lnet/lnet/lib-move.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-move.c
  33  *
  34  * Data movement routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <linux/pagemap.h>
  40
  41 #include <lnet/lib-lnet.h>
  42 #include <linux/nsproxy.h>
  43 #include <net/net_namespace.h>
  44
  45 static int local_nid_dist_zero = 1;
  46 module_param(local_nid_dist_zero, int, 0444);
  47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  48
  49 struct lnet_send_data {
  50         struct lnet_ni *sd_best_ni;
  51         struct lnet_peer_ni *sd_best_lpni;
  52         struct lnet_peer_ni *sd_final_dst_lpni;
  53         struct lnet_peer *sd_peer;
  54         struct lnet_peer *sd_gw_peer;
  55         struct lnet_peer_ni *sd_gw_lpni;
  56         struct lnet_peer_net *sd_peer_net;
  57         struct lnet_msg *sd_msg;
  58         lnet_nid_t sd_dst_nid;
  59         lnet_nid_t sd_src_nid;
  60         lnet_nid_t sd_rtr_nid;
  61         int sd_cpt;
  62         int sd_md_cpt;
  63         __u32 sd_send_case;
  64 };
  65
  66 static inline struct lnet_comm_count *
  67 get_stats_counts(struct lnet_element_stats *stats,
  68                  enum lnet_stats_type stats_type)
  69 {
  70         switch (stats_type) {
  71         case LNET_STATS_TYPE_SEND:
  72                 return &stats->el_send_stats;
  73         case LNET_STATS_TYPE_RECV:
  74                 return &stats->el_recv_stats;
  75         case LNET_STATS_TYPE_DROP:
  76                 return &stats->el_drop_stats;
  77         default:
  78                 CERROR("Unknown stats type\n");
  79         }
  80
  81         return NULL;
  82 }
  83
  84 void lnet_incr_stats(struct lnet_element_stats *stats,
  85                      enum lnet_msg_type msg_type,
  86                      enum lnet_stats_type stats_type)
  87 {
  88         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
  89         if (!counts)
  90                 return;
  91
  92         switch (msg_type) {
  93         case LNET_MSG_ACK:
  94                 atomic_inc(&counts->co_ack_count);
  95                 break;
  96         case LNET_MSG_PUT:
  97                 atomic_inc(&counts->co_put_count);
  98                 break;
  99         case LNET_MSG_GET:
 100                 atomic_inc(&counts->co_get_count);
 101                 break;
 102         case LNET_MSG_REPLY:
 103                 atomic_inc(&counts->co_reply_count);
 104                 break;
 105         case LNET_MSG_HELLO:
 106                 atomic_inc(&counts->co_hello_count);
 107                 break;
 108         default:
 109                 CERROR("There is a BUG in the code. Unknown message type\n");
 110                 break;
 111         }
 112 }
 113
 114 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 115                      enum lnet_stats_type stats_type)
 116 {
 117         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 118         if (!counts)
 119                 return 0;
 120
 121         return (atomic_read(&counts->co_ack_count) +
 122                 atomic_read(&counts->co_put_count) +
 123                 atomic_read(&counts->co_get_count) +
 124                 atomic_read(&counts->co_reply_count) +
 125                 atomic_read(&counts->co_hello_count));
 126 }
 127
 128 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 129                                 struct lnet_comm_count *counts)
 130 {
 131         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 132         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 133         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 134         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 135         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 136 }
 137
 138 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 139                               struct lnet_element_stats *stats)
 140 {
 141         struct lnet_comm_count *counts;
 142
 143         LASSERT(msg_stats);
 144         LASSERT(stats);
 145
 146         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 147         if (!counts)
 148                 return;
 149         assign_stats(&msg_stats->im_send_stats, counts);
 150
 151         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 152         if (!counts)
 153                 return;
 154         assign_stats(&msg_stats->im_recv_stats, counts);
 155
 156         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 157         if (!counts)
 158                 return;
 159         assign_stats(&msg_stats->im_drop_stats, counts);
 160 }
 161
 162 int
 163 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 164 {
 165         struct lnet_test_peer *tp;
 166         struct list_head *el;
 167         struct list_head *next;
 168         struct list_head  cull;
 169
 170         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 171         if (threshold != 0) {
 172                 /* Adding a new entry */
 173                 LIBCFS_ALLOC(tp, sizeof(*tp));
 174                 if (tp == NULL)
 175                         return -ENOMEM;
 176
 177                 tp->tp_nid = nid;
 178                 tp->tp_threshold = threshold;
 179
 180                 lnet_net_lock(0);
 181                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 182                 lnet_net_unlock(0);
 183                 return 0;
 184         }
 185
 186         /* removing entries */
 187         INIT_LIST_HEAD(&cull);
 188
 189         lnet_net_lock(0);
 190
 191         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 192                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 193
 194                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 195                     nid == LNET_NID_ANY ||      /* removing all entries */
 196                     tp->tp_nid == nid) {        /* matched this one */
 197                         list_move(&tp->tp_list, &cull);
 198                 }
 199         }
 200
 201         lnet_net_unlock(0);
 202
 203         while (!list_empty(&cull)) {
 204                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 205
 206                 list_del(&tp->tp_list);
 207                 LIBCFS_FREE(tp, sizeof(*tp));
 208         }
 209         return 0;
 210 }
 211
 212 static int
 213 fail_peer (lnet_nid_t nid, int outgoing)
 214 {
 215         struct lnet_test_peer *tp;
 216         struct list_head *el;
 217         struct list_head *next;
 218         struct list_head  cull;
 219         int               fail = 0;
 220
 221         INIT_LIST_HEAD(&cull);
 222
 223         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 224         lnet_net_lock(0);
 225
 226         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 227                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 228
 229                 if (tp->tp_threshold == 0) {
 230                         /* zombie entry */
 231                         if (outgoing) {
 232                                 /* only cull zombies on outgoing tests,
 233                                  * since we may be at interrupt priority on
 234                                  * incoming messages. */
 235                                 list_move(&tp->tp_list, &cull);
 236                         }
 237                         continue;
 238                 }
 239
 240                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 241                     nid == tp->tp_nid) {                /* fail this peer */
 242                         fail = 1;
 243
 244                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 245                                 tp->tp_threshold--;
 246                                 if (outgoing &&
 247                                     tp->tp_threshold == 0) {
 248                                         /* see above */
 249                                         list_move(&tp->tp_list, &cull);
 250                                 }
 251                         }
 252                         break;
 253                 }
 254         }
 255
 256         lnet_net_unlock(0);
 257
 258         while (!list_empty(&cull)) {
 259                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 260                 list_del(&tp->tp_list);
 261
 262                 LIBCFS_FREE(tp, sizeof(*tp));
 263         }
 264
 265         return fail;
 266 }
 267
 268 unsigned int
 269 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 270 {
 271         unsigned int nob = 0;
 272
 273         LASSERT(niov == 0 || iov != NULL);
 274         while (niov-- > 0)
 275                 nob += (iov++)->iov_len;
 276
 277         return (nob);
 278 }
 279 EXPORT_SYMBOL(lnet_iov_nob);
 280
 281 void
 282 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 283                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 284                   unsigned int nob)
 285 {
 286         /* NB diov, siov are READ-ONLY */
 287         unsigned int  this_nob;
 288
 289         if (nob == 0)
 290                 return;
 291
 292         /* skip complete frags before 'doffset' */
 293         LASSERT(ndiov > 0);
 294         while (doffset >= diov->iov_len) {
 295                 doffset -= diov->iov_len;
 296                 diov++;
 297                 ndiov--;
 298                 LASSERT(ndiov > 0);
 299         }
 300
 301         /* skip complete frags before 'soffset' */
 302         LASSERT(nsiov > 0);
 303         while (soffset >= siov->iov_len) {
 304                 soffset -= siov->iov_len;
 305                 siov++;
 306                 nsiov--;
 307                 LASSERT(nsiov > 0);
 308         }
 309
 310         do {
 311                 LASSERT(ndiov > 0);
 312                 LASSERT(nsiov > 0);
 313                 this_nob = MIN(diov->iov_len - doffset,
 314                                siov->iov_len - soffset);
 315                 this_nob = MIN(this_nob, nob);
 316
 317                 memcpy((char *)diov->iov_base + doffset,
 318                        (char *)siov->iov_base + soffset, this_nob);
 319                 nob -= this_nob;
 320
 321                 if (diov->iov_len > doffset + this_nob) {
 322                         doffset += this_nob;
 323                 } else {
 324                         diov++;
 325                         ndiov--;
 326                         doffset = 0;
 327                 }
 328
 329                 if (siov->iov_len > soffset + this_nob) {
 330                         soffset += this_nob;
 331                 } else {
 332                         siov++;
 333                         nsiov--;
 334                         soffset = 0;
 335                 }
 336         } while (nob > 0);
 337 }
 338 EXPORT_SYMBOL(lnet_copy_iov2iov);
 339
 340 int
 341 lnet_extract_iov(int dst_niov, struct kvec *dst,
 342                  int src_niov, struct kvec *src,
 343                  unsigned int offset, unsigned int len)
 344 {
 345         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 346          * for exactly 'len' bytes, and return the number of entries.
 347          * NB not destructive to 'src' */
 348         unsigned int    frag_len;
 349         unsigned int    niov;
 350
 351         if (len == 0)                           /* no data => */
 352                 return (0);                     /* no frags */
 353
 354         LASSERT(src_niov > 0);
 355         while (offset >= src->iov_len) {      /* skip initial frags */
 356                 offset -= src->iov_len;
 357                 src_niov--;
 358                 src++;
 359                 LASSERT(src_niov > 0);
 360         }
 361
 362         niov = 1;
 363         for (;;) {
 364                 LASSERT(src_niov > 0);
 365                 LASSERT((int)niov <= dst_niov);
 366
 367                 frag_len = src->iov_len - offset;
 368                 dst->iov_base = ((char *)src->iov_base) + offset;
 369
 370                 if (len <= frag_len) {
 371                         dst->iov_len = len;
 372                         return (niov);
 373                 }
 374
 375                 dst->iov_len = frag_len;
 376
 377                 len -= frag_len;
 378                 dst++;
 379                 src++;
 380                 niov++;
 381                 src_niov--;
 382                 offset = 0;
 383         }
 384 }
 385 EXPORT_SYMBOL(lnet_extract_iov);
 386
 387
 388 unsigned int
 389 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 390 {
 391         unsigned int  nob = 0;
 392
 393         LASSERT(niov == 0 || kiov != NULL);
 394         while (niov-- > 0)
 395                 nob += (kiov++)->kiov_len;
 396
 397         return (nob);
 398 }
 399 EXPORT_SYMBOL(lnet_kiov_nob);
 400
 401 void
 402 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 403                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 404                     unsigned int nob)
 405 {
 406         /* NB diov, siov are READ-ONLY */
 407         unsigned int    this_nob;
 408         char           *daddr = NULL;
 409         char           *saddr = NULL;
 410
 411         if (nob == 0)
 412                 return;
 413
 414         LASSERT (!in_interrupt ());
 415
 416         LASSERT (ndiov > 0);
 417         while (doffset >= diov->kiov_len) {
 418                 doffset -= diov->kiov_len;
 419                 diov++;
 420                 ndiov--;
 421                 LASSERT(ndiov > 0);
 422         }
 423
 424         LASSERT(nsiov > 0);
 425         while (soffset >= siov->kiov_len) {
 426                 soffset -= siov->kiov_len;
 427                 siov++;
 428                 nsiov--;
 429                 LASSERT(nsiov > 0);
 430         }
 431
 432         do {
 433                 LASSERT(ndiov > 0);
 434                 LASSERT(nsiov > 0);
 435                 this_nob = MIN(diov->kiov_len - doffset,
 436                                siov->kiov_len - soffset);
 437                 this_nob = MIN(this_nob, nob);
 438
 439                 if (daddr == NULL)
 440                         daddr = ((char *)kmap(diov->kiov_page)) +
 441                                 diov->kiov_offset + doffset;
 442                 if (saddr == NULL)
 443                         saddr = ((char *)kmap(siov->kiov_page)) +
 444                                 siov->kiov_offset + soffset;
 445
 446                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 447                  * However in practice at least one of the kiovs will be mapped
 448                  * kernel pages and the map/unmap will be NOOPs */
 449
 450                 memcpy (daddr, saddr, this_nob);
 451                 nob -= this_nob;
 452
 453                 if (diov->kiov_len > doffset + this_nob) {
 454                         daddr += this_nob;
 455                         doffset += this_nob;
 456                 } else {
 457                         kunmap(diov->kiov_page);
 458                         daddr = NULL;
 459                         diov++;
 460                         ndiov--;
 461                         doffset = 0;
 462                 }
 463
 464                 if (siov->kiov_len > soffset + this_nob) {
 465                         saddr += this_nob;
 466                         soffset += this_nob;
 467                 } else {
 468                         kunmap(siov->kiov_page);
 469                         saddr = NULL;
 470                         siov++;
 471                         nsiov--;
 472                         soffset = 0;
 473                 }
 474         } while (nob > 0);
 475
 476         if (daddr != NULL)
 477                 kunmap(diov->kiov_page);
 478         if (saddr != NULL)
 479                 kunmap(siov->kiov_page);
 480 }
 481 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 482
 483 void
 484 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 485                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 486                     unsigned int nob)
 487 {
 488         /* NB iov, kiov are READ-ONLY */
 489         unsigned int    this_nob;
 490         char           *addr = NULL;
 491
 492         if (nob == 0)
 493                 return;
 494
 495         LASSERT (!in_interrupt ());
 496
 497         LASSERT (niov > 0);
 498         while (iovoffset >= iov->iov_len) {
 499                 iovoffset -= iov->iov_len;
 500                 iov++;
 501                 niov--;
 502                 LASSERT(niov > 0);
 503         }
 504
 505         LASSERT(nkiov > 0);
 506         while (kiovoffset >= kiov->kiov_len) {
 507                 kiovoffset -= kiov->kiov_len;
 508                 kiov++;
 509                 nkiov--;
 510                 LASSERT(nkiov > 0);
 511         }
 512
 513         do {
 514                 LASSERT(niov > 0);
 515                 LASSERT(nkiov > 0);
 516                 this_nob = MIN(iov->iov_len - iovoffset,
 517                                kiov->kiov_len - kiovoffset);
 518                 this_nob = MIN(this_nob, nob);
 519
 520                 if (addr == NULL)
 521                         addr = ((char *)kmap(kiov->kiov_page)) +
 522                                 kiov->kiov_offset + kiovoffset;
 523
 524                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 525                 nob -= this_nob;
 526
 527                 if (iov->iov_len > iovoffset + this_nob) {
 528                         iovoffset += this_nob;
 529                 } else {
 530                         iov++;
 531                         niov--;
 532                         iovoffset = 0;
 533                 }
 534
 535                 if (kiov->kiov_len > kiovoffset + this_nob) {
 536                         addr += this_nob;
 537                         kiovoffset += this_nob;
 538                 } else {
 539                         kunmap(kiov->kiov_page);
 540                         addr = NULL;
 541                         kiov++;
 542                         nkiov--;
 543                         kiovoffset = 0;
 544                 }
 545
 546         } while (nob > 0);
 547
 548         if (addr != NULL)
 549                 kunmap(kiov->kiov_page);
 550 }
 551 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 552
 553 void
 554 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 555                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 556                    unsigned int nob)
 557 {
 558         /* NB kiov, iov are READ-ONLY */
 559         unsigned int    this_nob;
 560         char           *addr = NULL;
 561
 562         if (nob == 0)
 563                 return;
 564
 565         LASSERT (!in_interrupt ());
 566
 567         LASSERT (nkiov > 0);
 568         while (kiovoffset >= kiov->kiov_len) {
 569                 kiovoffset -= kiov->kiov_len;
 570                 kiov++;
 571                 nkiov--;
 572                 LASSERT(nkiov > 0);
 573         }
 574
 575         LASSERT(niov > 0);
 576         while (iovoffset >= iov->iov_len) {
 577                 iovoffset -= iov->iov_len;
 578                 iov++;
 579                 niov--;
 580                 LASSERT(niov > 0);
 581         }
 582
 583         do {
 584                 LASSERT(nkiov > 0);
 585                 LASSERT(niov > 0);
 586                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 587                                iov->iov_len - iovoffset);
 588                 this_nob = MIN(this_nob, nob);
 589
 590                 if (addr == NULL)
 591                         addr = ((char *)kmap(kiov->kiov_page)) +
 592                                 kiov->kiov_offset + kiovoffset;
 593
 594                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 595                 nob -= this_nob;
 596
 597                 if (kiov->kiov_len > kiovoffset + this_nob) {
 598                         addr += this_nob;
 599                         kiovoffset += this_nob;
 600                 } else {
 601                         kunmap(kiov->kiov_page);
 602                         addr = NULL;
 603                         kiov++;
 604                         nkiov--;
 605                         kiovoffset = 0;
 606                 }
 607
 608                 if (iov->iov_len > iovoffset + this_nob) {
 609                         iovoffset += this_nob;
 610                 } else {
 611                         iov++;
 612                         niov--;
 613                         iovoffset = 0;
 614                 }
 615         } while (nob > 0);
 616
 617         if (addr != NULL)
 618                 kunmap(kiov->kiov_page);
 619 }
 620 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 621
 622 int
 623 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 624                   int src_niov, lnet_kiov_t *src,
 625                   unsigned int offset, unsigned int len)
 626 {
 627         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 628          * for exactly 'len' bytes, and return the number of entries.
