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 bool
  67 lnet_msg_is_response(struct lnet_msg *msg)
  68 {
  69         return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
  70 }
  71
  72 static inline bool
  73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
  74 {
  75         if (md_options & LNET_MD_NO_TRACK_RESPONSE)
  76                 /* Explicitly disabled in MD options */
  77                 return false;
  78
  79         if (md_options & LNET_MD_TRACK_RESPONSE)
  80                 /* Explicity enabled in MD options */
  81                 return true;
  82
  83         if (lnet_response_tracking == 3)
  84                 /* Enabled for all message types */
  85                 return true;
  86
  87         if (msg_type == LNET_MSG_PUT)
  88                 return lnet_response_tracking == 2;
  89
  90         if (msg_type == LNET_MSG_GET)
  91                 return lnet_response_tracking == 1;
  92
  93         return false;
  94 }
  95
  96 static inline struct lnet_comm_count *
  97 get_stats_counts(struct lnet_element_stats *stats,
  98                  enum lnet_stats_type stats_type)
  99 {
 100         switch (stats_type) {
 101         case LNET_STATS_TYPE_SEND:
 102                 return &stats->el_send_stats;
 103         case LNET_STATS_TYPE_RECV:
 104                 return &stats->el_recv_stats;
 105         case LNET_STATS_TYPE_DROP:
 106                 return &stats->el_drop_stats;
 107         default:
 108                 CERROR("Unknown stats type\n");
 109         }
 110
 111         return NULL;
 112 }
 113
 114 void lnet_incr_stats(struct lnet_element_stats *stats,
 115                      enum lnet_msg_type msg_type,
 116                      enum lnet_stats_type stats_type)
 117 {
 118         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 119         if (!counts)
 120                 return;
 121
 122         switch (msg_type) {
 123         case LNET_MSG_ACK:
 124                 atomic_inc(&counts->co_ack_count);
 125                 break;
 126         case LNET_MSG_PUT:
 127                 atomic_inc(&counts->co_put_count);
 128                 break;
 129         case LNET_MSG_GET:
 130                 atomic_inc(&counts->co_get_count);
 131                 break;
 132         case LNET_MSG_REPLY:
 133                 atomic_inc(&counts->co_reply_count);
 134                 break;
 135         case LNET_MSG_HELLO:
 136                 atomic_inc(&counts->co_hello_count);
 137                 break;
 138         default:
 139                 CERROR("There is a BUG in the code. Unknown message type\n");
 140                 break;
 141         }
 142 }
 143
 144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 145                      enum lnet_stats_type stats_type)
 146 {
 147         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 148         if (!counts)
 149                 return 0;
 150
 151         return (atomic_read(&counts->co_ack_count) +
 152                 atomic_read(&counts->co_put_count) +
 153                 atomic_read(&counts->co_get_count) +
 154                 atomic_read(&counts->co_reply_count) +
 155                 atomic_read(&counts->co_hello_count));
 156 }
 157
 158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 159                                 struct lnet_comm_count *counts)
 160 {
 161         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 162         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 163         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 164         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 165         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 166 }
 167
 168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 169                               struct lnet_element_stats *stats)
 170 {
 171         struct lnet_comm_count *counts;
 172
 173         LASSERT(msg_stats);
 174         LASSERT(stats);
 175
 176         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 177         if (!counts)
 178                 return;
 179         assign_stats(&msg_stats->im_send_stats, counts);
 180
 181         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 182         if (!counts)
 183                 return;
 184         assign_stats(&msg_stats->im_recv_stats, counts);
 185
 186         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 187         if (!counts)
 188                 return;
 189         assign_stats(&msg_stats->im_drop_stats, counts);
 190 }
 191
 192 int
 193 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 194 {
 195         struct lnet_test_peer *tp;
 196         struct list_head *el;
 197         struct list_head *next;
 198         LIST_HEAD(cull);
 199
 200         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 201         if (threshold != 0) {
