LND to pass to lnet_finalize() when the sink
4855  * data has been received.
4856  *
4857  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4858  * lnet_finalize() is called on it, so the LND must call this first
4859  */
4860 struct lnet_msg *
4861 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4862 {
4863         struct lnet_msg *msg = lnet_msg_alloc();
4864         struct lnet_libmd *getmd = getmsg->msg_md;
4865         struct lnet_process_id peer_id = getmsg->msg_target;
4866         int cpt;
4867
4868         LASSERT(!getmsg->msg_target_is_router);
4869         LASSERT(!getmsg->msg_routing);
4870
4871         if (msg == NULL) {
4872                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4873                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4874                 goto drop;
4875         }
4876
4877         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4878         lnet_res_lock(cpt);
4879
4880         LASSERT(getmd->md_refcount > 0);
4881
4882         if (getmd->md_threshold == 0) {
4883                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4884                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4885                         getmd);
4886                 lnet_res_unlock(cpt);
4887                 goto drop;
4888         }
4889
4890         LASSERT(getmd->md_offset == 0);
4891
4892         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4893                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4894
4895         /* setup information for lnet_build_msg_event */
4896         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4897         msg->msg_from = peer_id.nid;
4898         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4899         msg->msg_hdr.src_nid = peer_id.nid;
4900         msg->msg_hdr.payload_length = getmd->md_length;
4901         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4902
4903         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4904         lnet_res_unlock(cpt);
4905
4906         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4907
4908         lnet_net_lock(cpt);
4909         lnet_msg_commit(msg, cpt);
4910         lnet_net_unlock(cpt);
4911
4912         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4913
4914         return msg;
4915
4916  drop:
4917         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4918
4919         lnet_net_lock(cpt);
4920         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4921         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4922         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4923                 getmd->md_length;
4924         lnet_net_unlock(cpt);
4925
4926         if (msg != NULL)
4927                 lnet_msg_free(msg);
4928
4929         return NULL;
4930 }
4931 EXPORT_SYMBOL(lnet_create_reply_msg);
4932
4933 void
4934 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4935                        unsigned int len)
4936 {
4937         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4938          * completed and I know it. */
4939         LASSERT(reply != NULL);
4940         LASSERT(reply->msg_type == LNET_MSG_GET);
4941         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4942
4943         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4944          * the end of my buffer, I might as well be dead. */
4945         LASSERT(len <= reply->msg_ev.mlength);
4946
4947         reply->msg_ev.mlength = len;
4948 }
4949 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4950
4951 /**
4952  * Initiate an asynchronous GET operation.
4953  *
4954  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4955  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4956  * the target node in the REPLY has been written to local MD.
4957  *
4958  * On the target node, an LNET_EVENT_GET is logged when the GET request
4959  * arrives and is accepted into a MD.
4960  *
4961  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4962  * \param mdh A handle for the MD that describes the memory into which the
4963  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4964  *
4965  * \retval  0      Success, and only in this case events will be generated
4966  * and logged to EQ (if it exists) of the MD.
4967  * \retval -EIO    Simulated failure.
4968  * \retval -ENOMEM Memory allocation failure.
4969  * \retval -ENOENT Invalid MD object.
4970  */
4971 int
4972 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4973         struct lnet_process_id target, unsigned int portal,
4974         __u64 match_bits, unsigned int offset, bool recovery)
4975 {
4976         struct lnet_msg *msg;
4977         struct lnet_libmd *md;
4978         struct lnet_rsp_tracker *rspt;
4979         int cpt;
4980         int rc;
4981
4982         LASSERT(the_lnet.ln_refcount > 0);
4983
4984         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4985             fail_peer(target.nid, 1))                   /* shall we now? */
4986         {
4987                 CERROR("Dropping GET to %s: simulated failure\n",
4988                        libcfs_id2str(target));
4989                 return -EIO;
4990         }
4991
4992         msg = lnet_msg_alloc();
4993         if (!msg) {
4994                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4995                        libcfs_id2str(target));
4996                 return -ENOMEM;
4997         }
4998
4999         cpt = lnet_cpt_of_cookie(mdh.cookie);
5000
5001         rspt = lnet_rspt_alloc(cpt);
5002         if (!rspt) {
5003                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5004                        libcfs_id2str(target));
5005                 return -ENOMEM;
5006         }
5007         INIT_LIST_HEAD(&rspt->rspt_on_list);
5008
5009         msg->msg_recovery = recovery;
5010
5011         lnet_res_lock(cpt);
5012
5013         md = lnet_handle2md(&mdh);
5014         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5015                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5016                        match_bits, portal, libcfs_id2str(target),
5017                        md == NULL ? -1 : md->md_threshold);
5018                 if (md != NULL && md->md_me != NULL)
5019                         CERROR("REPLY MD also attached to portal %d\n",
5020                                md->md_me->me_portal);
5021
5022                 lnet_res_unlock(cpt);
5023
5024                 lnet_msg_free(msg);
5025                 LIBCFS_FREE(rspt, sizeof(*rspt));
5026                 return -ENOENT;
5027         }
5028
5029         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5030
5031         lnet_msg_attach_md(msg, md, 0, 0);
5032
5033         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5034
5035         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5036         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5037         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5038         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5039
5040         /* NB handles only looked up by creator (no flips) */
5041         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5042                 the_lnet.ln_interface_cookie;
5043         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5044                 md->md_lh.lh_cookie;
5045
5046         lnet_res_unlock(cpt);
5047
5048         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5049
5050         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5051
5052         rc = lnet_send(self, msg, LNET_NID_ANY);
5053         if (rc < 0) {
5054                 CNETERR("Error sending GET to %s: %d\n",
5055                         libcfs_id2str(target), rc);
5056                 msg->msg_no_resend = true;
5057                 lnet_finalize(msg, rc);
5058         }
5059
5060         /* completion will be signalled by an event */
5061         return 0;
5062 }
5063 EXPORT_SYMBOL(LNetGet);
5064
5065 /**
5066  * Calculate distance to node at \a dstnid.