 629          * NB not destructive to 'src' */
 630         unsigned int    frag_len;
 631         unsigned int    niov;
 632
 633         if (len == 0)                           /* no data => */
 634                 return (0);                     /* no frags */
 635
 636         LASSERT(src_niov > 0);
 637         while (offset >= src->kiov_len) {      /* skip initial frags */
 638                 offset -= src->kiov_len;
 639                 src_niov--;
 640                 src++;
 641                 LASSERT(src_niov > 0);
 642         }
 643
 644         niov = 1;
 645         for (;;) {
 646                 LASSERT(src_niov > 0);
 647                 LASSERT((int)niov <= dst_niov);
 648
 649                 frag_len = src->kiov_len - offset;
 650                 dst->kiov_page = src->kiov_page;
 651                 dst->kiov_offset = src->kiov_offset + offset;
 652
 653                 if (len <= frag_len) {
 654                         dst->kiov_len = len;
 655                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 656                         return niov;
 657                 }
 658
 659                 dst->kiov_len = frag_len;
 660                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 661
 662                 len -= frag_len;
 663                 dst++;
 664                 src++;
 665                 niov++;
 666                 src_niov--;
 667                 offset = 0;
 668         }
 669 }
 670 EXPORT_SYMBOL(lnet_extract_kiov);
 671
 672 void
 673 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 674              int delayed, unsigned int offset, unsigned int mlen,
 675              unsigned int rlen)
 676 {
 677         unsigned int  niov = 0;
 678         struct kvec *iov = NULL;
 679         lnet_kiov_t  *kiov = NULL;
 680         int           rc;
 681
 682         LASSERT (!in_interrupt ());
 683         LASSERT (mlen == 0 || msg != NULL);
 684
 685         if (msg != NULL) {
 686                 LASSERT(msg->msg_receiving);
 687                 LASSERT(!msg->msg_sending);
 688                 LASSERT(rlen == msg->msg_len);
 689                 LASSERT(mlen <= msg->msg_len);
 690                 LASSERT(msg->msg_offset == offset);
 691                 LASSERT(msg->msg_wanted == mlen);
 692
 693                 msg->msg_receiving = 0;
 694
 695                 if (mlen != 0) {
 696                         niov = msg->msg_niov;
 697                         iov  = msg->msg_iov;
 698                         kiov = msg->msg_kiov;
 699
 700                         LASSERT (niov > 0);
 701                         LASSERT ((iov == NULL) != (kiov == NULL));
 702                 }
 703         }
 704
 705         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 706                                              niov, iov, kiov, offset, mlen,
 707                                              rlen);
 708         if (rc < 0)
 709                 lnet_finalize(msg, rc);
 710 }
 711
 712 static void
 713 lnet_setpayloadbuffer(struct lnet_msg *msg)
 714 {
 715         struct lnet_libmd *md = msg->msg_md;
 716
 717         LASSERT(msg->msg_len > 0);
 718         LASSERT(!msg->msg_routing);
 719         LASSERT(md != NULL);
 720         LASSERT(msg->msg_niov == 0);
 721         LASSERT(msg->msg_iov == NULL);
 722         LASSERT(msg->msg_kiov == NULL);
 723
 724         msg->msg_niov = md->md_niov;
 725         if ((md->md_options & LNET_MD_KIOV) != 0)
 726                 msg->msg_kiov = md->md_iov.kiov;
 727         else
 728                 msg->msg_iov = md->md_iov.iov;
 729 }
 730
 731 void
 732 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 733                unsigned int offset, unsigned int len)
 734 {
 735         msg->msg_type = type;
 736         msg->msg_target = target;
 737         msg->msg_len = len;
 738         msg->msg_offset = offset;
 739
 740         if (len != 0)
 741                 lnet_setpayloadbuffer(msg);
 742
 743         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 744         msg->msg_hdr.type           = cpu_to_le32(type);
 745         /* dest_nid will be overwritten by lnet_select_pathway() */
 746         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 747         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 748         /* src_nid will be set later */
 749         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 750         msg->msg_hdr.payload_length = cpu_to_le32(len);
 751 }
 752
 753 static void
 754 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 755 {
 756         void   *priv = msg->msg_private;
 757         int rc;
 758
 759         LASSERT (!in_interrupt ());
 760         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 761                  (msg->msg_txcredit && msg->msg_peertxcredit));
 762
 763         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 764         if (rc < 0) {
 765                 msg->msg_no_resend = true;
 766                 lnet_finalize(msg, rc);
 767         }
 768 }
 769
 770 static int
 771 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 772 {
 773         int     rc;
 774
 775         LASSERT(!msg->msg_sending);
 776         LASSERT(msg->msg_receiving);
 777         LASSERT(!msg->msg_rx_ready_delay);
 778         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 779
 780         msg->msg_rx_ready_delay = 1;
 781         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 782                                                   &msg->msg_private);
 783         if (rc != 0) {
 784                 CERROR("recv from %s / send to %s aborted: "
 785                        "eager_recv failed %d\n",
 786                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 787                        libcfs_id2str(msg->msg_target), rc);
 788                 LASSERT(rc < 0); /* required by my callers */
 789         }
 790
 791         return rc;
 792 }
 793
 794 static bool
 795 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 796 {
 797         time64_t deadline;
 798
 799         deadline = lpni->lpni_last_alive +
 800                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 801
 802         /*
 803          * assume peer_ni is alive as long as we're within the configured
 804          * peer timeout
 805          */
 806         if (deadline > now)
 807                 return false;
 808
 809         return true;
 810 }
 811
 812 /* NB: returns 1 when alive, 0 when dead, negative when error;
 813  *     may drop the lnet_net_lock */
 814 static int
 815 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 816                        struct lnet_msg *msg)
 817 {
 818         time64_t now = ktime_get_seconds();
 819
 820         if (!lnet_peer_aliveness_enabled(lpni))
 821                 return -ENODEV;
 822
 823         /*
 824          * If we're resending a message, let's attempt to send it even if
 825          * the peer is down to fulfill our resend quota on the message
 826          */
 827         if (msg->msg_retry_count > 0)
 828                 return 1;
 829
 830         /* try and send recovery messages irregardless */
 831         if (msg->msg_recovery)
 832                 return 1;
 833
 834         /* always send any responses */
 835         if (msg->msg_type == LNET_MSG_ACK ||
 836             msg->msg_type == LNET_MSG_REPLY)
 837                 return 1;
 838
 839         if (!lnet_is_peer_deadline_passed(lpni, now))
 840                 return true;
 841
 842         return lnet_is_peer_ni_alive(lpni);
 843 }
 844
 845 /**
 846  * \param msg The message to be sent.
 847  * \param do_send True if lnet_ni_send() should be called in this function.
 848  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 849  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 850  *
 851  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 852  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 853  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 854  * \retval -ECANCELED If the MD of the message has been unlinked.
 855  */
 856 static int
 857 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 858 {
 859         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 860         struct lnet_ni          *ni = msg->msg_txni;
 861         int                     cpt = msg->msg_tx_cpt;
 862         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 863
 864         /* non-lnet_send() callers have checked before */
 865         LASSERT(!do_send || msg->msg_tx_delayed);
 866         LASSERT(!msg->msg_receiving);
 867         LASSERT(msg->msg_tx_committed);
 868         /* can't get here if we're sending to the loopback interface */
 869         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 870
 871         /* NB 'lp' is always the next hop */
 872         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 873             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 874                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 875                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 876                         msg->msg_len;
 877                 lnet_net_unlock(cpt);
 878                 if (msg->msg_txpeer)
 879                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 880                                         msg->msg_type,
 881                                         LNET_STATS_TYPE_DROP);
 882                 if (msg->msg_txni)
 883                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 884                                         msg->msg_type,
 885                                         LNET_STATS_TYPE_DROP);
 886
 887                 CNETERR("Dropping message for %s: peer not alive\n",
 888                         libcfs_id2str(msg->msg_target));
 889                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 890                 if (do_send)
 891                         lnet_finalize(msg, -EHOSTUNREACH);
 892
 893                 lnet_net_lock(cpt);
 894                 return -EHOSTUNREACH;
 895         }
 896
 897         if (msg->msg_md != NULL &&
 898             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 899                 lnet_net_unlock(cpt);
 900
 901                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 902                         "called on the MD/ME.\n",
 903                         libcfs_id2str(msg->msg_target));
 904                 if (do_send) {
 905                         msg->msg_no_resend = true;
 906                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 907                                msg, libcfs_id2str(msg->msg_target));
 908                         lnet_finalize(msg, -ECANCELED);
 909                 }
 910
 911                 lnet_net_lock(cpt);
 912                 return -ECANCELED;
 913         }
 914
 915         if (!msg->msg_peertxcredit) {
 916                 spin_lock(&lp->lpni_lock);
 917                 LASSERT((lp->lpni_txcredits < 0) ==
 918                         !list_empty(&lp->lpni_txq));
 919
 920                 msg->msg_peertxcredit = 1;
 921                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 922                 lp->lpni_txcredits--;
 923
 924                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 925                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 926
 927                 if (lp->lpni_txcredits < 0) {
 928                         msg->msg_tx_delayed = 1;
 929                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 930                         spin_unlock(&lp->lpni_lock);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933                 spin_unlock(&lp->lpni_lock);
 934         }
 935
 936         if (!msg->msg_txcredit) {
 937                 LASSERT((tq->tq_credits < 0) ==
 938                         !list_empty(&tq->tq_delayed));
 939
 940                 msg->msg_txcredit = 1;
 941                 tq->tq_credits--;
 942                 atomic_dec(&ni->ni_tx_credits);
 943
 944                 if (tq->tq_credits < tq->tq_credits_min)
 945                         tq->tq_credits_min = tq->tq_credits;
 946
 947                 if (tq->tq_credits < 0) {
 948                         msg->msg_tx_delayed = 1;
 949                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 950                         return LNET_CREDIT_WAIT;
 951                 }
 952         }
 953
 954         /* unset the tx_delay flag as we're going to send it now */
 955         msg->msg_tx_delayed = 0;
 956
 957         if (do_send) {
 958                 lnet_net_unlock(cpt);
 959                 lnet_ni_send(ni, msg);
 960                 lnet_net_lock(cpt);
 961         }
 962         return LNET_CREDIT_OK;
 963 }
 964
 965
 966 static struct lnet_rtrbufpool *
 967 lnet_msg2bufpool(struct lnet_msg *msg)
 968 {
 969         struct lnet_rtrbufpool  *rbp;
 970         int                     cpt;
 971
 972         LASSERT(msg->msg_rx_committed);
 973
 974         cpt = msg->msg_rx_cpt;
 975         rbp = &the_lnet.ln_rtrpools[cpt][0];
 976
 977         LASSERT(msg->msg_len <= LNET_MTU);
 978         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 979                 rbp++;
 980                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 981         }
 982
 983         return rbp;
 984 }
 985
 986 static int
 987 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 988 {
 989         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 990          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 991          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 992          * received or OK to receive */
 993         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 994         struct lnet_peer *lp;
 995         struct lnet_rtrbufpool *rbp;
 996         struct lnet_rtrbuf *rb;
 997
 998         LASSERT(msg->msg_iov == NULL);
 999         LASSERT(msg->msg_kiov == NULL);
1000         LASSERT(msg->msg_niov == 0);
1001         LASSERT(msg->msg_routing);
1002         LASSERT(msg->msg_receiving);
1003         LASSERT(!msg->msg_sending);
1004         LASSERT(lpni->lpni_peer_net);
1005         LASSERT(lpni->lpni_peer_net->lpn_peer);
1006
1007         lp = lpni->lpni_peer_net->lpn_peer;
1008
1009         /* non-lnet_parse callers only receive delayed messages */
1010         LASSERT(!do_recv || msg->msg_rx_delayed);
1011
1012         if (!msg->msg_peerrtrcredit) {
1013                 /* lpni_lock protects the credit manipulation */
1014                 spin_lock(&lpni->lpni_lock);
1015                 /* lp_lock protects the lp_rtrq */
1016                 spin_lock(&lp->lp_lock);
1017
1018                 msg->msg_peerrtrcredit = 1;
1019                 lpni->lpni_rtrcredits--;
1020                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1021                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1022
1023                 if (lpni->lpni_rtrcredits < 0) {
1024                         /* must have checked eager_recv before here */
1025                         LASSERT(msg->msg_rx_ready_delay);
1026                         msg->msg_rx_delayed = 1;
1027                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1028                         spin_unlock(&lp->lp_lock);
1029                         spin_unlock(&lpni->lpni_lock);
1030                         return LNET_CREDIT_WAIT;
1031                 }
1032                 spin_unlock(&lp->lp_lock);
1033                 spin_unlock(&lpni->lpni_lock);
1034         }
1035
1036         rbp = lnet_msg2bufpool(msg);
1037
1038         if (!msg->msg_rtrcredit) {
1039                 msg->msg_rtrcredit = 1;
1040                 rbp->rbp_credits--;
1041                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1042                         rbp->rbp_mincredits = rbp->rbp_credits;
1043
1044                 if (rbp->rbp_credits < 0) {
1045                         /* must have checked eager_recv before here */
1046                         LASSERT(msg->msg_rx_ready_delay);
1047                         msg->msg_rx_delayed = 1;
1048                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1049                         return LNET_CREDIT_WAIT;
1050                 }
1051         }
1052
1053         LASSERT(!list_empty(&rbp->rbp_bufs));
1054         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1055         list_del(&rb->rb_list);
1056
1057         msg->msg_niov = rbp->rbp_npages;
1058         msg->msg_kiov = &rb->rb_kiov[0];
1059
1060         /* unset the msg-rx_delayed flag since we're receiving the message */
1061         msg->msg_rx_delayed = 0;
1062
1063         if (do_recv) {
1064                 int cpt = msg->msg_rx_cpt;
1065
1066                 lnet_net_unlock(cpt);
1067                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1068                              0, msg->msg_len, msg->msg_len);
1069                 lnet_net_lock(cpt);
1070         }
1071         return LNET_CREDIT_OK;
1072 }
1073
1074 void
1075 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1076 {
1077         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1078         struct lnet_ni          *txni = msg->msg_txni;
1079         struct lnet_msg         *msg2;
1080
1081         if (msg->msg_txcredit) {
1082                 struct lnet_ni       *ni = msg->msg_txni;
1083                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1084
1085                 /* give back NI txcredits */
1086                 msg->msg_txcredit = 0;
1087
1088                 LASSERT((tq->tq_credits < 0) ==
1089                         !list_empty(&tq->tq_delayed));
1090
1091                 tq->tq_credits++;
1092                 atomic_inc(&ni->ni_tx_credits);
1093                 if (tq->tq_credits <= 0) {
1094                         msg2 = list_entry(tq->tq_delayed.next,
1095                                           struct lnet_msg, msg_list);
1096                         list_del(&msg2->msg_list);
1097
1098                         LASSERT(msg2->msg_txni == ni);
1099                         LASSERT(msg2->msg_tx_delayed);
1100                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1101
1102                         (void) lnet_post_send_locked(msg2, 1);
1103                 }
1104         }
1105
1106         if (msg->msg_peertxcredit) {
1107                 /* give back peer txcredits */
1108                 msg->msg_peertxcredit = 0;
1109
1110                 spin_lock(&txpeer->lpni_lock);
1111                 LASSERT((txpeer->lpni_txcredits < 0) ==
1112                         !list_empty(&txpeer->lpni_txq));
1113
1114                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1115                 LASSERT(txpeer->lpni_txqnob >= 0);
1116
1117                 txpeer->lpni_txcredits++;
1118                 if (txpeer->lpni_txcredits <= 0) {
1119                         int msg2_cpt;
1120
1121                         msg2 = list_entry(txpeer->lpni_txq.next,
1122                                               struct lnet_msg, msg_list);
1123                         list_del(&msg2->msg_list);
1124                         spin_unlock(&txpeer->lpni_lock);
1125
1126                         LASSERT(msg2->msg_txpeer == txpeer);
1127                         LASSERT(msg2->msg_tx_delayed);
1128
1129                         msg2_cpt = msg2->msg_tx_cpt;
1130
1131                         /*
1132                          * The msg_cpt can be different from the msg2_cpt
1133                          * so we need to make sure we lock the correct cpt
1134                          * for msg2.
1135                          * Once we call lnet_post_send_locked() it is no
1136                          * longer safe to access msg2, since it could've
1137                          * been freed by lnet_finalize(), but we still
1138                          * need to relock the correct cpt, so we cache the
1139                          * msg2_cpt for the purpose of the check that
1140                          * follows the call to lnet_pose_send_locked().
1141                          */
1142                         if (msg2_cpt != msg->msg_tx_cpt) {
1143                                 lnet_net_unlock(msg->msg_tx_cpt);
1144                                 lnet_net_lock(msg2_cpt);
1145                         }
1146                         (void) lnet_post_send_locked(msg2, 1);
1147                         if (msg2_cpt != msg->msg_tx_cpt) {
1148                                 lnet_net_unlock(msg2_cpt);
1149                                 lnet_net_lock(msg->msg_tx_cpt);
1150                         }
1151                 } else {
1152                         spin_unlock(&txpeer->lpni_lock);
1153                 }
1154         }
1155
1156         if (txni != NULL) {
1157                 msg->msg_txni = NULL;
1158                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1159         }
1160
1161         if (txpeer != NULL) {
1162                 msg->msg_txpeer = NULL;
1163                 lnet_peer_ni_decref_locked(txpeer);
1164         }
1165 }
1166
1167 void
1168 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1169 {
1170         struct lnet_msg *msg;
1171
1172         if (list_empty(&rbp->rbp_msgs))
1173                 return;
1174         msg = list_entry(rbp->rbp_msgs.next,
1175                          struct lnet_msg, msg_list);
1176         list_del(&msg->msg_list);
1177
1178         (void)lnet_post_routed_recv_locked(msg, 1);
1179 }
1180
1181 void
1182 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1183 {
1184         struct lnet_msg *msg;
1185         struct lnet_msg *tmp;
1186
1187         lnet_net_unlock(cpt);
1188
1189         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1190                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1191                              0, 0, 0, msg->msg_hdr.payload_length);
1192                 list_del_init(&msg->msg_list);
1193                 msg->msg_no_resend = true;
1194                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1195                 lnet_finalize(msg, -ECANCELED);
1196         }
1197
1198         lnet_net_lock(cpt);
1199 }
1200
1201 void
1202 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1203 {
1204         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1205         struct lnet_peer *lp;
1206         struct lnet_ni *rxni = msg->msg_rxni;
1207         struct lnet_msg *msg2;
1208
1209         if (msg->msg_rtrcredit) {
1210                 /* give back global router credits */
1211                 struct lnet_rtrbuf *rb;
1212                 struct lnet_rtrbufpool *rbp;
1213
1214                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1215                  * there until it gets one allocated, or aborts the wait
1216                  * itself */
1217                 LASSERT(msg->msg_kiov != NULL);
1218
1219                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1220                 rbp = rb->rb_pool;
1221
1222                 msg->msg_kiov = NULL;
1223                 msg->msg_rtrcredit = 0;
1224
1225                 LASSERT(rbp == lnet_msg2bufpool(msg));
1226
1227                 LASSERT((rbp->rbp_credits > 0) ==
1228                         !list_empty(&rbp->rbp_bufs));
1229
1230                 /* If routing is now turned off, we just drop this buffer and
1231                  * don't bother trying to return credits.  */
1232                 if (!the_lnet.ln_routing) {
1233                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1234                         goto routing_off;
1235                 }
1236
1237                 /* It is possible that a user has lowered the desired number of
1238                  * buffers in this pool.  Make sure we never put back
1239                  * more buffers than the stated number. */
1240                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1241                         /* Discard this buffer so we don't have too
1242                          * many. */
1243                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1244                         rbp->rbp_nbuffers--;
1245                 } else {
1246                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1247                         rbp->rbp_credits++;
1248                         if (rbp->rbp_credits <= 0)
1249                                 lnet_schedule_blocked_locked(rbp);
1250                 }
1251         }
1252
1253 routing_off:
1254         if (msg->msg_peerrtrcredit) {
1255                 LASSERT(rxpeerni);
1256                 LASSERT(rxpeerni->lpni_peer_net);
1257                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1258
1259                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1260
1261                 /* give back peer router credits */
1262                 msg->msg_peerrtrcredit = 0;
1263
1264                 spin_lock(&rxpeerni->lpni_lock);
1265                 spin_lock(&lp->lp_lock);
1266
1267                 rxpeerni->lpni_rtrcredits++;
1268
1269                 /* drop all messages which are queued to be routed on that
1270                  * peer. */
1271                 if (!the_lnet.ln_routing) {
1272                         struct list_head drop;
1273                         INIT_LIST_HEAD(&drop);
1274                         list_splice_init(&lp->lp_rtrq, &drop);
1275                         spin_unlock(&lp->lp_lock);
1276                         spin_unlock(&rxpeerni->lpni_lock);
1277                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1278                 } else if (!list_empty(&lp->lp_rtrq)) {
1279                         int msg2_cpt;
1280
1281                         msg2 = list_entry(lp->lp_rtrq.next,
1282                                           struct lnet_msg, msg_list);
1283                         list_del(&msg2->msg_list);
1284                         msg2_cpt = msg2->msg_rx_cpt;
1285                         spin_unlock(&lp->lp_lock);
1286                         spin_unlock(&rxpeerni->lpni_lock);
1287                         /*
1288                          * messages on the lp_rtrq can be from any NID in
1289                          * the peer, which means they might have different
1290                          * cpts. We need to make sure we lock the right
1291                          * one.
1292                          */
1293                         if (msg2_cpt != msg->msg_rx_cpt) {
1294                                 lnet_net_unlock(msg->msg_rx_cpt);
1295                                 lnet_net_lock(msg2_cpt);
1296                         }
1297                         (void) lnet_post_routed_recv_locked(msg2, 1);
1298                         if (msg2_cpt != msg->msg_rx_cpt) {
1299                                 lnet_net_unlock(msg2_cpt);
1300                                 lnet_net_lock(msg->msg_rx_cpt);
1301                         }
1302                 } else {
1303                         spin_unlock(&lp->lp_lock);
1304                         spin_unlock(&rxpeerni->lpni_lock);
1305                 }
1306         }
1307         if (rxni != NULL) {
1308                 msg->msg_rxni = NULL;
1309                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1310         }
1311         if (rxpeerni != NULL) {
1312                 msg->msg_rxpeer = NULL;
1313                 lnet_peer_ni_decref_locked(rxpeerni);
1314         }
1315 }
1316
1317 static int
1318 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1319 {
1320         if (p1->lpni_txqnob < p2->lpni_txqnob)
1321                 return 1;
1322
1323         if (p1->lpni_txqnob > p2->lpni_txqnob)
1324                 return -1;
1325
1326         if (p1->lpni_txcredits > p2->lpni_txcredits)
1327                 return 1;
1328
1329         if (p1->lpni_txcredits < p2->lpni_txcredits)
1330                 return -1;
1331
1332         return 0;
1333 }
1334
1335 static struct lnet_peer_ni *
1336 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1337                     struct lnet_peer *peer,
1338                     struct lnet_peer_net *peer_net)
1339 {
1340         /*
1341          * Look at the peer NIs for the destination peer that connect
1342          * to the chosen net. If a peer_ni is preferred when using the
1343          * best_ni to communicate, we use that one. If there is no
1344          * preferred peer_ni, or there are multiple preferred peer_ni,
1345          * the available transmit credits are used. If the transmit
1346          * credits are equal, we round-robin over the peer_ni.