 202                 /* Adding a new entry */
 203                 LIBCFS_ALLOC(tp, sizeof(*tp));
 204                 if (tp == NULL)
 205                         return -ENOMEM;
 206
 207                 tp->tp_nid = nid;
 208                 tp->tp_threshold = threshold;
 209
 210                 lnet_net_lock(0);
 211                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 212                 lnet_net_unlock(0);
 213                 return 0;
 214         }
 215
 216         lnet_net_lock(0);
 217
 218         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 219                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 220
 221                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 222                     nid == LNET_NID_ANY ||      /* removing all entries */
 223                     tp->tp_nid == nid) {        /* matched this one */
 224                         list_move(&tp->tp_list, &cull);
 225                 }
 226         }
 227
 228         lnet_net_unlock(0);
 229
 230         while (!list_empty(&cull)) {
 231                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 232
 233                 list_del(&tp->tp_list);
 234                 LIBCFS_FREE(tp, sizeof(*tp));
 235         }
 236         return 0;
 237 }
 238
 239 static int
 240 fail_peer (lnet_nid_t nid, int outgoing)
 241 {
 242         struct lnet_test_peer *tp;
 243         struct list_head *el;
 244         struct list_head *next;
 245         LIST_HEAD(cull);
 246         int fail = 0;
 247
 248         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 249         lnet_net_lock(0);
 250
 251         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 252                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 253
 254                 if (tp->tp_threshold == 0) {
 255                         /* zombie entry */
 256                         if (outgoing) {
 257                                 /* only cull zombies on outgoing tests,
 258                                  * since we may be at interrupt priority on
 259                                  * incoming messages. */
 260                                 list_move(&tp->tp_list, &cull);
 261                         }
 262                         continue;
 263                 }
 264
 265                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 266                     nid == tp->tp_nid) {                /* fail this peer */
 267                         fail = 1;
 268
 269                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 270                                 tp->tp_threshold--;
 271                                 if (outgoing &&
 272                                     tp->tp_threshold == 0) {
 273                                         /* see above */
 274                                         list_move(&tp->tp_list, &cull);
 275                                 }
 276                         }
 277                         break;
 278                 }
 279         }
 280
 281         lnet_net_unlock(0);
 282
 283         while (!list_empty(&cull)) {
 284                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 285                 list_del(&tp->tp_list);
 286
 287                 LIBCFS_FREE(tp, sizeof(*tp));
 288         }
 289
 290         return fail;
 291 }
 292
 293 unsigned int
 294 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 295 {
 296         unsigned int nob = 0;
 297
 298         LASSERT(niov == 0 || iov != NULL);
 299         while (niov-- > 0)
 300                 nob += (iov++)->iov_len;
 301
 302         return (nob);
 303 }
 304 EXPORT_SYMBOL(lnet_iov_nob);
 305
 306 void
 307 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 308                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 309                   unsigned int nob)
 310 {
 311         /* NB diov, siov are READ-ONLY */
 312         unsigned int this_nob;
 313
 314         if (nob == 0)
 315                 return;
 316
 317         /* skip complete frags before 'doffset' */
 318         LASSERT(ndiov > 0);
 319         while (doffset >= diov->iov_len) {
 320                 doffset -= diov->iov_len;
 321                 diov++;
 322                 ndiov--;
 323                 LASSERT(ndiov > 0);
 324         }
 325
 326         /* skip complete frags before 'soffset' */
 327         LASSERT(nsiov > 0);
 328         while (soffset >= siov->iov_len) {
 329                 soffset -= siov->iov_len;
 330                 siov++;
 331                 nsiov--;
 332                 LASSERT(nsiov > 0);
 333         }
 334
 335         do {
 336                 LASSERT(ndiov > 0);
 337                 LASSERT(nsiov > 0);
 338                 this_nob = min3((unsigned int)diov->iov_len - doffset,