5067  *
5068  * \param dstnid Target NID.
5069  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5070  * is saved here.
5071  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5072  * here.
5073  *
5074  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5075  * local_nid_dist_zero is set, which is the default.
5076  * \retval positives Distance to target NID, i.e. number of hops plus one.
5077  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5078  */
5079 int
5080 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5081 {
5082         struct list_head        *e;
5083         struct lnet_ni *ni = NULL;
5084         struct lnet_remotenet *rnet;
5085         __u32                   dstnet = LNET_NIDNET(dstnid);
5086         int                     hops;
5087         int                     cpt;
5088         __u32                   order = 2;
5089         struct list_head        *rn_list;
5090
5091         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5092          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5093          * keep order 0 free for 0@lo and order 1 free for a local NID
5094          * match */
5095
5096         LASSERT(the_lnet.ln_refcount > 0);
5097
5098         cpt = lnet_net_lock_current();
5099
5100         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5101                 if (ni->ni_nid == dstnid) {
5102                         if (srcnidp != NULL)
5103                                 *srcnidp = dstnid;
5104                         if (orderp != NULL) {
5105                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5106                                         *orderp = 0;
5107                                 else
5108                                         *orderp = 1;
5109                         }
5110                         lnet_net_unlock(cpt);
5111
5112                         return local_nid_dist_zero ? 0 : 1;
5113                 }
5114
5115                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5116                         /* Check if ni was originally created in
5117                          * current net namespace.
5118                          * If not, assign order above 0xffff0000,
5119                          * to make this ni not a priority. */
5120                         if (current->nsproxy &&
5121                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5122                                         order += 0xffff0000;
5123                         if (srcnidp != NULL)
5124                                 *srcnidp = ni->ni_nid;
5125                         if (orderp != NULL)
5126                                 *orderp = order;
5127                         lnet_net_unlock(cpt);
5128                         return 1;
5129                 }
5130
5131                 order++;
5132         }
5133
5134         rn_list = lnet_net2rnethash(dstnet);
5135         list_for_each(e, rn_list) {
5136                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5137
5138                 if (rnet->lrn_net == dstnet) {
5139                         struct lnet_route *route;
5140                         struct lnet_route *shortest = NULL;
5141                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5142                         __u32 route_hops;
5143
5144                         LASSERT(!list_empty(&rnet->lrn_routes));
5145
5146                         list_for_each_entry(route, &rnet->lrn_routes,
5147                                             lr_list) {
5148                                 route_hops = route->lr_hops;
5149                                 if (route_hops == LNET_UNDEFINED_HOPS)
5150                                         route_hops = 1;
5151                                 if (shortest == NULL ||
5152                                     route_hops < shortest_hops) {
5153                                         shortest = route;
5154                                         shortest_hops = route_hops;
5155                                 }
5156                         }
5157
5158                         LASSERT(shortest != NULL);
5159                         hops = shortest_hops;
5160                         if (srcnidp != NULL) {
5161                                 struct lnet_net *net;
5162                                 net = lnet_get_net_locked(shortest->lr_lnet);
5163                                 LASSERT(net);
5164                                 ni = lnet_get_next_ni_locked(net, NULL);
5165                                 *srcnidp = ni->ni_nid;
5166                         }
5167                         if (orderp != NULL)
5168                                 *orderp = order;
5169                         lnet_net_unlock(cpt);
5170                         return hops + 1;
5171                 }
5172                 order++;
5173         }
5174
5175         lnet_net_unlock(cpt);
5176         return -EHOSTUNREACH;
5177 }
5178 EXPORT_SYMBOL(LNetDist);