1347          */
1348         struct lnet_peer_ni *lpni = NULL;
1349         struct lnet_peer_ni *best_lpni = NULL;
1350         int best_lpni_credits = INT_MIN;
1351         bool preferred = false;
1352         bool ni_is_pref;
1353         int best_lpni_healthv = 0;
1354         int lpni_healthv;
1355
1356         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1357                 /*
1358                  * if the best_ni we've chosen aleady has this lpni
1359                  * preferred, then let's use it
1360                  */
1361                 if (best_ni) {
1362                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1363                                                                 best_ni->ni_nid);
1364                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1365                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1366                 } else {
1367                         ni_is_pref = false;
1368                 }
1369
1370                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1371
1372                 if (best_lpni)
1373                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1374                                 libcfs_nid2str(lpni->lpni_nid),
1375                                 lpni->lpni_txcredits, best_lpni_credits,
1376                                 lpni->lpni_seq, best_lpni->lpni_seq);
1377
1378                 /* pick the healthiest peer ni */
1379                 if (lpni_healthv < best_lpni_healthv) {
1380                         continue;
1381                 } else if (lpni_healthv > best_lpni_healthv) {
1382                         best_lpni_healthv = lpni_healthv;
1383                 /* if this is a preferred peer use it */
1384                 } else if (!preferred && ni_is_pref) {
1385                         preferred = true;
1386                 } else if (preferred && !ni_is_pref) {
1387                         /*
1388                          * this is not the preferred peer so let's ignore
1389                          * it.
1390                          */
1391                         continue;
1392                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1393                         /*
1394                          * We already have a peer that has more credits
1395                          * available than this one. No need to consider
1396                          * this peer further.
1397                          */
1398                         continue;
1399                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1400                         /*
1401                          * The best peer found so far and the current peer
1402                          * have the same number of available credits let's
1403                          * make sure to select between them using Round
1404                          * Robin
1405                          */
1406                         if (best_lpni) {
1407                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1408                                         continue;
1409                         }
1410                 }
1411
1412                 best_lpni = lpni;
1413                 best_lpni_credits = lpni->lpni_txcredits;
1414         }
1415
1416         /* if we still can't find a peer ni then we can't reach it */
1417         if (!best_lpni) {
1418                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1419                         LNET_NIDNET(dst_nid);
1420                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1421                                 libcfs_net2str(net_id));
1422                 return NULL;
1423         }
1424
1425         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1426                libcfs_nid2str(best_lpni->lpni_nid));
1427
1428         return best_lpni;
1429 }
1430
1431 /*
1432  * Prerequisite: the best_ni should already be set in the sd
1433  */
1434 static inline struct lnet_peer_ni *
1435 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1436                            struct lnet_peer *peer, __u32 net_id)
1437 {
1438         struct lnet_peer_net *peer_net;
1439
1440         /*
1441          * The gateway is Multi-Rail capable so now we must select the
1442          * proper peer_ni
1443          */
1444         peer_net = lnet_peer_get_net_locked(peer, net_id);
1445
1446         if (!peer_net) {
1447                 CERROR("gateway peer %s has no NI on net %s\n",
1448                        libcfs_nid2str(peer->lp_primary_nid),
1449                        libcfs_net2str(net_id));
1450                 return NULL;
1451         }
1452
1453         return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1454 }
1455
1456 static int
1457 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1458                     struct lnet_peer_ni **best_lpni)
1459 {
1460         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1461         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1462         struct lnet_peer *lp1 = r1->lr_gateway;
1463         struct lnet_peer *lp2 = r2->lr_gateway;
1464         struct lnet_peer_ni *lpni1;
1465         struct lnet_peer_ni *lpni2;
1466         int rc;
1467
1468         lpni1 = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, lp1,
1469                                            r1->lr_lnet);
1470         lpni2 = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY, lp2,
1471                                            r2->lr_lnet);
1472         LASSERT(lpni1 && lpni2);
1473
1474         if (r1->lr_priority < r2->lr_priority) {
1475                 *best_lpni = lpni1;
1476                 return 1;
1477         }
1478
1479         if (r1->lr_priority > r2->lr_priority) {
1480                 *best_lpni = lpni2;
1481                 return -1;
1482         }
1483
1484         if (r1_hops < r2_hops) {
1485                 *best_lpni = lpni1;
1486                 return 1;
1487         }
1488
1489         if (r1_hops > r2_hops) {
1490                 *best_lpni = lpni2;
1491                 return -1;
1492         }
1493
1494         rc = lnet_compare_peers(lpni1, lpni2);
1495         if (rc == 1) {
1496                 *best_lpni = lpni1;
1497                 return rc;
1498         } else if (rc == -1) {
1499                 *best_lpni = lpni2;
1500                 return rc;
1501         }
1502
1503         if (r1->lr_seq - r2->lr_seq <= 0) {
1504                 *best_lpni = lpni1;
1505                 return 1;
1506         }
1507
1508         *best_lpni = lpni2;
1509         return -1;
1510 }
1511
1512 static struct lnet_route *
1513 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1514                        struct lnet_route **prev_route,
1515                        struct lnet_peer_ni **gwni)
1516 {
1517         struct lnet_peer_ni *best_gw_ni = NULL;
1518         struct lnet_route *best_route;
1519         struct lnet_route *last_route;
1520         struct lnet_remotenet *rnet;
1521         struct lnet_peer *lp_best;
1522         struct lnet_route *route;
1523         struct lnet_peer *lp;
1524         int rc;
1525
1526         rnet = lnet_find_rnet_locked(remote_net);
1527         if (rnet == NULL)
1528                 return NULL;
1529
1530         lp_best = NULL;
1531         best_route = last_route = NULL;
1532         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1533                 lp = route->lr_gateway;
1534
1535                 if (!lnet_is_route_alive(route))
1536                         continue;
1537
1538                 if (lp_best == NULL) {
1539                         best_route = last_route = route;
1540                         lp_best = lp;
1541                         best_gw_ni = lnet_find_best_lpni_on_net(NULL,
1542                                                                 LNET_NID_ANY,
1543                                                                 route->lr_gateway,
1544                                                                 route->lr_lnet);
1545                         LASSERT(best_gw_ni);
1546                         continue;
1547                 }
1548
1549                 /* no protection on below fields, but it's harmless */
1550                 if (last_route->lr_seq - route->lr_seq < 0)
1551                         last_route = route;
1552
1553                 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1554                 if (rc < 0)
1555                         continue;
1556
1557                 best_route = route;
1558                 lp_best = lp;
1559         }
1560
1561         *prev_route = last_route;
1562         *gwni = best_gw_ni;
1563
1564         return best_route;
1565 }
1566
1567 static struct lnet_ni *
1568 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1569                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1570                  int md_cpt)
1571 {
1572         struct lnet_ni *ni = NULL;
1573         unsigned int shortest_distance;
1574         int best_credits;
1575         int best_healthv;
1576
1577         /*
1578          * If there is no peer_ni that we can send to on this network,
1579          * then there is no point in looking for a new best_ni here.
1580         */
1581         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1582                 return best_ni;
1583
1584         if (best_ni == NULL) {
1585                 shortest_distance = UINT_MAX;
1586                 best_credits = INT_MIN;
1587                 best_healthv = 0;
1588         } else {
1589                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1590                                                      best_ni->ni_dev_cpt);
1591                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1592                 best_healthv = atomic_read(&best_ni->ni_healthv);
1593         }
1594
1595         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1596                 unsigned int distance;
1597                 int ni_credits;
1598                 int ni_healthv;
1599                 int ni_fatal;
1600
1601                 ni_credits = atomic_read(&ni->ni_tx_credits);
1602                 ni_healthv = atomic_read(&ni->ni_healthv);
1603                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1604
1605                 /*
1606                  * calculate the distance from the CPT on which
1607                  * the message memory is allocated to the CPT of
1608                  * the NI's physical device
1609                  */
1610                 distance = cfs_cpt_distance(lnet_cpt_table(),
1611                                             md_cpt,
1612                                             ni->ni_dev_cpt);
1613
1614                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1615                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1616                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1617                         : "not seleced", best_credits, shortest_distance,
1618                         (best_ni) ? best_ni->ni_seq : 0);
1619
1620                 /*
1621                  * All distances smaller than the NUMA range
1622                  * are treated equally.
1623                  */
1624                 if (distance < lnet_numa_range)
1625                         distance = lnet_numa_range;
1626
1627                 /*
1628                  * Select on health, shorter distance, available
1629                  * credits, then round-robin.
1630                  */
1631                 if (ni_fatal) {
1632                         continue;
1633                 } else if (ni_healthv < best_healthv) {
1634                         continue;
1635                 } else if (ni_healthv > best_healthv) {
1636                         best_healthv = ni_healthv;
1637                         /*
1638                          * If we're going to prefer this ni because it's
1639                          * the healthiest, then we should set the
1640                          * shortest_distance in the algorithm in case
1641                          * there are multiple NIs with the same health but
1642                          * different distances.
1643                          */
1644                         if (distance < shortest_distance)
1645                                 shortest_distance = distance;
1646                 } else if (distance > shortest_distance) {
1647                         continue;
1648                 } else if (distance < shortest_distance) {
1649                         shortest_distance = distance;
1650                 } else if (ni_credits < best_credits) {
1651                         continue;
1652                 } else if (ni_credits == best_credits) {
1653                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1654                                 continue;
1655                 }
1656                 best_ni = ni;
1657                 best_credits = ni_credits;
1658         }
1659
1660         CDEBUG(D_NET, "selected best_ni %s\n",
1661                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1662
1663         return best_ni;
1664 }
1665
1666 /*
1667  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1668  * because such traffic is required to perform discovery. We therefore
1669  * exclude all GET and PUT on that portal. We also exclude all ACK and
1670  * REPLY traffic, but that is because the portal is not tracked in the
1671  * message structure for these message types. We could restrict this
1672  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1673  */
1674 static bool
1675 lnet_msg_discovery(struct lnet_msg *msg)
1676 {
1677         if (msg->msg_type == LNET_MSG_PUT) {
1678                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1679                         return true;
1680         } else if (msg->msg_type == LNET_MSG_GET) {
1681                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1682                         return true;
1683         }
1684         return false;
1685 }
1686
1687 #define SRC_SPEC        0x0001
1688 #define SRC_ANY         0x0002
1689 #define LOCAL_DST       0x0004
1690 #define REMOTE_DST      0x0008
1691 #define MR_DST          0x0010
1692 #define NMR_DST         0x0020
1693 #define SND_RESP        0x0040
1694
1695 /* The following to defines are used for return codes */
1696 #define REPEAT_SEND     0x1000
1697 #define PASS_THROUGH    0x2000
1698
1699 /* The different cases lnet_select pathway needs to handle */
1700 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1701 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1702 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1703 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1704 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1705 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1706 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1707 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1708
1709 static int
1710 lnet_handle_lo_send(struct lnet_send_data *sd)
1711 {
1712         struct lnet_msg *msg = sd->sd_msg;
1713         int cpt = sd->sd_cpt;
1714
1715         /* No send credit hassles with LOLND */
1716         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1717         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1718         if (!msg->msg_routing)
1719                 msg->msg_hdr.src_nid =
1720                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1721         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1722         lnet_msg_commit(msg, cpt);
1723         msg->msg_txni = the_lnet.ln_loni;
1724
1725         return LNET_CREDIT_OK;
1726 }
1727
1728 static int
1729 lnet_handle_send(struct lnet_send_data *sd)
1730 {
1731         struct lnet_ni *best_ni = sd->sd_best_ni;
1732         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1733         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1734         struct lnet_msg *msg = sd->sd_msg;
1735         int cpt2;
1736         __u32 send_case = sd->sd_send_case;
1737         int rc;
1738         __u32 routing = send_case & REMOTE_DST;
1739          struct lnet_rsp_tracker *rspt;
1740
1741         /*
1742          * Increment sequence number of the selected peer so that we
1743          * pick the next one in Round Robin.
1744          */
1745         best_lpni->lpni_seq++;
1746
1747         /*
1748          * grab a reference on the peer_ni so it sticks around even if
1749          * we need to drop and relock the lnet_net_lock below.
1750          */
1751         lnet_peer_ni_addref_locked(best_lpni);
1752
1753         /*
1754          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1755          * message. This ensures that we get a CPT that is correct for
1756          * the NI when the NI has been restricted to a subset of all CPTs.
1757          * If the selected CPT differs from the one currently locked, we
1758          * must unlock and relock the lnet_net_lock(), and then check whether
1759          * the configuration has changed. We don't have a hold on the best_ni
1760          * yet, and it may have vanished.
1761          */
1762         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1763         if (sd->sd_cpt != cpt2) {
1764                 __u32 seq = lnet_get_dlc_seq_locked();
1765                 lnet_net_unlock(sd->sd_cpt);
1766                 sd->sd_cpt = cpt2;
1767                 lnet_net_lock(sd->sd_cpt);
1768                 if (seq != lnet_get_dlc_seq_locked()) {
1769                         lnet_peer_ni_decref_locked(best_lpni);
1770                         return REPEAT_SEND;
1771                 }
1772         }
1773
1774         /*
1775          * store the best_lpni in the message right away to avoid having
1776          * to do the same operation under different conditions
1777          */
1778         msg->msg_txpeer = best_lpni;
1779         msg->msg_txni = best_ni;
1780
1781         /*
1782          * grab a reference for the best_ni since now it's in use in this
1783          * send. The reference will be dropped in lnet_finalize()
1784          */
1785         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1786
1787         /*
1788          * Always set the target.nid to the best peer picked. Either the
1789          * NID will be one of the peer NIDs selected, or the same NID as
1790          * what was originally set in the target or it will be the NID of
1791          * a router if this message should be routed
1792          */
1793         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1794
1795         /*
1796          * lnet_msg_commit assigns the correct cpt to the message, which
1797          * is used to decrement the correct refcount on the ni when it's
1798          * time to return the credits
1799          */
1800         lnet_msg_commit(msg, sd->sd_cpt);
1801
1802         /*
1803          * If we are routing the message then we keep the src_nid that was
1804          * set by the originator. If we are not routing then we are the
1805          * originator and set it here.
1806          */
1807         if (!msg->msg_routing)
1808                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1809
1810         if (routing) {
1811                 msg->msg_target_is_router = 1;
1812                 msg->msg_target.pid = LNET_PID_LUSTRE;
1813                 /*
1814                  * since we're routing we want to ensure that the
1815                  * msg_hdr.dest_nid is set to the final destination. When
1816                  * the router receives this message it knows how to route
1817                  * it.
1818                  *
1819                  * final_dst_lpni is set at the beginning of the
1820                  * lnet_select_pathway() function and is never changed.
1821                  * It's safe to use it here.
1822                  */
1823                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1824         } else {
1825                 /*
1826                  * if we're not routing set the dest_nid to the best peer
1827                  * ni NID that we picked earlier in the algorithm.
1828                  */
1829                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1830         }
1831
1832         /*
1833          * if we have response tracker block update it with the next hop
1834          * nid
1835          */
1836         if (msg->msg_md) {
1837                 rspt = msg->msg_md->md_rspt_ptr;
1838                 if (rspt) {
1839                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1840                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1841                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1842                 }
1843         }
1844
1845         rc = lnet_post_send_locked(msg, 0);
1846
1847         if (!rc)
1848                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1849                        libcfs_nid2str(msg->msg_hdr.src_nid),
1850                        libcfs_nid2str(msg->msg_txni->ni_nid),
1851                        libcfs_nid2str(sd->sd_src_nid),
1852                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1853                        libcfs_nid2str(sd->sd_dst_nid),
1854                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1855                        libcfs_nid2str(sd->sd_rtr_nid),
1856                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1857
1858         return rc;
1859 }
1860
1861 static inline void
1862 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1863 {
1864         if (sd->sd_send_case & NMR_DST &&
1865             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1866             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1867             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1868                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1869                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1870                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1871                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1872                                                  sd->sd_best_ni->ni_nid);
1873         }
1874 }
1875
1876 /*
1877  * Source Specified
1878  * Local Destination
1879  * non-mr peer
1880  *
1881  * use the source and destination NIDs as the pathway
1882  */
1883 static int
1884 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1885 {
1886         /* the destination lpni is set before we get here. */
1887
1888         /* find local NI */
1889         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1890         if (!sd->sd_best_ni) {
1891                 CERROR("Can't send to %s: src %s is not a "
1892                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1893                                 libcfs_nid2str(sd->sd_src_nid));
1894                 return -EINVAL;
1895         }
1896
1897         /*
1898          * the preferred NID will only be set for NMR peers
1899          */
1900         lnet_set_non_mr_pref_nid(sd);
1901
1902         return lnet_handle_send(sd);
1903 }
1904
1905 /*
1906  * Source Specified
1907  * Local Destination
1908  * MR Peer
1909  *
1910  * Don't run the selection algorithm on the peer NIs. By specifying the
1911  * local NID, we're also saying that we should always use the destination NID
1912  * provided. This handles the case where we should be using the same
1913  * destination NID for the all the messages which belong to the same RPC
1914  * request.
1915  */
1916 static int
1917 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1918 {
1919         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1920         if (!sd->sd_best_ni) {
1921                 CERROR("Can't send to %s: src %s is not a "
1922                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1923                                 libcfs_nid2str(sd->sd_src_nid));
1924                 return -EINVAL;
1925         }
1926
1927         if (sd->sd_best_lpni &&
1928             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1929                 return lnet_handle_lo_send(sd);
1930         else if (sd->sd_best_lpni)
1931                 return lnet_handle_send(sd);
1932
1933         CERROR("can't send to %s. no NI on %s\n",
1934                libcfs_nid2str(sd->sd_dst_nid),
1935                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1936
1937         return -EHOSTUNREACH;
1938 }
1939
1940 struct lnet_ni *
1941 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1942                               struct lnet_peer *peer,
1943                               struct lnet_peer_net *peer_net,
1944                               int cpt,
1945                               bool incr_seq)
1946 {
1947         struct lnet_net *local_net;
1948         struct lnet_ni *best_ni;
1949
1950         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1951         if (!local_net)
1952                 return NULL;
1953
1954         /*
1955          * Iterate through the NIs in this local Net and select
1956          * the NI to send from. The selection is determined by
1957          * these 3 criterion in the following priority:
1958          *      1. NUMA
1959          *      2. NI available credits
1960          *      3. Round Robin
1961          */
1962         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1963                                    peer, peer_net, cpt);
1964
1965         if (incr_seq && best_ni)
1966                 best_ni->ni_seq++;
1967
1968         return best_ni;
1969 }
1970
1971 static int
1972 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1973                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1974                              int cpt)
1975 {
1976         struct lnet_peer *peer;
1977         lnet_nid_t primary_nid;
1978         int rc;
1979
1980         lnet_peer_ni_addref_locked(lpni);
1981
1982         peer = lpni->lpni_peer_net->lpn_peer;
1983
1984         if (lnet_peer_gw_discovery(peer)) {
1985                 lnet_peer_ni_decref_locked(lpni);
1986                 return 0;
1987         }
1988
1989         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1990                 lnet_peer_ni_decref_locked(lpni);
1991                 return 0;
1992         }
1993
1994         rc = lnet_discover_peer_locked(lpni, cpt, false);
1995         if (rc) {
1996                 lnet_peer_ni_decref_locked(lpni);
1997                 return rc;
1998         }
1999         /* The peer may have changed. */
2000         peer = lpni->lpni_peer_net->lpn_peer;
2001         spin_lock(&peer->lp_lock);
2002         if (lnet_peer_is_uptodate_locked(peer)) {
2003                 spin_unlock(&peer->lp_lock);
2004                 lnet_peer_ni_decref_locked(lpni);
2005                 return 0;
2006         }
2007         /* queue message and return */
2008         msg->msg_rtr_nid_param = rtr_nid;
2009         msg->msg_sending = 0;
2010         msg->msg_txpeer = NULL;
2011         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2012         primary_nid = peer->lp_primary_nid;
2013         spin_unlock(&peer->lp_lock);
2014
2015         lnet_peer_ni_decref_locked(lpni);
2016
2017         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2018                 msg, libcfs_nid2str(primary_nid));
2019
2020         return LNET_DC_WAIT;
2021 }
2022
2023 static int
2024 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2025                              lnet_nid_t dst_nid,
2026                              struct lnet_peer_ni **gw_lpni,
2027                              struct lnet_peer **gw_peer)
2028 {
2029         int rc;
2030         __u32 local_lnet;
2031         struct lnet_peer *gw;
2032         struct lnet_peer *lp;
2033         struct lnet_peer_net *lpn;
2034         struct lnet_peer_net *best_lpn = NULL;
2035         struct lnet_remotenet *rnet;
2036         struct lnet_route *best_route = NULL;
2037         struct lnet_route *last_route = NULL;
2038         struct lnet_peer_ni *lpni = NULL;
2039         struct lnet_peer_ni *gwni = NULL;
2040         lnet_nid_t src_nid = sd->sd_src_nid;
2041
2042         /* If a router nid was specified then we are replying to a GET or
2043          * sending an ACK. In this case we use the gateway associated with the
2044          * specified router nid.