 339                                 (unsigned int)siov->iov_len - soffset,
 340                                 nob);
 341
 342                 memcpy((char *)diov->iov_base + doffset,
 343                        (char *)siov->iov_base + soffset, this_nob);
 344                 nob -= this_nob;
 345
 346                 if (diov->iov_len > doffset + this_nob) {
 347                         doffset += this_nob;
 348                 } else {
 349                         diov++;
 350                         ndiov--;
 351                         doffset = 0;
 352                 }
 353
 354                 if (siov->iov_len > soffset + this_nob) {
 355                         soffset += this_nob;
 356                 } else {
 357                         siov++;
 358                         nsiov--;
 359                         soffset = 0;
 360                 }
 361         } while (nob > 0);
 362 }
 363 EXPORT_SYMBOL(lnet_copy_iov2iov);
 364
 365 unsigned int
 366 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 367 {
 368         unsigned int  nob = 0;
 369
 370         LASSERT(niov == 0 || kiov != NULL);
 371         while (niov-- > 0)
 372                 nob += (kiov++)->bv_len;
 373
 374         return (nob);
 375 }
 376 EXPORT_SYMBOL(lnet_kiov_nob);
 377
 378 void
 379 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
 380                     unsigned int doffset,
 381                     unsigned int nsiov, struct bio_vec *siov,
 382                     unsigned int soffset,
 383                     unsigned int nob)
 384 {
 385         /* NB diov, siov are READ-ONLY */
 386         unsigned int    this_nob;
 387         char           *daddr = NULL;
 388         char           *saddr = NULL;
 389
 390         if (nob == 0)
 391                 return;
 392
 393         LASSERT (!in_interrupt ());
 394
 395         LASSERT (ndiov > 0);
 396         while (doffset >= diov->bv_len) {
 397                 doffset -= diov->bv_len;
 398                 diov++;
 399                 ndiov--;
 400                 LASSERT(ndiov > 0);
 401         }
 402
 403         LASSERT(nsiov > 0);
 404         while (soffset >= siov->bv_len) {
 405                 soffset -= siov->bv_len;
 406                 siov++;
 407                 nsiov--;
 408                 LASSERT(nsiov > 0);
 409         }
 410
 411         do {
 412                 LASSERT(ndiov > 0);
 413                 LASSERT(nsiov > 0);
 414                 this_nob = min3(diov->bv_len - doffset,
 415                                 siov->bv_len - soffset,
 416                                 nob);
 417
 418                 if (daddr == NULL)
 419                         daddr = ((char *)kmap(diov->bv_page)) +
 420                                 diov->bv_offset + doffset;
 421                 if (saddr == NULL)
 422                         saddr = ((char *)kmap(siov->bv_page)) +
 423                                 siov->bv_offset + soffset;
 424
 425                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 426                  * However in practice at least one of the kiovs will be mapped
 427                  * kernel pages and the map/unmap will be NOOPs */
 428
 429                 memcpy (daddr, saddr, this_nob);
 430                 nob -= this_nob;
 431
 432                 if (diov->bv_len > doffset + this_nob) {
 433                         daddr += this_nob;
 434                         doffset += this_nob;
 435                 } else {
 436                         kunmap(diov->bv_page);
 437                         daddr = NULL;
 438                         diov++;
 439                         ndiov--;
 440                         doffset = 0;
 441                 }
 442
 443                 if (siov->bv_len > soffset + this_nob) {
 444                         saddr += this_nob;
 445                         soffset += this_nob;
 446                 } else {
 447                         kunmap(siov->bv_page);
 448                         saddr = NULL;
 449                         siov++;
 450                         nsiov--;
 451                         soffset = 0;
 452                 }
 453         } while (nob > 0);
 454
 455         if (daddr != NULL)
 456                 kunmap(diov->bv_page);
 457         if (saddr != NULL)
 458                 kunmap(siov->bv_page);
 459 }
 460 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 461
 462 void
 463 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 464                     unsigned int nkiov, struct bio_vec *kiov,
 465                     unsigned int kiovoffset,
 466                     unsigned int nob)
 467 {
 468         /* NB iov, kiov are READ-ONLY */
 469         unsigned int    this_nob;
 470         char           *addr = NULL;
 471
 472         if (nob == 0)
 473                 return;
 474