2045          */
2046         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2047                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2048                 if (!gwni) {
2049                         CERROR("No peer NI for gateway %s\n",
2050                                libcfs_nid2str(sd->sd_rtr_nid));
2051                         return -EHOSTUNREACH;
2052                 }
2053                 gw = gwni->lpni_peer_net->lpn_peer;
2054                 lnet_peer_ni_decref_locked(gwni);
2055                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2056         } else {
2057                 /* we've already looked up the initial lpni using dst_nid */
2058                 lpni = sd->sd_best_lpni;
2059                 /* the peer tree must be in existence */
2060                 LASSERT(lpni && lpni->lpni_peer_net &&
2061                         lpni->lpni_peer_net->lpn_peer);
2062                 lp = lpni->lpni_peer_net->lpn_peer;
2063
2064                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2065                         /* is this remote network reachable?  */
2066                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2067                         if (!rnet)
2068                                 continue;
2069
2070                         if (!best_lpn)
2071                                 best_lpn = lpn;
2072
2073                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2074                                 continue;
2075
2076                         best_lpn = lpn;
2077                 }
2078
2079                 if (!best_lpn) {
2080                         CERROR("peer %s has no available nets\n",
2081                                libcfs_nid2str(sd->sd_dst_nid));
2082                         return -EHOSTUNREACH;
2083                 }
2084
2085                 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2086                                                               sd->sd_dst_nid,
2087                                                               lp,
2088                                                               best_lpn->lpn_net_id);
2089                 if (!sd->sd_best_lpni) {
2090                         CERROR("peer %s down\n",
2091                                libcfs_nid2str(sd->sd_dst_nid));
2092                         return -EHOSTUNREACH;
2093                 }
2094
2095                 best_route = lnet_find_route_locked(NULL, best_lpn->lpn_net_id,
2096                                                     &last_route, &gwni);
2097                 if (!best_route) {
2098                         CERROR("no route to %s from %s\n",
2099                                libcfs_nid2str(dst_nid),
2100                                libcfs_nid2str(src_nid));
2101                         return -EHOSTUNREACH;
2102                 }
2103
2104                 if (!gwni) {
2105                         CERROR("Internal Error. Route expected to %s from %s\n",
2106                                libcfs_nid2str(dst_nid),
2107                                libcfs_nid2str(src_nid));
2108                         return -EFAULT;
2109                 }
2110
2111                 gw = best_route->lr_gateway;
2112                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2113                 local_lnet = best_route->lr_lnet;
2114
2115         }
2116
2117         /*
2118          * Discover this gateway if it hasn't already been discovered.
2119          * This means we might delay the message until discovery has
2120          * completed
2121          */
2122         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2123         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2124                                           sd->sd_cpt);
2125         if (rc)
2126                 return rc;
2127
2128         if (!sd->sd_best_ni)
2129                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2130                                         lnet_peer_get_net_locked(gw,
2131                                                                  local_lnet),
2132                                         sd->sd_md_cpt,
2133                                         true);
2134
2135         if (!sd->sd_best_ni) {
2136                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2137                        libcfs_net2str(local_lnet),
2138                        libcfs_nid2str(sd->sd_src_nid));
2139                 return -EFAULT;
2140         }
2141
2142         *gw_lpni = gwni;
2143         *gw_peer = gw;
2144
2145         /*
2146          * increment the sequence numbers since now we're sure we're
2147          * going to use this path
2148          */
2149         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2150                 LASSERT(best_route && last_route);
2151                 best_route->lr_seq = last_route->lr_seq + 1;
2152                 best_lpn->lpn_seq++;
2153         }
2154
2155         return 0;
2156 }
2157
2158 /*
2159  * Handle two cases:
2160  *
2161  * Case 1:
2162  *  Source specified
2163  *  Remote destination
2164  *  Non-MR destination
2165  *
2166  * Case 2:
2167  *  Source specified
2168  *  Remote destination
2169  *  MR destination
2170  *
2171  * The handling of these two cases is similar. Even though the destination
2172  * can be MR or non-MR, we'll deal directly with the router.
2173  */
2174 static int
2175 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2176 {
2177         int rc;
2178         struct lnet_peer_ni *gw_lpni = NULL;
2179         struct lnet_peer *gw_peer = NULL;
2180
2181         /* find local NI */
2182         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2183         if (!sd->sd_best_ni) {
2184                 CERROR("Can't send to %s: src %s is not a "
2185                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2186                                 libcfs_nid2str(sd->sd_src_nid));
2187                 return -EINVAL;
2188         }
2189
2190         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2191                                      &gw_peer);
2192         if (rc)
2193                 return rc;
2194
2195         if (sd->sd_send_case & NMR_DST)
2196                 /*
2197                 * since the final destination is non-MR let's set its preferred
2198                 * NID before we send
2199                 */
2200                 lnet_set_non_mr_pref_nid(sd);
2201
2202         /*
2203          * We're going to send to the gw found so let's set its
2204          * info
2205          */
2206         sd->sd_peer = gw_peer;
2207         sd->sd_best_lpni = gw_lpni;
2208
2209         return lnet_handle_send(sd);
2210 }
2211
2212 struct lnet_ni *
2213 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2214                                bool discovery)
2215 {
2216         struct lnet_peer_net *peer_net = NULL;
2217         struct lnet_ni *best_ni = NULL;
2218
2219         /*
2220          * The peer can have multiple interfaces, some of them can be on
2221          * the local network and others on a routed network. We should
2222          * prefer the local network. However if the local network is not
2223          * available then we need to try the routed network
2224          */
2225
2226         /* go through all the peer nets and find the best_ni */
2227         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2228                 /*
2229                  * The peer's list of nets can contain non-local nets. We
2230                  * want to only examine the local ones.
2231                  */
2232                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2233                         continue;
2234                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2235                                                    peer_net, md_cpt, false);
2236
2237                 /*
2238                  * if this is a discovery message and lp_disc_net_id is
2239                  * specified then use that net to send the discovery on.
2240                  */
2241                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2242                     discovery)
2243                         break;
2244         }
2245
2246         if (best_ni)
2247                 /* increment sequence number so we can round robin */
2248                 best_ni->ni_seq++;
2249
2250         return best_ni;
2251 }
2252
2253 static struct lnet_ni *
2254 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2255 {
2256         struct lnet_ni *best_ni = NULL;
2257         struct lnet_peer_net *peer_net;
2258         struct lnet_peer *peer = sd->sd_peer;
2259         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2260         struct lnet_peer_ni *lpni;
2261         int cpt = sd->sd_cpt;
2262
2263         /*
2264          * We must use a consistent source address when sending to a
2265          * non-MR peer. However, a non-MR peer can have multiple NIDs
2266          * on multiple networks, and we may even need to talk to this
2267          * peer on multiple networks -- certain types of
2268          * load-balancing configuration do this.
2269          *
2270          * So we need to pick the NI the peer prefers for this
2271          * particular network.
2272          */
2273
2274         /* Get the target peer_ni */
2275         peer_net = lnet_peer_get_net_locked(peer,
2276                         LNET_NIDNET(best_lpni->lpni_nid));
2277         LASSERT(peer_net != NULL);
2278         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2279                                 lpni_peer_nis) {
2280                 if (lpni->lpni_pref_nnids == 0)
2281                         continue;
2282                 LASSERT(lpni->lpni_pref_nnids == 1);
2283                 best_ni = lnet_nid2ni_locked(
2284                                 lpni->lpni_pref.nid, cpt);
2285                 break;
2286         }
2287
2288         return best_ni;
2289 }
2290
2291 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2292 static int
2293 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2294 {
2295         struct lnet_ni *best_ni = NULL;
2296         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2297
2298         /*
2299          * We must use a consistent source address when sending to a
2300          * non-MR peer. However, a non-MR peer can have multiple NIDs
2301          * on multiple networks, and we may even need to talk to this
2302          * peer on multiple networks -- certain types of
2303          * load-balancing configuration do this.
2304          *
2305          * So we need to pick the NI the peer prefers for this
2306          * particular network.
2307          */
2308
2309         best_ni = lnet_find_existing_preferred_best_ni(sd);
2310
2311         /* if best_ni is still not set just pick one */
2312         if (!best_ni) {
2313                 best_ni =
2314                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2315                                                 sd->sd_best_lpni->lpni_peer_net,
2316                                                 sd->sd_md_cpt, true);
2317                 /* If there is no best_ni we don't have a route */
2318                 if (!best_ni) {
2319                         CERROR("no path to %s from net %s\n",
2320                                 libcfs_nid2str(best_lpni->lpni_nid),
2321                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2322                         return -EHOSTUNREACH;
2323                 }
2324         }
2325
2326         sd->sd_best_ni = best_ni;
2327
2328         /* Set preferred NI if necessary. */
2329         lnet_set_non_mr_pref_nid(sd);
2330
2331         return 0;
2332 }
2333
2334
2335 /*
2336  * Source not specified
2337  * Local destination
2338  * Non-MR Peer
2339  *
2340  * always use the same source NID for NMR peers
2341  * If we've talked to that peer before then we already have a preferred
2342  * source NI associated with it. Otherwise, we select a preferred local NI
2343  * and store it in the peer
2344  */
2345 static int
2346 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2347 {
2348         int rc;
2349
2350         /* sd->sd_best_lpni is already set to the final destination */
2351
2352         /*
2353          * At this point we should've created the peer ni and peer. If we
2354          * can't find it, then something went wrong. Instead of assert
2355          * output a relevant message and fail the send
2356          */
2357         if (!sd->sd_best_lpni) {
2358                 CERROR("Internal fault. Unable to send msg %s to %s. "
2359                        "NID not known\n",
2360                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2361                        libcfs_nid2str(sd->sd_dst_nid));
2362                 return -EFAULT;
2363         }
2364
2365         rc = lnet_select_preferred_best_ni(sd);
2366         if (!rc)
2367                 rc = lnet_handle_send(sd);
2368
2369         return rc;
2370 }
2371
2372 static int
2373 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2374 {
2375         /*
2376          * NOTE we've already handled the remote peer case. So we only
2377          * need to worry about the local case here.
2378          *
2379          * if we're sending a response, ACK or reply, we need to send it
2380          * to the destination NID given to us. At this point we already
2381          * have the peer_ni we're suppose to send to, so just find the
2382          * best_ni on the peer net and use that. Since we're sending to an
2383          * MR peer then we can just run the selection algorithm on our
2384          * local NIs and pick the best one.
2385          */
2386         if (sd->sd_send_case & SND_RESP) {
2387                 sd->sd_best_ni =
2388                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2389                                                 sd->sd_best_lpni->lpni_peer_net,
2390                                                 sd->sd_md_cpt, true);
2391
2392                 if (!sd->sd_best_ni) {
2393                         /*
2394                          * We're not going to deal with not able to send
2395                          * a response to the provided final destination
2396                          */
2397                         CERROR("Can't send response to %s. "
2398                                "No local NI available\n",
2399                                 libcfs_nid2str(sd->sd_dst_nid));
2400                         return -EHOSTUNREACH;
2401                 }
2402
2403                 return lnet_handle_send(sd);
2404         }
2405
2406         /*
2407          * If we get here that means we're sending a fresh request, PUT or
2408          * GET, so we need to run our standard selection algorithm.
2409          * First find the best local interface that's on any of the peer's
2410          * networks.
2411          */
2412         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2413                                         sd->sd_md_cpt,
2414                                         lnet_msg_discovery(sd->sd_msg));
2415         if (sd->sd_best_ni) {
2416                 sd->sd_best_lpni =
2417                   lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2418                                              sd->sd_peer,
2419                                              sd->sd_best_ni->ni_net->net_id);
2420
2421                 /*
2422                  * if we're successful in selecting a peer_ni on the local
2423                  * network, then send to it. Otherwise fall through and
2424                  * try and see if we can reach it over another routed
2425                  * network
2426                  */
2427                 if (sd->sd_best_lpni &&
2428                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
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_lo_send(sd);
2436                 } else if (sd->sd_best_lpni) {
2437                         /*
2438                          * in case we initially started with a routed
2439                          * destination, let's reset to local
2440                          */
2441                         sd->sd_send_case &= ~REMOTE_DST;
2442                         sd->sd_send_case |= LOCAL_DST;
2443                         return lnet_handle_send(sd);
2444                 }
2445
2446                 CERROR("Internal Error. Expected to have a best_lpni: "
2447                        "%s -> %s\n",
2448                        libcfs_nid2str(sd->sd_src_nid),
2449                        libcfs_nid2str(sd->sd_dst_nid));
2450
2451                 return -EFAULT;
2452         }
2453
2454         /*
2455          * Peer doesn't have a local network. Let's see if there is
2456          * a remote network we can reach it on.
2457          */
2458         return PASS_THROUGH;
2459 }
2460
2461 /*
2462  * Case 1:
2463  *      Source NID not specified
2464  *      Local destination
2465  *      MR peer
2466  *
2467  * Case 2:
2468  *      Source NID not speified
2469  *      Remote destination
2470  *      MR peer
2471  *
2472  * In both of these cases if we're sending a response, ACK or REPLY, then
2473  * we need to send to the destination NID provided.
2474  *
2475  * In the remote case let's deal with MR routers.
2476  *
2477  */
2478
2479 static int
2480 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2481 {
2482         int rc = 0;
2483         struct lnet_peer *gw_peer = NULL;
2484         struct lnet_peer_ni *gw_lpni = NULL;
2485
2486         /*
2487          * handle sending a response to a remote peer here so we don't
2488          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2489          */
2490         if (sd->sd_send_case & REMOTE_DST &&
2491             sd->sd_send_case & SND_RESP) {
2492                 struct lnet_peer_ni *gw;
2493                 struct lnet_peer *gw_peer;
2494
2495                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2496                                                   &gw_peer);
2497                 if (rc < 0) {
2498                         CERROR("Can't send response to %s. "
2499                                "No route available\n",
2500                                 libcfs_nid2str(sd->sd_dst_nid));
2501                         return -EHOSTUNREACH;
2502                 } else if (rc > 0) {
2503                         return rc;
2504                 }
2505
2506                 sd->sd_best_lpni = gw;
2507                 sd->sd_peer = gw_peer;
2508
2509                 return lnet_handle_send(sd);
2510         }
2511
2512         /*
2513          * Even though the NID for the peer might not be on a local network,
2514          * since the peer is MR there could be other interfaces on the
2515          * local network. In that case we'd still like to prefer the local
2516          * network over the routed network. If we're unable to do that
2517          * then we select the best router among the different routed networks,
2518          * and if the router is MR then we can deal with it as such.
2519          */
2520         rc = lnet_handle_any_mr_dsta(sd);
2521         if (rc != PASS_THROUGH)
2522                 return rc;
2523
2524         /*
2525          * Now that we must route to the destination, we must consider the
2526          * MR case, where the destination has multiple interfaces, some of
2527          * which we can route to and others we do not. For this reason we
2528          * need to select the destination which we can route to and if
2529          * there are multiple, we need to round robin.
2530          */
2531         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2532                                           &gw_peer);
2533         if (rc)
2534                 return rc;
2535
2536         sd->sd_send_case &= ~LOCAL_DST;
2537         sd->sd_send_case |= REMOTE_DST;
2538
2539         sd->sd_peer = gw_peer;
2540         sd->sd_best_lpni = gw_lpni;
2541
2542         return lnet_handle_send(sd);
2543 }
2544
2545 /*
2546  * Source not specified
2547  * Remote destination
2548  * Non-MR peer
2549  *
2550  * Must send to the specified peer NID using the same source NID that
2551  * we've used before. If it's the first time to talk to that peer then
2552  * find the source NI and assign it as preferred to that peer
2553  */
2554 static int
2555 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2556 {
2557         int rc;
2558         struct lnet_peer_ni *gw_lpni = NULL;
2559         struct lnet_peer *gw_peer = NULL;
2560
2561         /*
2562          * Let's set if we have a preferred NI to talk to this NMR peer
2563          */
2564         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2565
2566         /*
2567          * find the router and that'll find the best NI if we didn't find
2568          * it already.
2569          */
2570         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2571                                           &gw_peer);
2572         if (rc)
2573                 return rc;
2574
2575         /*
2576          * set the best_ni we've chosen as the preferred one for
2577          * this peer
2578          */
2579         lnet_set_non_mr_pref_nid(sd);
2580
2581         /* we'll be sending to the gw */
2582         sd->sd_best_lpni = gw_lpni;
2583         sd->sd_peer = gw_peer;
2584
2585         return lnet_handle_send(sd);
2586 }
2587
2588 static int
2589 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2590 {
2591         /*
2592          * turn off the SND_RESP bit.
2593          * It will be checked in the case handling
2594          */
2595         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2596
2597         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2598                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2599                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2600                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2601                 libcfs_nid2str(sd->sd_dst_nid),
2602                 (send_case & LOCAL_DST) ? "local" : "routed");
2603
2604         switch (send_case) {
2605         /*
2606          * For all cases where the source is specified, we should always
2607          * use the destination NID, whether it's an MR destination or not,
2608          * since we're continuing a series of related messages for the
2609          * same RPC
2610          */
2611         case SRC_SPEC_LOCAL_NMR_DST:
2612                 return lnet_handle_spec_local_nmr_dst(sd);
2613         case SRC_SPEC_LOCAL_MR_DST:
2614                 return lnet_handle_spec_local_mr_dst(sd);
2615         case SRC_SPEC_ROUTER_NMR_DST:
2616         case SRC_SPEC_ROUTER_MR_DST:
2617                 return lnet_handle_spec_router_dst(sd);
2618         case SRC_ANY_LOCAL_NMR_DST:
2619                 return lnet_handle_any_local_nmr_dst(sd);
2620         case SRC_ANY_LOCAL_MR_DST:
2621         case SRC_ANY_ROUTER_MR_DST:
2622                 return lnet_handle_any_mr_dst(sd);
2623         case SRC_ANY_ROUTER_NMR_DST:
2624                 return lnet_handle_any_router_nmr_dst(sd);
2625         default:
2626                 CERROR("Unknown send case\n");
2627                 return -1;
2628         }
2629 }
2630
2631 static int
2632 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2633                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2634 {
2635         struct lnet_peer_ni     *lpni;
2636         struct lnet_peer        *peer;
2637         struct lnet_send_data   send_data;
2638         int                     cpt, rc;
2639         int                     md_cpt;
2640         __u32                   send_case = 0;
2641
2642         memset(&send_data, 0, sizeof(send_data));
2643
2644         /*
2645          * get an initial CPT to use for locking. The idea here is not to
2646          * serialize the calls to select_pathway, so that as many
2647          * operations can run concurrently as possible. To do that we use
2648          * the CPT where this call is being executed. Later on when we
2649          * determine the CPT to use in lnet_message_commit, we switch the
2650          * lock and check if there was any configuration change.  If none,
2651          * then we proceed, if there is, then we restart the operation.
2652          */
2653         cpt = lnet_net_lock_current();
2654
2655         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2656         if (md_cpt == CFS_CPT_ANY)
2657                 md_cpt = cpt;
2658
2659 again:
2660
2661         /*
2662          * If we're sending to ourselves then there is no need to go through
2663          * any selection. We can shortcut the entire process and send over
2664          * lolnd.
2665          *
2666          * However, we make two exceptions to this rule:
2667          * 1. If the src_nid is specified then our API defines that we must send
2668          *    via that interface.
2669          * 2. Recovery messages must be sent to the lnet_ni that is being
2670          *    recovered.
2671          */
2672         send_data.sd_msg = msg;
2673         send_data.sd_cpt = cpt;
2674         if (src_nid == LNET_NID_ANY && !msg->msg_recovery &&
2675             lnet_nid2ni_locked(dst_nid, cpt)) {
2676                 rc = lnet_handle_lo_send(&send_data);
2677                 lnet_net_unlock(cpt);
2678                 return rc;
2679         }
2680
2681         /*
2682          * find an existing peer_ni, or create one and mark it as having been
2683          * created due to network traffic. This call will create the
2684          * peer->peer_net->peer_ni tree.
2685          */
2686         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2687         if (IS_ERR(lpni)) {
2688                 lnet_net_unlock(cpt);
2689                 return PTR_ERR(lpni);
2690         }
2691
2692         /*
2693          * Cache the original src_nid. If we need to resend the message
2694          * then we'll need to know whether the src_nid was originally
2695          * specified for this message. If it was originally specified,
2696          * then we need to keep using the same src_nid since it's
2697          * continuing the same sequence of messages.
2698          */
2699         msg->msg_src_nid_param = src_nid;
2700
2701         /*
2702          * If necessary, perform discovery on the peer that owns this peer_ni.
2703          * Note, this can result in the ownership of this peer_ni changing
2704          * to another peer object.
2705          */
2706         rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2707         if (rc) {
2708                 lnet_peer_ni_decref_locked(lpni);
2709                 lnet_net_unlock(cpt);
2710                 return rc;
2711         }
2712         lnet_peer_ni_decref_locked(lpni);
2713
2714         peer = lpni->lpni_peer_net->lpn_peer;
2715
2716         /*
2717          * Identify the different send cases
2718          */
2719         if (src_nid == LNET_NID_ANY)
2720                 send_case |= SRC_ANY;
2721         else
2722                 send_case |= SRC_SPEC;
2723
2724         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2725                 send_case |= LOCAL_DST;
2726         else
2727                 send_case |= REMOTE_DST;
2728
2729         /*
2730          * if this is a non-MR peer or if we're recovering a peer ni then
2731          * let's consider this an NMR case so we can hit the destination
2732          * NID.
2733          */
2734         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2735                 send_case |= NMR_DST;
2736         else
2737                 send_case |= MR_DST;
2738
2739         if (msg->msg_type == LNET_MSG_REPLY ||
2740             msg->msg_type == LNET_MSG_ACK)
2741                 send_case |= SND_RESP;
2742
2743         /* assign parameters to the send_data */
2744         send_data.sd_rtr_nid = rtr_nid;
2745         send_data.sd_src_nid = src_nid;
2746         send_data.sd_dst_nid = dst_nid;
2747         send_data.sd_best_lpni = lpni;
2748         /*
2749          * keep a pointer to the final destination in case we're going to
2750          * route, so we'll need to access it later
2751          */
2752         send_data.sd_final_dst_lpni = lpni;
2753         send_data.sd_peer = peer;
2754         send_data.sd_md_cpt = md_cpt;
2755         send_data.sd_send_case = send_case;
2756
2757         rc = lnet_handle_send_case_locked(&send_data);
2758
2759         /*
2760          * Update the local cpt since send_data.sd_cpt might've been
2761          * updated as a result of calling lnet_handle_send_case_locked().
2762          */
2763         cpt = send_data.sd_cpt;
2764
2765         if (rc == REPEAT_SEND)
2766                 goto again;
2767
2768         lnet_net_unlock(cpt);
2769
2770         return rc;
2771 }
2772
2773 int
2774 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2775 {
2776         lnet_nid_t              dst_nid = msg->msg_target.nid;
2777         int                     rc;
2778
2779         /*
2780          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2781          * but we might want to use pre-determined router for ACK/REPLY
2782          * in the future
2783          */
2784         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2785         LASSERT(msg->msg_txpeer == NULL);
2786         LASSERT(msg->msg_txni == NULL);
2787         LASSERT(!msg->msg_sending);
2788         LASSERT(!msg->msg_target_is_router);
2789         LASSERT(!msg->msg_receiving);
2790
2791         msg->msg_sending = 1;
2792
2793         LASSERT(!msg->msg_tx_committed);
2794
2795         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2796         if (rc < 0) {
2797                 if (rc == -EHOSTUNREACH)
2798                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2799                 else
2800                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2801                 return rc;
2802         }
2803
2804         if (rc == LNET_CREDIT_OK)
2805                 lnet_ni_send(msg->msg_txni, msg);
2806
2807         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2808         return 0;
2809 }
2810
2811 enum lnet_mt_event_type {
2812         MT_TYPE_LOCAL_NI = 0,
2813         MT_TYPE_PEER_NI
2814 };
2815
2816 struct lnet_mt_event_info {
2817         enum lnet_mt_event_type mt_type;
2818         lnet_nid_t mt_nid;
2819 };
2820
2821 /* called with res_lock held */
2822 void
2823 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2824 {
2825         struct lnet_rsp_tracker *rspt;
2826
2827         /*
2828          * msg has a refcount on the MD so the MD is not going away.
2829          * The rspt queue for the cpt is protected by
2830          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2831          */
2832         if (!md->md_rspt_ptr)
2833                 return;
2834
2835         rspt = md->md_rspt_ptr;
2836
2837         /* debug code */
2838         LASSERT(rspt->rspt_cpt == cpt);
2839
2840         md->md_rspt_ptr = NULL;
2841
2842         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2843                 /*
2844                  * The monitor thread has invalidated this handle because the
2845                  * response timed out, but it failed to lookup the MD. That
2846                  * means this response tracker is on the zombie list. We can
2847                  * safely remove it under the resource lock (held by caller) and
2848                  * free the response tracker block.
2849                  */
2850                 list_del(&rspt->rspt_on_list);
2851                 lnet_rspt_free(rspt, cpt);
2852         } else {
2853                 /*
2854                  * invalidate the handle to indicate that a response has been
2855                  * received, which will then lead the monitor thread to clean up
2856                  * the rspt block.
2857                  */
2858                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2859         }
2860 }
2861
2862 void
2863 lnet_clean_zombie_rstqs(void)
2864 {
2865         struct lnet_rsp_tracker *rspt, *tmp;
2866         int i;
2867
2868         cfs_cpt_for_each(i, lnet_cpt_table()) {
2869                 list_for_each_entry_safe(rspt, tmp,
2870                                          the_lnet.ln_mt_zombie_rstqs[i],
2871                                          rspt_on_list) {
2872                         list_del(&rspt->rspt_on_list);
2873                         lnet_rspt_free(rspt, i);
2874                 }
2875         }
2876
2877         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2878 }
2879
2880 static void
2881 lnet_finalize_expired_responses(void)
2882 {
2883         struct lnet_libmd *md;
2884         struct list_head local_queue;
2885         struct lnet_rsp_tracker *rspt, *tmp;
2886         ktime_t now;
2887         int i;
2888
2889         if (the_lnet.ln_mt_rstq == NULL)
2890                 return;
2891
2892         cfs_cpt_for_each(i, lnet_cpt_table()) {
2893                 INIT_LIST_HEAD(&local_queue);
2894
2895                 lnet_net_lock(i);
2896                 if (!the_lnet.ln_mt_rstq[i]) {
2897                         lnet_net_unlock(i);
2898                         continue;
2899                 }
2900                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2901                 lnet_net_unlock(i);
2902
2903                 now = ktime_get();
2904
2905                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2906                         /*
2907                          * The rspt mdh will be invalidated when a response
2908                          * is received or whenever we want to discard the
2909                          * block the monitor thread will walk the queue
2910                          * and clean up any rsts with an invalid mdh.
2911                          * The monitor thread will walk the queue until
2912                          * the first unexpired rspt block. This means that
2913                          * some rspt blocks which received their
2914                          * corresponding responses will linger in the
2915                          * queue until they are cleaned up eventually.
2916                          */
2917                         lnet_res_lock(i);
2918                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2919                                 lnet_res_unlock(i);
2920                                 list_del(&rspt->rspt_on_list);
2921                                 lnet_rspt_free(rspt, i);
2922                                 continue;
2923                         }
2924
2925                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2926                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2927                                 struct lnet_peer_ni *lpni;
2928                                 lnet_nid_t nid;
2929
2930                                 md = lnet_handle2md(&rspt->rspt_mdh);
2931                                 if (!md) {
2932                                         /* MD has been queued for unlink, but
2933                                          * rspt hasn't been detached (Note we've
2934                                          * checked above that the rspt_mdh is
2935                                          * valid). Since we cannot lookup the MD
2936                                          * we're unable to detach the rspt
2937                                          * ourselves. Thus, move the rspt to the
2938                                          * zombie list where we'll wait for
2939                                          * either:
2940                                          *   1. The remaining operations on the
2941                                          *   MD to complete. In this case the
2942                                          *   final operation will result in
2943                                          *   lnet_msg_detach_md()->
2944                                          *   lnet_detach_rsp_tracker() where
2945                                          *   we will clean up this response
2946                                          *   tracker.
2947                                          *   2. LNet to shutdown. In this case
2948                                          *   we'll wait until after all LND Nets
2949                                          *   have shutdown and then we can
2950                                          *   safely free any remaining response
2951                                          *   tracker blocks on the zombie list.
2952                                          * Note: We need to hold the resource
2953                                          * lock when adding to the zombie list
2954                                          * because we may have concurrent access
2955                                          * with lnet_detach_rsp_tracker().
2956                                          */
2957                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2958                                         list_move(&rspt->rspt_on_list,
2959                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2960                                         lnet_res_unlock(i);
2961                                         continue;
2962                                 }
2963                                 LASSERT(md->md_rspt_ptr == rspt);
2964                                 md->md_rspt_ptr = NULL;
2965                                 lnet_res_unlock(i);
2966
2967                                 LNetMDUnlink(rspt->rspt_mdh);
2968
2969                                 nid = rspt->rspt_next_hop_nid;
2970
2971                                 list_del(&rspt->rspt_on_list);
2972                                 lnet_rspt_free(rspt, i);
2973
2974                                 /* If we're shutting down we just want to clean
2975                                  * up the rspt blocks
2976                                  */
2977                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2978                                         continue;
2979
2980                                 lnet_net_lock(i);
2981                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2982                                 lnet_net_unlock(i);
2983
2984                                 CDEBUG(D_NET,
2985                                        "Response timeout: md = %p: nid = %s\n",
2986                                        md, libcfs_nid2str(nid));
2987
2988                                 /*
2989                                  * If there is a timeout on the response
2990                                  * from the next hop decrement its health
2991                                  * value so that we don't use it
2992                                  */
2993                                 lnet_net_lock(0);
2994                                 lpni = lnet_find_peer_ni_locked(nid);
2995                                 if (lpni) {
2996                                         lnet_handle_remote_failure_locked(lpni);
2997                                         lnet_peer_ni_decref_locked(lpni);
2998                                 }
2999                                 lnet_net_unlock(0);
3000                         } else {
3001                                 lnet_res_unlock(i);
3002                                 break;
3003                         }
3004                 }
3005
3006                 if (!list_empty(&local_queue)) {
3007                         lnet_net_lock(i);
3008                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3009                         lnet_net_unlock(i);
3010                 }
3011         }
3012 }
3013
3014 static void
3015 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3016 {
3017         struct lnet_msg *msg;
3018
3019         while (!list_empty(resendq)) {
3020                 struct lnet_peer_ni *lpni;
3021
3022                 msg = list_entry(resendq->next, struct lnet_msg,
3023                                  msg_list);
3024
3025                 list_del_init(&msg->msg_list);
3026
3027                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3028                 if (!lpni) {
3029                         lnet_net_unlock(cpt);
3030                         CERROR("Expected that a peer is already created for %s\n",
3031                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3032                         msg->msg_no_resend = true;
3033                         lnet_finalize(msg, -EFAULT);
3034                         lnet_net_lock(cpt);
3035                 } else {
3036                         struct lnet_peer *peer;
3037                         int rc;
3038                         lnet_nid_t src_nid = LNET_NID_ANY;
3039
3040                         /*
3041                          * if this message is not being routed and the
3042                          * peer is non-MR then we must use the same
3043                          * src_nid that was used in the original send.
3044                          * Otherwise if we're routing the message (IE
3045                          * we're a router) then we can use any of our
3046                          * local interfaces. It doesn't matter to the
3047                          * final destination.
3048                          */
3049                         peer = lpni->lpni_peer_net->lpn_peer;
3050                         if (!msg->msg_routing &&
3051                             !lnet_peer_is_multi_rail(peer))
3052                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3053
3054                         /*
3055                          * If we originally specified a src NID, then we
3056                          * must attempt to reuse it in the resend as well.
3057                          */
3058                         if (msg->msg_src_nid_param != LNET_NID_ANY)
3059                                 src_nid = msg->msg_src_nid_param;
3060                         lnet_peer_ni_decref_locked(lpni);
3061
3062                         lnet_net_unlock(cpt);
3063                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3064                                libcfs_nid2str(src_nid),
3065                                libcfs_id2str(msg->msg_target),
3066                                lnet_msgtyp2str(msg->msg_type),
3067                                msg->msg_recovery,
3068                                msg->msg_retry_count);
3069                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3070                         if (rc) {
3071                                 CERROR("Error sending %s to %s: %d\n",
3072                                        lnet_msgtyp2str(msg->msg_type),
3073                                        libcfs_id2str(msg->msg_target), rc);
3074                                 msg->msg_no_resend = true;
3075                                 lnet_finalize(msg, rc);
3076                         }
3077                         lnet_net_lock(cpt);
3078                         if (!rc)
3079                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3080                 }
3081         }
3082 }
3083
3084 static void
3085 lnet_resend_pending_msgs(void)
3086 {
3087         int i;
3088
3089         cfs_cpt_for_each(i, lnet_cpt_table()) {
3090                 lnet_net_lock(i);
3091                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3092                 lnet_net_unlock(i);
3093         }
3094 }
3095
3096 /* called with cpt and ni_lock held */
3097 static void
3098 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3099 {
3100         struct lnet_handle_md recovery_mdh;
3101
3102         LNetInvalidateMDHandle(&recovery_mdh);
3103
3104         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3105             force) {
3106                 recovery_mdh = ni->ni_ping_mdh;
3107                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3108         }
3109         lnet_ni_unlock(ni);
3110         lnet_net_unlock(cpt);
3111         if (!LNetMDHandleIsInvalid(recovery_mdh))
3112                 LNetMDUnlink(recovery_mdh);
3113         lnet_net_lock(cpt);
3114         lnet_ni_lock(ni);
3115 }
3116
3117 static void
3118 lnet_recover_local_nis(void)
3119 {
3120         struct lnet_mt_event_info *ev_info;
3121         struct list_head processed_list;
3122         struct list_head local_queue;
3123         struct lnet_handle_md mdh;
3124         struct lnet_ni *tmp;
3125         struct lnet_ni *ni;
3126         lnet_nid_t nid;
3127         int healthv;
3128         int rc;
3129
3130         INIT_LIST_HEAD(&local_queue);
3131         INIT_LIST_HEAD(&processed_list);
3132
3133         /*
3134          * splice the recovery queue on a local queue. We will iterate
3135          * through the local queue and update it as needed. Once we're
3136          * done with the traversal, we'll splice the local queue back on
3137          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3138          * will be traversed in the next iteration.
3139          */
3140         lnet_net_lock(0);
3141         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3142                          &local_queue);
3143         lnet_net_unlock(0);
3144
3145         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3146                 /*
3147                  * if an NI is being deleted or it is now healthy, there
3148                  * is no need to keep it around in the recovery queue.
3149                  * The monitor thread is the only thread responsible for
3150                  * removing the NI from the recovery queue.
3151                  * Multiple threads can be adding NIs to the recovery
3152                  * queue.
3153                  */
3154                 healthv = atomic_read(&ni->ni_healthv);
3155
3156                 lnet_net_lock(0);
3157                 lnet_ni_lock(ni);
3158                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3159                     healthv == LNET_MAX_HEALTH_VALUE) {
3160                         list_del_init(&ni->ni_recovery);
3161                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3162                         lnet_ni_unlock(ni);
3163                         lnet_ni_decref_locked(ni, 0);
3164                         lnet_net_unlock(0);
3165                         continue;
3166                 }
3167
3168                 /*
3169                  * if the local NI failed recovery we must unlink the md.
3170                  * But we want to keep the local_ni on the recovery queue
3171                  * so we can continue the attempts to recover it.
3172                  */
3173                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3174                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3175                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3176                 }
3177
3178                 lnet_ni_unlock(ni);
3179                 lnet_net_unlock(0);
3180
3181
3182                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3183                        libcfs_nid2str(ni->ni_nid));
3184
3185                 lnet_ni_lock(ni);
3186                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3187                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3188                         lnet_ni_unlock(ni);
3189
3190                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3191                         if (!ev_info) {
3192                                 CERROR("out of memory. Can't recover %s\n",
3193                                        libcfs_nid2str(ni->ni_nid));
3194                                 lnet_ni_lock(ni);
3195                                 ni->ni_recovery_state &=
3196                                   ~LNET_NI_RECOVERY_PENDING;
3197                                 lnet_ni_unlock(ni);
3198                                 continue;
3199                         }
3200
3201                         mdh = ni->ni_ping_mdh;
3202                         /*
3203                          * Invalidate the ni mdh in case it's deleted.
3204                          * We'll unlink the mdh in this case below.
3205                          */
3206                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3207                         nid = ni->ni_nid;
3208
3209                         /*
3210                          * remove the NI from the local queue and drop the
3211                          * reference count to it while we're recovering
3212                          * it. The reason for that, is that the NI could
3213                          * be deleted, and the way the code is structured
3214                          * is if we don't drop the NI, then the deletion
3215                          * code will enter a loop waiting for the
3216                          * reference count to be removed while holding the
3217                          * ln_mutex_lock(). When we look up the peer to
3218                          * send to in lnet_select_pathway() we will try to
3219                          * lock the ln_mutex_lock() as well, leading to
3220                          * a deadlock. By dropping the refcount and
3221                          * removing it from the list, we allow for the NI
3222                          * to be removed, then we use the cached NID to
3223                          * look it up again. If it's gone, then we just
3224                          * continue examining the rest of the queue.
3225                          */
3226                         lnet_net_lock(0);
3227                         list_del_init(&ni->ni_recovery);
3228                         lnet_ni_decref_locked(ni, 0);
3229                         lnet_net_unlock(0);
3230
3231                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3232                         ev_info->mt_nid = nid;
3233                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3234                                             ev_info, the_lnet.ln_mt_eqh, true);
3235                         /* lookup the nid again */
3236                         lnet_net_lock(0);
3237                         ni = lnet_nid2ni_locked(nid, 0);
3238                         if (!ni) {
3239                                 /*
3240                                  * the NI has been deleted when we dropped
3241                                  * the ref count
3242                                  */
3243                                 lnet_net_unlock(0);
3244                                 LNetMDUnlink(mdh);
3245                                 continue;
3246                         }
3247                         /*
3248                          * Same note as in lnet_recover_peer_nis(). When
3249                          * we're sending the ping, the NI is free to be
3250                          * deleted or manipulated. By this point it
3251                          * could've been added back on the recovery queue,
3252                          * and a refcount taken on it.