 475         LASSERT (!in_interrupt ());
 476
 477         LASSERT (niov > 0);
 478         while (iovoffset >= iov->iov_len) {
 479                 iovoffset -= iov->iov_len;
 480                 iov++;
 481                 niov--;
 482                 LASSERT(niov > 0);
 483         }
 484
 485         LASSERT(nkiov > 0);
 486         while (kiovoffset >= kiov->bv_len) {
 487                 kiovoffset -= kiov->bv_len;
 488                 kiov++;
 489                 nkiov--;
 490                 LASSERT(nkiov > 0);
 491         }
 492
 493         do {
 494                 LASSERT(niov > 0);
 495                 LASSERT(nkiov > 0);
 496                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 497                                 (unsigned int)kiov->bv_len - kiovoffset,
 498                                 nob);
 499
 500                 if (addr == NULL)
 501                         addr = ((char *)kmap(kiov->bv_page)) +
 502                                 kiov->bv_offset + kiovoffset;
 503
 504                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 505                 nob -= this_nob;
 506
 507                 if (iov->iov_len > iovoffset + this_nob) {
 508                         iovoffset += this_nob;
 509                 } else {
 510                         iov++;
 511                         niov--;
 512                         iovoffset = 0;
 513                 }
 514
 515                 if (kiov->bv_len > kiovoffset + this_nob) {
 516                         addr += this_nob;
 517                         kiovoffset += this_nob;
 518                 } else {
 519                         kunmap(kiov->bv_page);
 520                         addr = NULL;
 521                         kiov++;
 522                         nkiov--;
 523                         kiovoffset = 0;
 524                 }
 525
 526         } while (nob > 0);
 527
 528         if (addr != NULL)
 529                 kunmap(kiov->bv_page);
 530 }
 531 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 532
 533 void
 534 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
 535                    unsigned int kiovoffset,
 536                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 537                    unsigned int nob)
 538 {
 539         /* NB kiov, iov are READ-ONLY */
 540         unsigned int    this_nob;
 541         char           *addr = NULL;
 542
 543         if (nob == 0)
 544                 return;
 545
 546         LASSERT (!in_interrupt ());
 547
 548         LASSERT (nkiov > 0);
 549         while (kiovoffset >= kiov->bv_len) {
 550                 kiovoffset -= kiov->bv_len;
 551                 kiov++;
 552                 nkiov--;
 553                 LASSERT(nkiov > 0);
 554         }
 555
 556         LASSERT(niov > 0);
 557         while (iovoffset >= iov->iov_len) {
 558                 iovoffset -= iov->iov_len;
 559                 iov++;
 560                 niov--;
 561                 LASSERT(niov > 0);
 562         }
 563
 564         do {
 565                 LASSERT(nkiov > 0);
 566                 LASSERT(niov > 0);
 567                 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
 568                                 (unsigned int)iov->iov_len - iovoffset,
 569                                 nob);
 570
 571                 if (addr == NULL)
 572                         addr = ((char *)kmap(kiov->bv_page)) +
 573                                 kiov->bv_offset + kiovoffset;
 574
 575                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 576                 nob -= this_nob;
 577
 578                 if (kiov->bv_len > kiovoffset + this_nob) {
 579                         addr += this_nob;
 580                         kiovoffset += this_nob;
 581                 } else {
 582                         kunmap(kiov->bv_page);
 583                         addr = NULL;
 584                         kiov++;
 585                         nkiov--;
 586                         kiovoffset = 0;
 587                 }
 588
 589                 if (iov->iov_len > iovoffset + this_nob) {
 590                         iovoffset += this_nob;
 591                 } else {
 592                         iov++;
 593                         niov--;
 594                         iovoffset = 0;
 595                 }
 596         } while (nob > 0);
 597
 598         if (addr != NULL)
 599                 kunmap(kiov->bv_page);
 600 }
 601 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 602
 603 int
 604 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
 605                   int src_niov, struct bio_vec *src,
 606                   unsigned int offset, unsigned int len)
 607 {
 608         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 609          * for exactly 'len' bytes, and return the number of entries.