3253                          * So we can't just add it blindly again or we'll
3254                          * corrupt the queue. We must check under lock if
3255                          * it's not on any list and if not then add it
3256                          * to the processed list, which will eventually be
3257                          * spliced back on to the recovery queue.
3258                          */
3259                         ni->ni_ping_mdh = mdh;
3260                         if (list_empty(&ni->ni_recovery)) {
3261                                 list_add_tail(&ni->ni_recovery, &processed_list);
3262                                 lnet_ni_addref_locked(ni, 0);
3263                         }
3264                         lnet_net_unlock(0);
3265
3266                         lnet_ni_lock(ni);
3267                         if (rc)
3268                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3269                 }
3270                 lnet_ni_unlock(ni);
3271         }
3272
3273         /*
3274          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3275          * reexamined in the next iteration.
3276          */
3277         list_splice_init(&processed_list, &local_queue);
3278         lnet_net_lock(0);
3279         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3280         lnet_net_unlock(0);
3281 }
3282
3283 static int
3284 lnet_resendqs_create(void)
3285 {
3286         struct list_head **resendqs;
3287         resendqs = lnet_create_array_of_queues();
3288
3289         if (!resendqs)
3290                 return -ENOMEM;
3291
3292         lnet_net_lock(LNET_LOCK_EX);
3293         the_lnet.ln_mt_resendqs = resendqs;
3294         lnet_net_unlock(LNET_LOCK_EX);
3295
3296         return 0;
3297 }
3298
3299 static void
3300 lnet_clean_local_ni_recoveryq(void)
3301 {
3302         struct lnet_ni *ni;
3303
3304         /* This is only called when the monitor thread has stopped */
3305         lnet_net_lock(0);
3306
3307         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3308                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3309                                 struct lnet_ni, ni_recovery);
3310                 list_del_init(&ni->ni_recovery);
3311                 lnet_ni_lock(ni);
3312                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3313                 lnet_ni_unlock(ni);
3314                 lnet_ni_decref_locked(ni, 0);
3315         }
3316
3317         lnet_net_unlock(0);
3318 }
3319
3320 static void
3321 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3322                                      bool force)
3323 {
3324         struct lnet_handle_md recovery_mdh;
3325
3326         LNetInvalidateMDHandle(&recovery_mdh);
3327
3328         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3329                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3330                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3331         }
3332         spin_unlock(&lpni->lpni_lock);
3333         lnet_net_unlock(cpt);
3334         if (!LNetMDHandleIsInvalid(recovery_mdh))
3335                 LNetMDUnlink(recovery_mdh);
3336         lnet_net_lock(cpt);
3337         spin_lock(&lpni->lpni_lock);
3338 }
3339
3340 static void
3341 lnet_clean_peer_ni_recoveryq(void)
3342 {
3343         struct lnet_peer_ni *lpni, *tmp;
3344
3345         lnet_net_lock(LNET_LOCK_EX);
3346
3347         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3348                                  lpni_recovery) {
3349                 list_del_init(&lpni->lpni_recovery);
3350                 spin_lock(&lpni->lpni_lock);
3351                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3352                 spin_unlock(&lpni->lpni_lock);
3353                 lnet_peer_ni_decref_locked(lpni);
3354         }
3355
3356         lnet_net_unlock(LNET_LOCK_EX);
3357 }
3358
3359 static void
3360 lnet_clean_resendqs(void)
3361 {
3362         struct lnet_msg *msg, *tmp;
3363         struct list_head msgs;
3364         int i;
3365
3366         INIT_LIST_HEAD(&msgs);
3367
3368         cfs_cpt_for_each(i, lnet_cpt_table()) {
3369                 lnet_net_lock(i);
3370                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3371                 lnet_net_unlock(i);
3372                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3373                         list_del_init(&msg->msg_list);
3374                         msg->msg_no_resend = true;
3375                         lnet_finalize(msg, -ESHUTDOWN);
3376                 }
3377         }
3378
3379         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3380 }
3381
3382 static void
3383 lnet_recover_peer_nis(void)
3384 {
3385         struct lnet_mt_event_info *ev_info;
3386         struct list_head processed_list;
3387         struct list_head local_queue;
3388         struct lnet_handle_md mdh;
3389         struct lnet_peer_ni *lpni;
3390         struct lnet_peer_ni *tmp;
3391         lnet_nid_t nid;
3392         int healthv;
3393         int rc;
3394
3395         INIT_LIST_HEAD(&local_queue);
3396         INIT_LIST_HEAD(&processed_list);
3397
3398         /*
3399          * Always use cpt 0 for locking across all interactions with
3400          * ln_mt_peerNIRecovq
3401          */
3402         lnet_net_lock(0);
3403         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3404                          &local_queue);
3405         lnet_net_unlock(0);
3406
3407         list_for_each_entry_safe(lpni, tmp, &local_queue,
3408                                  lpni_recovery) {
3409                 /*
3410                  * The same protection strategy is used here as is in the
3411                  * local recovery case.
3412                  */
3413                 lnet_net_lock(0);
3414                 healthv = atomic_read(&lpni->lpni_healthv);
3415                 spin_lock(&lpni->lpni_lock);
3416                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3417                     healthv == LNET_MAX_HEALTH_VALUE) {
3418                         list_del_init(&lpni->lpni_recovery);
3419                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3420                         spin_unlock(&lpni->lpni_lock);
3421                         lnet_peer_ni_decref_locked(lpni);
3422                         lnet_net_unlock(0);
3423                         continue;
3424                 }
3425
3426                 /*
3427                  * If the peer NI has failed recovery we must unlink the
3428                  * md. But we want to keep the peer ni on the recovery
3429                  * queue so we can try to continue recovering it
3430                  */
3431                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3432                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3433                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3434                 }
3435
3436                 spin_unlock(&lpni->lpni_lock);
3437                 lnet_net_unlock(0);
3438
3439                 /*
3440                  * NOTE: we're racing with peer deletion from user space.
3441                  * It's possible that a peer is deleted after we check its
3442                  * state. In this case the recovery can create a new peer
3443                  */
3444                 spin_lock(&lpni->lpni_lock);
3445                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3446                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3447                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3448                         spin_unlock(&lpni->lpni_lock);
3449
3450                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3451                         if (!ev_info) {
3452                                 CERROR("out of memory. Can't recover %s\n",
3453                                        libcfs_nid2str(lpni->lpni_nid));
3454                                 spin_lock(&lpni->lpni_lock);
3455                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3456                                 spin_unlock(&lpni->lpni_lock);
3457                                 continue;
3458                         }
3459
3460                         /* look at the comments in lnet_recover_local_nis() */
3461                         mdh = lpni->lpni_recovery_ping_mdh;
3462                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3463                         nid = lpni->lpni_nid;
3464                         lnet_net_lock(0);
3465                         list_del_init(&lpni->lpni_recovery);
3466                         lnet_peer_ni_decref_locked(lpni);
3467                         lnet_net_unlock(0);
3468
3469                         ev_info->mt_type = MT_TYPE_PEER_NI;
3470                         ev_info->mt_nid = nid;
3471                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3472                                             ev_info, the_lnet.ln_mt_eqh, true);
3473                         lnet_net_lock(0);
3474                         /*
3475                          * lnet_find_peer_ni_locked() grabs a refcount for
3476                          * us. No need to take it explicitly.
3477                          */
3478                         lpni = lnet_find_peer_ni_locked(nid);
3479                         if (!lpni) {
3480                                 lnet_net_unlock(0);
3481                                 LNetMDUnlink(mdh);
3482                                 continue;
3483                         }
3484
3485                         lpni->lpni_recovery_ping_mdh = mdh;
3486                         /*
3487                          * While we're unlocked the lpni could've been
3488                          * readded on the recovery queue. In this case we
3489                          * don't need to add it to the local queue, since
3490                          * it's already on there and the thread that added
3491                          * it would've incremented the refcount on the
3492                          * peer, which means we need to decref the refcount
3493                          * that was implicitly grabbed by find_peer_ni_locked.
3494                          * Otherwise, if the lpni is still not on
3495                          * the recovery queue, then we'll add it to the
3496                          * processed list.
3497                          */
3498                         if (list_empty(&lpni->lpni_recovery))
3499                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3500                         else
3501                                 lnet_peer_ni_decref_locked(lpni);
3502                         lnet_net_unlock(0);
3503
3504                         spin_lock(&lpni->lpni_lock);
3505                         if (rc)
3506                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3507                 }
3508                 spin_unlock(&lpni->lpni_lock);
3509         }
3510
3511         list_splice_init(&processed_list, &local_queue);
3512         lnet_net_lock(0);
3513         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3514         lnet_net_unlock(0);
3515 }
3516
3517 static int
3518 lnet_monitor_thread(void *arg)
3519 {
3520         time64_t recovery_timeout = 0;
3521         time64_t rsp_timeout = 0;
3522         int interval;
3523         time64_t now;
3524
3525         wait_for_completion(&the_lnet.ln_started);
3526         /*
3527          * The monitor thread takes care of the following:
3528          *  1. Checks the aliveness of routers
3529          *  2. Checks if there are messages on the resend queue to resend
3530          *     them.
3531          *  3. Check if there are any NIs on the local recovery queue and
3532          *     pings them
3533          *  4. Checks if there are any NIs on the remote recovery queue
3534          *     and pings them.
3535          */
3536         cfs_block_allsigs();
3537
3538         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3539                 now = ktime_get_real_seconds();
3540
3541                 if (lnet_router_checker_active())
3542                         lnet_check_routers();
3543
3544                 lnet_resend_pending_msgs();
3545
3546                 if (now >= rsp_timeout) {
3547                         lnet_finalize_expired_responses();
3548                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3549                 }
3550
3551                 if (now >= recovery_timeout) {
3552                         lnet_recover_local_nis();
3553                         lnet_recover_peer_nis();
3554                         recovery_timeout = now + lnet_recovery_interval;
3555                 }
3556
3557                 /*
3558                  * TODO do we need to check if we should sleep without
3559                  * timeout?  Technically, an active system will always
3560                  * have messages in flight so this check will always
3561                  * evaluate to false. And on an idle system do we care
3562                  * if we wake up every 1 second? Although, we've seen
3563                  * cases where we get a complaint that an idle thread
3564                  * is waking up unnecessarily.
3565                  *
3566                  * Take into account the current net_count when you wake
3567                  * up for alive router checking, since we need to check
3568                  * possibly as many networks as we have configured.
3569                  */
3570                 interval = min(lnet_recovery_interval,
3571                                min((unsigned int) alive_router_check_interval /
3572                                         lnet_current_net_count,
3573                                    lnet_transaction_timeout / 2));
3574                 wait_for_completion_interruptible_timeout(
3575                         &the_lnet.ln_mt_wait_complete,
3576                         cfs_time_seconds(interval));
3577                 /* Must re-init the completion before testing anything,
3578                  * including ln_mt_state.
3579                  */
3580                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3581         }
3582
3583         /* Shutting down */
3584         lnet_net_lock(LNET_LOCK_EX);
3585         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3586         lnet_net_unlock(LNET_LOCK_EX);
3587
3588         /* signal that the monitor thread is exiting */
3589         up(&the_lnet.ln_mt_signal);
3590
3591         return 0;
3592 }
3593
3594 /*
3595  * lnet_send_ping
3596  * Sends a ping.
3597  * Returns == 0 if success
3598  * Returns > 0 if LNetMDBind or prior fails
3599  * Returns < 0 if LNetGet fails
3600  */
3601 int
3602 lnet_send_ping(lnet_nid_t dest_nid,
3603                struct lnet_handle_md *mdh, int nnis,
3604                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3605 {
3606         struct lnet_md md = { NULL };
3607         struct lnet_process_id id;
3608         struct lnet_ping_buffer *pbuf;
3609         int rc;
3610
3611         if (dest_nid == LNET_NID_ANY) {
3612                 rc = -EHOSTUNREACH;
3613                 goto fail_error;
3614         }
3615
3616         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3617         if (!pbuf) {
3618                 rc = ENOMEM;
3619                 goto fail_error;
3620         }
3621
3622         /* initialize md content */
3623         md.start     = &pbuf->pb_info;
3624         md.length    = LNET_PING_INFO_SIZE(nnis);
3625         md.threshold = 2; /* GET/REPLY */
3626         md.max_size  = 0;
3627         md.options   = LNET_MD_TRUNCATE;
3628         md.user_ptr  = user_data;
3629         md.eq_handle = eqh;
3630
3631         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3632         if (rc) {
3633                 lnet_ping_buffer_decref(pbuf);
3634                 CERROR("Can't bind MD: %d\n", rc);
3635                 rc = -rc; /* change the rc to positive */
3636                 goto fail_error;
3637         }
3638         id.pid = LNET_PID_LUSTRE;
3639         id.nid = dest_nid;
3640
3641         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3642                      LNET_RESERVED_PORTAL,
3643                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3644
3645         if (rc)
3646                 goto fail_unlink_md;
3647
3648         return 0;
3649
3650 fail_unlink_md:
3651         LNetMDUnlink(*mdh);
3652         LNetInvalidateMDHandle(mdh);
3653 fail_error:
3654         return rc;
3655 }
3656
3657 static void
3658 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3659                            int status, bool unlink_event)
3660 {
3661         lnet_nid_t nid = ev_info->mt_nid;
3662
3663         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3664                 struct lnet_ni *ni;
3665
3666                 lnet_net_lock(0);
3667                 ni = lnet_nid2ni_locked(nid, 0);
3668                 if (!ni) {
3669                         lnet_net_unlock(0);
3670                         return;
3671                 }
3672                 lnet_ni_lock(ni);
3673                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3674                 if (status)
3675                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3676                 lnet_ni_unlock(ni);
3677                 lnet_net_unlock(0);
3678
3679                 if (status != 0) {
3680                         CERROR("local NI (%s) recovery failed with %d\n",
3681                                libcfs_nid2str(nid), status);
3682                         return;
3683                 }
3684                 /*
3685                  * need to increment healthv for the ni here, because in
3686                  * the lnet_finalize() path we don't have access to this
3687                  * NI. And in order to get access to it, we'll need to
3688                  * carry forward too much information.
3689                  * In the peer case, it'll naturally be incremented
3690                  */
3691                 if (!unlink_event)
3692                         lnet_inc_healthv(&ni->ni_healthv);
3693         } else {
3694                 struct lnet_peer_ni *lpni;
3695                 int cpt;
3696
3697                 cpt = lnet_net_lock_current();
3698                 lpni = lnet_find_peer_ni_locked(nid);
3699                 if (!lpni) {
3700                         lnet_net_unlock(cpt);
3701                         return;
3702                 }
3703                 spin_lock(&lpni->lpni_lock);
3704                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3705                 if (status)
3706                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3707                 spin_unlock(&lpni->lpni_lock);
3708                 lnet_peer_ni_decref_locked(lpni);
3709                 lnet_net_unlock(cpt);
3710
3711                 if (status != 0)
3712                         CERROR("peer NI (%s) recovery failed with %d\n",
3713                                libcfs_nid2str(nid), status);
3714         }
3715 }
3716
3717 void
3718 lnet_mt_event_handler(struct lnet_event *event)
3719 {
3720         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3721         struct lnet_ping_buffer *pbuf;
3722
3723         /* TODO: remove assert */
3724         LASSERT(event->type == LNET_EVENT_REPLY ||
3725                 event->type == LNET_EVENT_SEND ||
3726                 event->type == LNET_EVENT_UNLINK);
3727
3728         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3729                event->status);
3730
3731         switch (event->type) {
3732         case LNET_EVENT_UNLINK:
3733                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3734                        libcfs_nid2str(ev_info->mt_nid));
3735                 /* fallthrough */
3736         case LNET_EVENT_REPLY:
3737                 lnet_handle_recovery_reply(ev_info, event->status,
3738                                            event->type == LNET_EVENT_UNLINK);
3739                 break;
3740         case LNET_EVENT_SEND:
3741                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3742                                libcfs_nid2str(ev_info->mt_nid),
3743                                (event->status) ? "unsuccessfully" :
3744                                "successfully", event->status);
3745                 break;
3746         default:
3747                 CERROR("Unexpected event: %d\n", event->type);
3748                 break;
3749         }
3750         if (event->unlinked) {
3751                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3752                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3753                 lnet_ping_buffer_decref(pbuf);
3754         }
3755 }
3756
3757 static int
3758 lnet_rsp_tracker_create(void)
3759 {
3760         struct list_head **rstqs;
3761         rstqs = lnet_create_array_of_queues();
3762
3763         if (!rstqs)
3764                 return -ENOMEM;
3765
3766         the_lnet.ln_mt_rstq = rstqs;
3767
3768         return 0;
3769 }
3770
3771 static void
3772 lnet_rsp_tracker_clean(void)
3773 {
3774         lnet_finalize_expired_responses();
3775
3776         cfs_percpt_free(the_lnet.ln_mt_rstq);
3777         the_lnet.ln_mt_rstq = NULL;
3778 }
3779
3780 int lnet_monitor_thr_start(void)
3781 {
3782         int rc = 0;
3783         struct task_struct *task;
3784
3785         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3786                 return -EALREADY;
3787
3788         rc = lnet_resendqs_create();
3789         if (rc)
3790                 return rc;
3791
3792         rc = lnet_rsp_tracker_create();
3793         if (rc)
3794                 goto clean_queues;
3795
3796         sema_init(&the_lnet.ln_mt_signal, 0);
3797
3798         lnet_net_lock(LNET_LOCK_EX);
3799         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3800         lnet_net_unlock(LNET_LOCK_EX);
3801         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3802         if (IS_ERR(task)) {
3803                 rc = PTR_ERR(task);
3804                 CERROR("Can't start monitor thread: %d\n", rc);
3805                 goto clean_thread;
3806         }
3807
3808         return 0;
3809
3810 clean_thread:
3811         lnet_net_lock(LNET_LOCK_EX);
3812         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3813         lnet_net_unlock(LNET_LOCK_EX);
3814         /* block until event callback signals exit */
3815         down(&the_lnet.ln_mt_signal);
3816         /* clean up */
3817         lnet_net_lock(LNET_LOCK_EX);
3818         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3819         lnet_net_unlock(LNET_LOCK_EX);
3820         lnet_rsp_tracker_clean();
3821         lnet_clean_local_ni_recoveryq();
3822         lnet_clean_peer_ni_recoveryq();
3823         lnet_clean_resendqs();
3824         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3825         return rc;
3826 clean_queues:
3827         lnet_rsp_tracker_clean();
3828         lnet_clean_local_ni_recoveryq();
3829         lnet_clean_peer_ni_recoveryq();
3830         lnet_clean_resendqs();
3831         return rc;
3832 }
3833
3834 void lnet_monitor_thr_stop(void)
3835 {
3836         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3837                 return;
3838
3839         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3840         lnet_net_lock(LNET_LOCK_EX);
3841         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3842         lnet_net_unlock(LNET_LOCK_EX);
3843
3844         /* tell the monitor thread that we're shutting down */
3845         complete(&the_lnet.ln_mt_wait_complete);
3846
3847         /* block until monitor thread signals that it's done */
3848         down(&the_lnet.ln_mt_signal);
3849         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3850
3851         /* perform cleanup tasks */
3852         lnet_rsp_tracker_clean();
3853         lnet_clean_local_ni_recoveryq();
3854         lnet_clean_peer_ni_recoveryq();
3855         lnet_clean_resendqs();
3856 }
3857
3858 void
3859 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3860                   __u32 msg_type)
3861 {
3862         lnet_net_lock(cpt);
3863         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3864         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3865         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3866         lnet_net_unlock(cpt);
3867
3868         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3869 }
3870
3871 static void
3872 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3873 {
3874         struct lnet_hdr *hdr = &msg->msg_hdr;
3875
3876         if (msg->msg_wanted != 0)
3877                 lnet_setpayloadbuffer(msg);
3878
3879         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3880
3881         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3882          * it back into the ACK during lnet_finalize() */
3883         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3884                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3885
3886         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3887                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3888 }
3889
3890 static int
3891 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3892 {
3893         struct lnet_hdr         *hdr = &msg->msg_hdr;
3894         struct lnet_match_info  info;
3895         int                     rc;
3896         bool                    ready_delay;
3897
3898         /* Convert put fields to host byte order */
3899         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3900         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3901         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3902
3903         /* Primary peer NID. */
3904         info.mi_id.nid  = msg->msg_initiator;
3905         info.mi_id.pid  = hdr->src_pid;
3906         info.mi_opc     = LNET_MD_OP_PUT;
3907         info.mi_portal  = hdr->msg.put.ptl_index;
3908         info.mi_rlength = hdr->payload_length;
3909         info.mi_roffset = hdr->msg.put.offset;
3910         info.mi_mbits   = hdr->msg.put.match_bits;
3911         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3912
3913         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3914         ready_delay = msg->msg_rx_ready_delay;
3915
3916  again:
3917         rc = lnet_ptl_match_md(&info, msg);
3918         switch (rc) {
3919         default:
3920                 LBUG();
3921
3922         case LNET_MATCHMD_OK:
3923                 lnet_recv_put(ni, msg);
3924                 return 0;
3925
3926         case LNET_MATCHMD_NONE:
3927                 if (ready_delay)
3928                         /* no eager_recv or has already called it, should
3929                          * have been attached on delayed list */
3930                         return 0;
3931
3932                 rc = lnet_ni_eager_recv(ni, msg);
3933                 if (rc == 0) {
3934                         ready_delay = true;
3935                         goto again;
3936                 }
3937                 /* fall through */
3938
3939         case LNET_MATCHMD_DROP:
3940                 CNETERR("Dropping PUT from %s portal %d match %llu"
3941                         " offset %d length %d: %d\n",
3942                         libcfs_id2str(info.mi_id), info.mi_portal,
3943                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3944
3945                 return -ENOENT; /* -ve: OK but no match */
3946         }
3947 }
3948
3949 static int
3950 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3951 {
3952         struct lnet_match_info info;
3953         struct lnet_hdr *hdr = &msg->msg_hdr;
3954         struct lnet_process_id source_id;
3955         struct lnet_handle_wire reply_wmd;
3956         int rc;
3957
3958         /* Convert get fields to host byte order */
3959         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3960         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3961         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3962         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3963
3964         source_id.nid = hdr->src_nid;
3965         source_id.pid = hdr->src_pid;
3966         /* Primary peer NID */
3967         info.mi_id.nid  = msg->msg_initiator;
3968         info.mi_id.pid  = hdr->src_pid;
3969         info.mi_opc     = LNET_MD_OP_GET;
3970         info.mi_portal  = hdr->msg.get.ptl_index;
3971         info.mi_rlength = hdr->msg.get.sink_length;
3972         info.mi_roffset = hdr->msg.get.src_offset;
3973         info.mi_mbits   = hdr->msg.get.match_bits;
3974         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3975
3976         rc = lnet_ptl_match_md(&info, msg);
3977         if (rc == LNET_MATCHMD_DROP) {
3978                 CNETERR("Dropping GET from %s portal %d match %llu"
3979                         " offset %d length %d\n",
3980                         libcfs_id2str(info.mi_id), info.mi_portal,
3981                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3982                 return -ENOENT; /* -ve: OK but no match */
3983         }
3984
3985         LASSERT(rc == LNET_MATCHMD_OK);
3986
3987         lnet_build_msg_event(msg, LNET_EVENT_GET);
3988
3989         reply_wmd = hdr->msg.get.return_wmd;
3990
3991         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3992                        msg->msg_offset, msg->msg_wanted);
3993
3994         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3995
3996         if (rdma_get) {
3997                 /* The LND completes the REPLY from her recv procedure */
3998                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3999                              msg->msg_offset, msg->msg_len, msg->msg_len);
4000                 return 0;
4001         }
4002
4003         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4004         msg->msg_receiving = 0;
4005
4006         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4007         if (rc < 0) {
4008                 /* didn't get as far as lnet_ni_send() */
4009                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4010                        libcfs_nid2str(ni->ni_nid),
4011                        libcfs_id2str(info.mi_id), rc);
4012
4013                 lnet_finalize(msg, rc);
4014         }
4015
4016         return 0;
4017 }
4018
4019 static int
4020 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4021 {
4022         void *private = msg->msg_private;
4023         struct lnet_hdr *hdr = &msg->msg_hdr;
4024         struct lnet_process_id src = {0};
4025         struct lnet_libmd *md;
4026         int rlength;
4027         int mlength;
4028         int cpt;
4029
4030         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4031         lnet_res_lock(cpt);
4032
4033         src.nid = hdr->src_nid;
4034         src.pid = hdr->src_pid;
4035
4036         /* NB handles only looked up by creator (no flips) */
4037         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4038         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4039                 CNETERR("%s: Dropping REPLY from %s for %s "
4040                         "MD %#llx.%#llx\n",
4041                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4042                         (md == NULL) ? "invalid" : "inactive",
4043                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4044                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4045                 if (md != NULL && md->md_me != NULL)
4046                         CERROR("REPLY MD also attached to portal %d\n",
4047                                md->md_me->me_portal);
4048
4049                 lnet_res_unlock(cpt);
4050                 return -ENOENT; /* -ve: OK but no match */
4051         }
4052
4053         LASSERT(md->md_offset == 0);
4054
4055         rlength = hdr->payload_length;
4056         mlength = MIN(rlength, (int)md->md_length);
4057
4058         if (mlength < rlength &&
4059             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4060                 CNETERR("%s: Dropping REPLY from %s length %d "
4061                         "for MD %#llx would overflow (%d)\n",
4062                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4063                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4064                         mlength);
4065                 lnet_res_unlock(cpt);
4066                 return -ENOENT; /* -ve: OK but no match */
4067         }
4068
4069         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4070                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4071                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4072
4073         lnet_msg_attach_md(msg, md, 0, mlength);
4074
4075         if (mlength != 0)
4076                 lnet_setpayloadbuffer(msg);
4077
4078         lnet_res_unlock(cpt);
4079
4080         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4081
4082         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4083         return 0;
4084 }
4085
4086 static int
4087 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4088 {
4089         struct lnet_hdr *hdr = &msg->msg_hdr;
4090         struct lnet_process_id src = {0};
4091         struct lnet_libmd *md;
4092         int cpt;
4093
4094         src.nid = hdr->src_nid;
4095         src.pid = hdr->src_pid;
4096
4097         /* Convert ack fields to host byte order */
4098         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4099         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4100
4101         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4102         lnet_res_lock(cpt);
4103
4104         /* NB handles only looked up by creator (no flips) */
4105         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4106         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4107                 /* Don't moan; this is expected */
4108                 CDEBUG(D_NET,
4109                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4110                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4111                        (md == NULL) ? "invalid" : "inactive",
4112                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4113                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4114                 if (md != NULL && md->md_me != NULL)
4115                         CERROR("Source MD also attached to portal %d\n",
4116                                md->md_me->me_portal);
4117
4118                 lnet_res_unlock(cpt);
4119                 return -ENOENT;                  /* -ve! */
4120         }
4121
4122         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4123                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4124                hdr->msg.ack.dst_wmd.wh_object_cookie);
4125
4126         lnet_msg_attach_md(msg, md, 0, 0);
4127
4128         lnet_res_unlock(cpt);
4129
4130         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4131
4132         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4133         return 0;
4134 }
4135
4136 /**
4137  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4138  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4139  * \retval -ve                  error code
4140  */
4141 int
4142 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4143 {
4144         int     rc = 0;
4145
4146         if (!the_lnet.ln_routing)
4147                 return -ECANCELED;
4148
4149         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4150             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4151                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4152                         msg->msg_rx_ready_delay = 1;
4153                 } else {
4154                         lnet_net_unlock(msg->msg_rx_cpt);
4155                         rc = lnet_ni_eager_recv(ni, msg);
4156                         lnet_net_lock(msg->msg_rx_cpt);
4157                 }
4158         }
4159
4160         if (rc == 0)
4161                 rc = lnet_post_routed_recv_locked(msg, 0);
4162         return rc;
4163 }
4164
4165 int
4166 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4167 {
4168         int     rc;
4169
4170         switch (msg->msg_type) {
4171         case LNET_MSG_ACK:
4172                 rc = lnet_parse_ack(ni, msg);
4173                 break;
4174         case LNET_MSG_PUT:
4175                 rc = lnet_parse_put(ni, msg);
4176                 break;
4177         case LNET_MSG_GET:
4178                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4179                 break;
4180         case LNET_MSG_REPLY:
4181                 rc = lnet_parse_reply(ni, msg);
4182                 break;
4183         default: /* prevent an unused label if !kernel */
4184                 LASSERT(0);
4185                 return -EPROTO;
4186         }
4187
4188         LASSERT(rc == 0 || rc == -ENOENT);
4189         return rc;
4190 }
4191
4192 char *
4193 lnet_msgtyp2str (int type)
4194 {
4195         switch (type) {
4196         case LNET_MSG_ACK:
4197                 return ("ACK");
4198         case LNET_MSG_PUT:
4199                 return ("PUT");
4200         case LNET_MSG_GET:
4201                 return ("GET");
4202         case LNET_MSG_REPLY:
4203                 return ("REPLY");
4204         case LNET_MSG_HELLO:
4205                 return ("HELLO");
4206         default:
4207                 return ("<UNKNOWN>");
4208         }
4209 }
4210
4211 void
4212 lnet_print_hdr(struct lnet_hdr *hdr)
4213 {
4214         struct lnet_process_id src = {
4215                 .nid = hdr->src_nid,
4216                 .pid = hdr->src_pid,
4217         };
4218         struct lnet_process_id dst = {
4219                 .nid = hdr->dest_nid,
4220                 .pid = hdr->dest_pid,
4221         };
4222         char *type_str = lnet_msgtyp2str(hdr->type);
4223
4224         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4225         CWARN("    From %s\n", libcfs_id2str(src));
4226         CWARN("    To   %s\n", libcfs_id2str(dst));
4227
4228         switch (hdr->type) {
4229         default:
4230                 break;
4231
4232         case LNET_MSG_PUT:
4233                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4234                       "match bits %llu\n",
4235                       hdr->msg.put.ptl_index,
4236                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4237                       hdr->msg.put.ack_wmd.wh_object_cookie,
4238                       hdr->msg.put.match_bits);
4239                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4240                       hdr->payload_length, hdr->msg.put.offset,
4241                       hdr->msg.put.hdr_data);
4242                 break;
4243
4244         case LNET_MSG_GET:
4245                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4246                       "match bits %llu\n", hdr->msg.get.ptl_index,
4247                       hdr->msg.get.return_wmd.wh_interface_cookie,
4248                       hdr->msg.get.return_wmd.wh_object_cookie,
4249                       hdr->msg.get.match_bits);
4250                 CWARN("    Length %d, src offset %d\n",
4251                       hdr->msg.get.sink_length,
4252                       hdr->msg.get.src_offset);
4253                 break;
4254
4255         case LNET_MSG_ACK:
4256                 CWARN("    dst md %#llx.%#llx, "
4257                       "manipulated length %d\n",
4258                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4259                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4260                       hdr->msg.ack.mlength);
4261                 break;
4262
4263         case LNET_MSG_REPLY:
4264                 CWARN("    dst md %#llx.%#llx, "
4265                       "length %d\n",
4266                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4267                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4268                       hdr->payload_length);
4269         }
4270
4271 }
4272
4273 int
4274 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4275            void *private, int rdma_req)
4276 {
4277         struct lnet_peer_ni *lpni;
4278         struct lnet_msg *msg;
4279         __u32 payload_length;
4280         lnet_pid_t dest_pid;
4281         lnet_nid_t dest_nid;
4282         lnet_nid_t src_nid;
4283         bool push = false;
4284         int for_me;
4285         __u32 type;
4286         int rc = 0;
4287         int cpt;
4288
4289         LASSERT (!in_interrupt ());
4290
4291         type = le32_to_cpu(hdr->type);
4292         src_nid = le64_to_cpu(hdr->src_nid);
4293         dest_nid = le64_to_cpu(hdr->dest_nid);
4294         dest_pid = le32_to_cpu(hdr->dest_pid);
4295         payload_length = le32_to_cpu(hdr->payload_length);
4296
4297         for_me = (ni->ni_nid == dest_nid);
4298         cpt = lnet_cpt_of_nid(from_nid, ni);
4299
4300         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4301                 libcfs_nid2str(dest_nid),
4302                 libcfs_nid2str(ni->ni_nid),
4303                 libcfs_nid2str(src_nid),
4304                 lnet_msgtyp2str(type),
4305                 (for_me) ? "for me" : "routed");
4306
4307         switch (type) {
4308         case LNET_MSG_ACK:
4309         case LNET_MSG_GET:
4310                 if (payload_length > 0) {
4311                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4312                                libcfs_nid2str(from_nid),
4313                                libcfs_nid2str(src_nid),
4314                                lnet_msgtyp2str(type), payload_length);
4315                         return -EPROTO;
4316                 }
4317                 break;
4318
4319         case LNET_MSG_PUT:
4320         case LNET_MSG_REPLY:
4321                 if (payload_length >
4322                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4323                         CERROR("%s, src %s: bad %s payload %d "
4324                                "(%d max expected)\n",
4325                                libcfs_nid2str(from_nid),
4326                                libcfs_nid2str(src_nid),
4327                                lnet_msgtyp2str(type),
4328                                payload_length,
4329                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4330                         return -EPROTO;
4331                 }
4332                 break;
4333
4334         default:
4335                 CERROR("%s, src %s: Bad message type 0x%x\n",
4336                        libcfs_nid2str(from_nid),
4337                        libcfs_nid2str(src_nid), type);
4338                 return -EPROTO;
4339         }
4340
4341         if (the_lnet.ln_routing &&
4342             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4343                 lnet_ni_lock(ni);
4344                 spin_lock(&ni->ni_net->net_lock);
4345                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4346                 spin_unlock(&ni->ni_net->net_lock);
4347                 if (ni->ni_status != NULL &&
4348                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4349                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4350                         push = true;
4351                 }
4352                 lnet_ni_unlock(ni);
4353         }
4354
4355         if (push)
4356                 lnet_push_update_to_peers(1);
4357
4358         /* Regard a bad destination NID as a protocol error.  Senders should
4359          * know what they're doing; if they don't they're misconfigured, buggy
4360          * or malicious so we chop them off at the knees :) */
4361
4362         if (!for_me) {
4363                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4364                         /* should have gone direct */
4365                         CERROR("%s, src %s: Bad dest nid %s "
4366                                "(should have been sent direct)\n",
4367                                 libcfs_nid2str(from_nid),
4368                                 libcfs_nid2str(src_nid),
4369                                 libcfs_nid2str(dest_nid));
4370                         return -EPROTO;
4371                 }
4372
4373                 if (lnet_islocalnid(dest_nid)) {
4374                         /* dest is another local NI; sender should have used
4375                          * this node's NID on its own network */
4376                         CERROR("%s, src %s: Bad dest nid %s "
4377                                "(it's my nid but on a different network)\n",
4378                                 libcfs_nid2str(from_nid),
4379                                 libcfs_nid2str(src_nid),
4380                                 libcfs_nid2str(dest_nid));
4381                         return -EPROTO;
4382                 }
4383
4384                 if (rdma_req && type == LNET_MSG_GET) {
4385                         CERROR("%s, src %s: Bad optimized GET for %s "
4386                                "(final destination must be me)\n",
4387                                 libcfs_nid2str(from_nid),
4388                                 libcfs_nid2str(src_nid),
4389                                 libcfs_nid2str(dest_nid));
4390                         return -EPROTO;
4391                 }
4392
4393                 if (!the_lnet.ln_routing) {
4394                         CERROR("%s, src %s: Dropping message for %s "
4395                                "(routing not enabled)\n",
4396                                 libcfs_nid2str(from_nid),
4397                                 libcfs_nid2str(src_nid),
4398                                 libcfs_nid2str(dest_nid));
4399                         goto drop;
4400                 }
4401         }
4402
4403         /* Message looks OK; we're not going to return an error, so we MUST
4404          * call back lnd_recv() come what may... */
4405
4406         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4407             fail_peer(src_nid, 0)) {                    /* shall we now? */
4408                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4409                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4410                        lnet_msgtyp2str(type));
4411                 goto drop;
4412         }
4413
4414         if (!list_empty(&the_lnet.ln_drop_rules) &&
4415             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4416                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4417                               "silent message loss\n",
4418                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4419                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4420                 goto drop;
4421         }
4422
4423         if (lnet_drop_asym_route && for_me &&
4424             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4425                 struct lnet_net *net;
4426                 struct lnet_remotenet *rnet;
4427                 bool found = true;
4428
4429                 /* we are dealing with a routed message,
4430                  * so see if route to reach src_nid goes through from_nid
4431                  */
4432                 lnet_net_lock(cpt);
4433                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4434                 if (!net) {
4435                         lnet_net_unlock(cpt);
4436                         CERROR("net %s not found\n",
4437                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4438                         return -EPROTO;
4439                 }
4440
4441                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4442                 if (rnet) {
4443                         struct lnet_peer *gw = NULL;
4444                         struct lnet_peer_ni *lpni = NULL;
4445                         struct lnet_route *route;
4446
4447                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4448                                 found = false;
4449                                 gw = route->lr_gateway;
4450                                 if (route->lr_lnet != net->net_id)
4451                                         continue;
4452                                 /*
4453                                  * if the nid is one of the gateway's NIDs
4454                                  * then this is a valid gateway
4455                                  */
4456                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4457                                                 NULL, lpni)) != NULL) {
4458                                         if (lpni->lpni_nid == from_nid) {
4459                                                 found = true;
4460                                                 break;
4461                                         }
4462                                 }
4463                         }
4464                 }
4465                 lnet_net_unlock(cpt);
4466                 if (!found) {
4467                         /* we would not use from_nid to route a message to
4468                          * src_nid
4469                          * => asymmetric routing detected but forbidden
4470                          */
4471                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4472                                libcfs_nid2str(from_nid),
4473                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4474                         goto drop;
4475                 }
4476         }
4477
4478         msg = lnet_msg_alloc();
4479         if (msg == NULL) {
4480                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4481                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4482                        lnet_msgtyp2str(type));
4483                 goto drop;
4484         }
4485
4486         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4487          * pointers NULL etc */
4488
4489         msg->msg_type = type;
4490         msg->msg_private = private;
4491         msg->msg_receiving = 1;
4492         msg->msg_rdma_get = rdma_req;
4493         msg->msg_len = msg->msg_wanted = payload_length;
4494         msg->msg_offset = 0;
4495         msg->msg_hdr = *hdr;
4496         /* for building message event */
4497         msg->msg_from = from_nid;
4498         if (!for_me) {
4499                 msg->msg_target.pid     = dest_pid;
4500                 msg->msg_target.nid     = dest_nid;
4501                 msg->msg_routing        = 1;
4502
4503         } else {
4504                 /* convert common msg->hdr fields to host byteorder */
4505                 msg->msg_hdr.type       = type;
4506                 msg->msg_hdr.src_nid    = src_nid;
4507                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4508                 msg->msg_hdr.dest_nid   = dest_nid;
4509                 msg->msg_hdr.dest_pid   = dest_pid;
4510                 msg->msg_hdr.payload_length = payload_length;
4511         }
4512
4513         lnet_net_lock(cpt);
4514         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4515         if (IS_ERR(lpni)) {
4516                 lnet_net_unlock(cpt);
4517                 CERROR("%s, src %s: Dropping %s "
4518                        "(error %ld looking up sender)\n",
4519                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4520                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4521                 lnet_msg_free(msg);
4522                 if (rc == -ESHUTDOWN)
4523                         /* We are shutting down.  Don't do anything more */
4524                         return 0;
4525                 goto drop;
4526         }
4527
4528         if (the_lnet.ln_routing)
4529                 lpni->lpni_last_alive = ktime_get_seconds();
4530
4531         msg->msg_rxpeer = lpni;
4532         msg->msg_rxni = ni;
4533         lnet_ni_addref_locked(ni, cpt);
4534         /* Multi-Rail: Primary NID of source. */
4535         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4536
4537         /*
4538          * mark the status of this lpni as UP since we received a message
4539          * from it. The ping response reports back the ns_status which is
4540          * marked on the remote as up or down and we cache it here.
4541          */
4542         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4543
4544         lnet_msg_commit(msg, cpt);
4545
4546         /* message delay simulation */
4547         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4548                      lnet_delay_rule_match_locked(hdr, msg))) {
4549                 lnet_net_unlock(cpt);
4550                 return 0;
4551         }
4552
4553         if (!for_me) {
4554                 rc = lnet_parse_forward_locked(ni, msg);
4555                 lnet_net_unlock(cpt);
4556
4557                 if (rc < 0)
4558                         goto free_drop;
4559
4560                 if (rc == LNET_CREDIT_OK) {
4561                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4562                                      0, payload_length, payload_length);
4563                 }
4564                 return 0;
4565         }
4566
4567         lnet_net_unlock(cpt);
4568
4569         rc = lnet_parse_local(ni, msg);
4570         if (rc != 0)
4571                 goto free_drop;
4572         return 0;
4573
4574  free_drop:
4575         LASSERT(msg->msg_md == NULL);
4576         lnet_finalize(msg, rc);
4577
4578  drop:
4579         lnet_drop_message(ni, cpt, private, payload_length, type);
4580         return 0;
4581 }
4582 EXPORT_SYMBOL(lnet_parse);
4583
4584 void
4585 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4586 {
4587         while (!list_empty(head)) {
4588                 struct lnet_process_id id = {0};
4589                 struct lnet_msg *msg;
4590
4591                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4592                 list_del(&msg->msg_list);
4593
4594                 id.nid = msg->msg_hdr.src_nid;
4595                 id.pid = msg->msg_hdr.src_pid;
4596
4597                 LASSERT(msg->msg_md == NULL);
4598                 LASSERT(msg->msg_rx_delayed);
4599                 LASSERT(msg->msg_rxpeer != NULL);
4600                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4601
4602                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4603                       " offset %d length %d: %s\n",
4604                       libcfs_id2str(id),
4605                       msg->msg_hdr.msg.put.ptl_index,
4606                       msg->msg_hdr.msg.put.match_bits,
4607                       msg->msg_hdr.msg.put.offset,
4608                       msg->msg_hdr.payload_length, reason);
4609
4610                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4611                  * called lnet_drop_message(), so I just hang onto msg as well
4612                  * until that's done */
4613
4614                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4615                                   msg->msg_private, msg->msg_len,
4616                                   msg->msg_type);
4617
4618                 msg->msg_no_resend = true;
4619                 /*
4620                  * NB: message will not generate event because w/o attached MD,
4621                  * but we still should give error code so lnet_msg_decommit()
4622                  * can skip counters operations and other checks.
4623                  */
4624                 lnet_finalize(msg, -ENOENT);
4625         }
4626 }
4627
4628 void
4629 lnet_recv_delayed_msg_list(struct list_head *head)
4630 {
4631         while (!list_empty(head)) {
4632                 struct lnet_msg *msg;
4633                 struct lnet_process_id id;
4634
4635                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4636                 list_del(&msg->msg_list);
4637
4638                 /* md won't disappear under me, since each msg
4639                  * holds a ref on it */
4640
4641                 id.nid = msg->msg_hdr.src_nid;
4642                 id.pid = msg->msg_hdr.src_pid;
4643
4644                 LASSERT(msg->msg_rx_delayed);
4645                 LASSERT(msg->msg_md != NULL);
4646                 LASSERT(msg->msg_rxpeer != NULL);
4647                 LASSERT(msg->msg_rxni != NULL);
4648                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4649
4650                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4651                        "match %llu offset %d length %d.\n",
4652                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4653                         msg->msg_hdr.msg.put.match_bits,
4654                         msg->msg_hdr.msg.put.offset,
4655                         msg->msg_hdr.payload_length);
4656
4657                 lnet_recv_put(msg->msg_rxni, msg);
4658         }
4659 }
4660
4661 static void
4662 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4663                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4664 {
4665         s64 timeout_ns;
4666         bool new_entry = true;
4667         struct lnet_rsp_tracker *local_rspt;
4668
4669         /*
4670          * MD has a refcount taken by message so it's not going away.
4671          * The MD however can be looked up. We need to secure the access
4672          * to the md_rspt_ptr by taking the res_lock.
4673          * The rspt can be accessed without protection up to when it gets
4674          * added to the list.
4675          */
4676
4677         lnet_res_lock(cpt);
4678         local_rspt = md->md_rspt_ptr;
4679         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4680         if (local_rspt != NULL) {
4681                 /*
4682                  * we already have an rspt attached to the md, so we'll
4683                  * update the deadline on that one.
4684                  */
4685                 lnet_rspt_free(rspt, cpt);
4686                 new_entry = false;
4687         } else {
4688                 /* new md */
4689                 rspt->rspt_mdh = mdh;
4690                 rspt->rspt_cpt = cpt;
4691                 /* store the rspt so we can access it when we get the REPLY */
4692                 md->md_rspt_ptr = rspt;
4693                 local_rspt = rspt;
4694         }
4695         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4696
4697         /*
4698          * add to the list of tracked responses. It's added to tail of the
4699          * list in order to expire all the older entries first.
4700          */
4701         lnet_net_lock(cpt);
4702         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4703                 list_del_init(&local_rspt->rspt_on_list);
4704         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4705         lnet_net_unlock(cpt);
4706         lnet_res_unlock(cpt);
4707 }
4708
4709 /**
4710  * Initiate an asynchronous PUT operation.
4711  *
4712  * There are several events associated with a PUT: completion of the send on
4713  * the initiator node (LNET_EVENT_SEND), and when the send completes
4714  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4715  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4716  * used at the target node to indicate the completion of incoming data
4717  * delivery.
4718  *
4719  * The local events will be logged in the EQ associated with the MD pointed to
4720  * by \a mdh handle. Using a MD without an associated EQ results in these
4721  * events being discarded. In this case, the caller must have another
4722  * mechanism (e.g., a higher level protocol) for determining when it is safe
4723  * to modify the memory region associated with the MD.
4724  *
4725  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4726  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4727  *
4728  * \param self Indicates the NID of a local interface through which to send
4729  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4730  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4731  * must be "free floating" (See LNetMDBind()).
4732  * \param ack Controls whether an acknowledgment is requested.
4733  * Acknowledgments are only sent when they are requested by the initiating
4734  * process and the target MD enables them.
4735  * \param target A process identifier for the target process.
4736  * \param portal The index in the \a target's portal table.
4737  * \param match_bits The match bits to use for MD selection at the target
4738  * process.
4739  * \param offset The offset into the target MD (only used when the target
4740  * MD has the LNET_MD_MANAGE_REMOTE option set).
4741  * \param hdr_data 64 bits of user data that can be included in the message
4742  * header. This data is written to an event queue entry at the target if an
4743  * EQ is present on the matching MD.
4744  *
4745  * \retval  0      Success, and only in this case events will be generated
4746  * and logged to EQ (if it exists).
4747  * \retval -EIO    Simulated failure.
4748  * \retval -ENOMEM Memory allocation failure.
4749  * \retval -ENOENT Invalid MD object.
4750  *
4751  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4752  */
4753 int
4754 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4755         struct lnet_process_id target, unsigned int portal,
4756         __u64 match_bits, unsigned int offset,
4757         __u64 hdr_data)
4758 {
4759         struct lnet_msg *msg;
4760         struct lnet_libmd *md;
4761         int cpt;
4762         int rc;
4763         struct lnet_rsp_tracker *rspt = NULL;
4764
4765         LASSERT(the_lnet.ln_refcount > 0);
4766
4767         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4768             fail_peer(target.nid, 1)) {                 /* shall we now? */
4769                 CERROR("Dropping PUT to %s: simulated failure\n",
4770                        libcfs_id2str(target));
4771                 return -EIO;
4772         }
4773
4774         msg = lnet_msg_alloc();
4775         if (msg == NULL) {
4776                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4777                        libcfs_id2str(target));
4778                 return -ENOMEM;
4779         }
4780         msg->msg_vmflush = !!memory_pressure_get();
4781
4782         cpt = lnet_cpt_of_cookie(mdh.cookie);
4783
4784         if (ack == LNET_ACK_REQ) {
4785                 rspt = lnet_rspt_alloc(cpt);
4786                 if (!rspt) {
4787                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4788                                 libcfs_id2str(target));
4789                         return -ENOMEM;
4790                 }
4791                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4792         }
4793
4794         lnet_res_lock(cpt);
4795
4796         md = lnet_handle2md(&mdh);
4797         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4798                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4799                        match_bits, portal, libcfs_id2str(target),
4800                        md == NULL ? -1 : md->md_threshold);
4801                 if (md != NULL && md->md_me != NULL)
4802                         CERROR("Source MD also attached to portal %d\n",
4803                                md->md_me->me_portal);
4804                 lnet_res_unlock(cpt);
4805
4806                 lnet_rspt_free(rspt, cpt);
4807                 lnet_msg_free(msg);
4808                 return -ENOENT;
4809         }
4810
4811         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4812
4813         lnet_msg_attach_md(msg, md, 0, 0);
4814
4815         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4816
4817         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4818         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4819         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4820         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4821
4822         /* NB handles only looked up by creator (no flips) */
4823         if (ack == LNET_ACK_REQ) {
4824                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4825                         the_lnet.ln_interface_cookie;
4826                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4827                         md->md_lh.lh_cookie;
4828         } else {
4829                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4830                         LNET_WIRE_HANDLE_COOKIE_NONE;
4831                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4832                         LNET_WIRE_HANDLE_COOKIE_NONE;
4833         }
4834
4835         lnet_res_unlock(cpt);
4836
4837         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4838
4839         if (ack == LNET_ACK_REQ)
4840                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4841
4842         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4843                                  CFS_FAIL_ONCE))
4844                 rc = -EIO;
4845         else
4846                 rc = lnet_send(self, msg, LNET_NID_ANY);
4847
4848         if (rc != 0) {
4849                 CNETERR("Error sending PUT to %s: %d\n",
4850                         libcfs_id2str(target), rc);
4851                 msg->msg_no_resend = true;
4852                 lnet_finalize(msg, rc);
4853         }
4854
4855         /* completion will be signalled by an event */
4856         return 0;
4857 }
4858 EXPORT_SYMBOL(LNetPut);
4859
4860 /*
4861  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4862  * returns a msg for the LND to pass to lnet_finalize() when the sink
4863  * data has been received.
4864  *
4865  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4866  * lnet_finalize() is called on it, so the LND must call this first
4867  */
4868 struct lnet_msg *
4869 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4870 {
4871         struct lnet_msg *msg = lnet_msg_alloc();
4872         struct lnet_libmd *getmd = getmsg->msg_md;
4873         struct lnet_process_id peer_id = getmsg->msg_target;
4874         int cpt;
4875
4876         LASSERT(!getmsg->msg_target_is_router);
4877         LASSERT(!getmsg->msg_routing);
4878
4879         if (msg == NULL) {
4880                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4881                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4882                 goto drop;
4883         }
4884
4885         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4886         lnet_res_lock(cpt);
4887
4888         LASSERT(getmd->md_refcount > 0);
4889
4890         if (getmd->md_threshold == 0) {
4891                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4892                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4893                         getmd);
4894                 lnet_res_unlock(cpt);
4895                 goto drop;
4896         }
4897
4898         LASSERT(getmd->md_offset == 0);
4899
4900         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4901                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4902
4903         /* setup information for lnet_build_msg_event */
4904         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4905         msg->msg_from = peer_id.nid;
4906         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4907         msg->msg_hdr.src_nid = peer_id.nid;
4908         msg->msg_hdr.payload_length = getmd->md_length;
4909         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4910
4911         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4912         lnet_res_unlock(cpt);
4913
4914         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4915
4916         lnet_net_lock(cpt);
4917         lnet_msg_commit(msg, cpt);
4918         lnet_net_unlock(cpt);
4919
4920         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4921
4922         return msg;
4923
4924  drop:
4925         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4926
4927         lnet_net_lock(cpt);
4928         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4929         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4930         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4931                 getmd->md_length;
4932         lnet_net_unlock(cpt);
4933
4934         if (msg != NULL)
4935                 lnet_msg_free(msg);
4936
4937         return NULL;
4938 }
4939 EXPORT_SYMBOL(lnet_create_reply_msg);
4940
4941 void
4942 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4943                        unsigned int len)
4944 {
4945         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4946          * completed and I know it. */
4947         LASSERT(reply != NULL);
4948         LASSERT(reply->msg_type == LNET_MSG_GET);
4949         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4950
4951         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4952          * the end of my buffer, I might as well be dead. */
4953         LASSERT(len <= reply->msg_ev.mlength);
4954
4955         reply->msg_ev.mlength = len;
4956 }
4957 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4958
4959 /**
4960  * Initiate an asynchronous GET operation.
4961  *
4962  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4963  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4964  * the target node in the REPLY has been written to local MD.
4965  *
4966  * On the target node, an LNET_EVENT_GET is logged when the GET request
4967  * arrives and is accepted into a MD.
4968  *
4969  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4970  * \param mdh A handle for the MD that describes the memory into which the
4971  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4972  *
4973  * \retval  0      Success, and only in this case events will be generated
4974  * and logged to EQ (if it exists) of the MD.
4975  * \retval -EIO    Simulated failure.
4976  * \retval -ENOMEM Memory allocation failure.
4977  * \retval -ENOENT Invalid MD object.
4978  */
4979 int
4980 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4981         struct lnet_process_id target, unsigned int portal,
4982         __u64 match_bits, unsigned int offset, bool recovery)
4983 {
4984         struct lnet_msg *msg;
4985         struct lnet_libmd *md;
4986         struct lnet_rsp_tracker *rspt;
4987         int cpt;
4988         int rc;
4989
4990         LASSERT(the_lnet.ln_refcount > 0);
4991
4992         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4993             fail_peer(target.nid, 1))                   /* shall we now? */
4994         {
4995                 CERROR("Dropping GET to %s: simulated failure\n",
4996                        libcfs_id2str(target));
4997                 return -EIO;
4998         }
4999
5000         msg = lnet_msg_alloc();
5001         if (!msg) {
5002                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5003                        libcfs_id2str(target));
5004                 return -ENOMEM;
5005         }
5006
5007         cpt = lnet_cpt_of_cookie(mdh.cookie);
5008
5009         rspt = lnet_rspt_alloc(cpt);
5010         if (!rspt) {
5011                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5012                        libcfs_id2str(target));
5013                 return -ENOMEM;
5014         }
5015         INIT_LIST_HEAD(&rspt->rspt_on_list);
5016
5017         msg->msg_recovery = recovery;
5018
5019         lnet_res_lock(cpt);
5020
5021         md = lnet_handle2md(&mdh);
5022         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5023                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5024                        match_bits, portal, libcfs_id2str(target),
5025                        md == NULL ? -1 : md->md_threshold);
5026                 if (md != NULL && md->md_me != NULL)
5027                         CERROR("REPLY MD also attached to portal %d\n",
5028                                md->md_me->me_portal);
5029
5030                 lnet_res_unlock(cpt);
5031
5032                 lnet_msg_free(msg);
5033                 lnet_rspt_free(rspt, cpt);
5034                 return -ENOENT;
5035         }
5036
5037         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5038
5039         lnet_msg_attach_md(msg, md, 0, 0);
5040
5041         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5042
5043         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5044         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5045         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5046         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5047
5048         /* NB handles only looked up by creator (no flips) */
5049         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5050                 the_lnet.ln_interface_cookie;
5051         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5052                 md->md_lh.lh_cookie;
5053
5054         lnet_res_unlock(cpt);
5055
5056         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5057
5058         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5059
5060         rc = lnet_send(self, msg, LNET_NID_ANY);
5061         if (rc < 0) {
5062                 CNETERR("Error sending GET to %s: %d\n",
5063                         libcfs_id2str(target), rc);
5064                 msg->msg_no_resend = true;
5065                 lnet_finalize(msg, rc);
5066         }
5067
5068         /* completion will be signalled by an event */
5069         return 0;
5070 }
5071 EXPORT_SYMBOL(LNetGet);
5072
5073 /**
5074  * Calculate distance to node at \a dstnid.
5075  *
5076  * \param dstnid Target NID.
5077  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5078  * is saved here.
5079  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5080  * here.
5081  *
5082  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5083  * local_nid_dist_zero is set, which is the default.
5084  * \retval positives Distance to target NID, i.e. number of hops plus one.
5085  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5086  */
5087 int
5088 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5089 {
5090         struct list_head        *e;
5091         struct lnet_ni *ni = NULL;
5092         struct lnet_remotenet *rnet;
5093         __u32                   dstnet = LNET_NIDNET(dstnid);
5094         int                     hops;
5095         int                     cpt;
5096         __u32                   order = 2;
5097         struct list_head        *rn_list;
5098
5099         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5100          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5101          * keep order 0 free for 0@lo and order 1 free for a local NID
5102          * match */
5103
5104         LASSERT(the_lnet.ln_refcount > 0);
5105
5106         cpt = lnet_net_lock_current();
5107
5108         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5109                 if (ni->ni_nid == dstnid) {
5110                         if (srcnidp != NULL)
5111                                 *srcnidp = dstnid;
5112                         if (orderp != NULL) {
5113                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5114                                         *orderp = 0;
5115                                 else
5116                                         *orderp = 1;
5117                         }
5118                         lnet_net_unlock(cpt);
5119
5120                         return local_nid_dist_zero ? 0 : 1;
5121                 }
5122
5123                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5124                         /* Check if ni was originally created in
5125                          * current net namespace.
5126                          * If not, assign order above 0xffff0000,
5127                          * to make this ni not a priority. */
5128                         if (current->nsproxy &&
5129                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5130                                         order += 0xffff0000;
5131                         if (srcnidp != NULL)
5132                                 *srcnidp = ni->ni_nid;
5133                         if (orderp != NULL)
5134                                 *orderp = order;
5135                         lnet_net_unlock(cpt);
5136                         return 1;
5137                 }
5138
5139                 order++;
5140         }
5141
5142         rn_list = lnet_net2rnethash(dstnet);
5143         list_for_each(e, rn_list) {
5144                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5145
5146                 if (rnet->lrn_net == dstnet) {
5147                         struct lnet_route *route;
5148                         struct lnet_route *shortest = NULL;
5149                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5150                         __u32 route_hops;
5151
5152                         LASSERT(!list_empty(&rnet->lrn_routes));
5153
5154                         list_for_each_entry(route, &rnet->lrn_routes,
5155                                             lr_list) {
5156                                 route_hops = route->lr_hops;
5157                                 if (route_hops == LNET_UNDEFINED_HOPS)
5158                                         route_hops = 1;
5159                                 if (shortest == NULL ||
5160                                     route_hops < shortest_hops) {
5161                                         shortest = route;
5162                                         shortest_hops = route_hops;
5163                                 }
5164                         }
5165
5166                         LASSERT(shortest != NULL);
5167                         hops = shortest_hops;
5168                         if (srcnidp != NULL) {
5169                                 struct lnet_net *net;
5170                                 net = lnet_get_net_locked(shortest->lr_lnet);
5171                                 LASSERT(net);
5172                                 ni = lnet_get_next_ni_locked(net, NULL);
5173                                 *srcnidp = ni->ni_nid;
5174                         }
5175                         if (orderp != NULL)
5176                                 *orderp = order;
5177                         lnet_net_unlock(cpt);
5178                         return hops + 1;
5179                 }
5180                 order++;
5181         }
5182
5183         lnet_net_unlock(cpt);
5184         return -EHOSTUNREACH;
5185 }
5186 EXPORT_SYMBOL(LNetDist);