 610          * NB not destructive to 'src' */
 611         unsigned int    frag_len;
 612         unsigned int    niov;
 613
 614         if (len == 0)                           /* no data => */
 615                 return (0);                     /* no frags */
 616
 617         LASSERT(src_niov > 0);
 618         while (offset >= src->bv_len) {      /* skip initial frags */
 619                 offset -= src->bv_len;
 620                 src_niov--;
 621                 src++;
 622                 LASSERT(src_niov > 0);
 623         }
 624
 625         niov = 1;
 626         for (;;) {
 627                 LASSERT(src_niov > 0);
 628                 LASSERT((int)niov <= dst_niov);
 629
 630                 frag_len = src->bv_len - offset;
 631                 dst->bv_page = src->bv_page;
 632                 dst->bv_offset = src->bv_offset + offset;
 633
 634                 if (len <= frag_len) {
 635                         dst->bv_len = len;
 636                         LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 637                         return niov;
 638                 }
 639
 640                 dst->bv_len = frag_len;
 641                 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 642
 643                 len -= frag_len;
 644                 dst++;
 645                 src++;
 646                 niov++;
 647                 src_niov--;
 648                 offset = 0;
 649         }
 650 }
 651 EXPORT_SYMBOL(lnet_extract_kiov);
 652
 653 void
 654 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 655              int delayed, unsigned int offset, unsigned int mlen,
 656              unsigned int rlen)
 657 {
 658         unsigned int niov = 0;
 659         struct kvec *iov = NULL;
 660         struct bio_vec  *kiov = NULL;
 661         int rc;
 662
 663         LASSERT (!in_interrupt ());
 664         LASSERT (mlen == 0 || msg != NULL);
 665
 666         if (msg != NULL) {
 667                 LASSERT(msg->msg_receiving);
 668                 LASSERT(!msg->msg_sending);
 669                 LASSERT(rlen == msg->msg_len);
 670                 LASSERT(mlen <= msg->msg_len);
 671                 LASSERT(msg->msg_offset == offset);
 672                 LASSERT(msg->msg_wanted == mlen);
 673
 674                 msg->msg_receiving = 0;
 675
 676                 if (mlen != 0) {
 677                         niov = msg->msg_niov;
 678                         kiov = msg->msg_kiov;
 679
 680                         LASSERT (niov > 0);
 681                         LASSERT ((iov == NULL) != (kiov == NULL));
 682                 }
 683         }
 684
 685         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 686                                              niov, kiov, offset, mlen,
 687                                              rlen);
 688         if (rc < 0)
 689                 lnet_finalize(msg, rc);
 690 }
 691
 692 static void
 693 lnet_setpayloadbuffer(struct lnet_msg *msg)
 694 {
 695         struct lnet_libmd *md = msg->msg_md;
 696
 697         LASSERT(msg->msg_len > 0);
 698         LASSERT(!msg->msg_routing);
 699         LASSERT(md != NULL);
 700         LASSERT(msg->msg_niov == 0);
 701         LASSERT(msg->msg_kiov == NULL);
 702
 703         msg->msg_niov = md->md_niov;
 704         msg->msg_kiov = md->md_kiov;
 705 }
 706
 707 void
 708 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 709                unsigned int offset, unsigned int len)
 710 {
 711         msg->msg_type = type;
 712         msg->msg_target = target;
 713         msg->msg_len = len;
 714         msg->msg_offset = offset;
 715
 716         if (len != 0)
 717                 lnet_setpayloadbuffer(msg);
 718
 719         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 720         msg->msg_hdr.type           = cpu_to_le32(type);
 721         /* dest_nid will be overwritten by lnet_select_pathway() */
 722         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 723         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 724         /* src_nid will be set later */
 725         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 726         msg->msg_hdr.payload_length = cpu_to_le32(len);
 727 }
 728
 729 static void
 730 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 731 {
 732         void *priv = msg->msg_private;
 733         int rc;
 734
 735         LASSERT(!in_interrupt());
 736         LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
 737                 (msg->msg_txcredit && msg->msg_peertxcredit));
 738
 739         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 740         if (rc < 0) {
 741                 msg->msg_no_resend = true;
 742                 lnet_finalize(msg, rc);
 743         }
 744 }
 745
 746 static int
 747 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 748 {
 749         int     rc;
 750
 751         LASSERT(!msg->msg_sending);
 752         LASSERT(msg->msg_receiving);
 753         LASSERT(!msg->msg_rx_ready_delay);
 754         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 755
 756         msg->msg_rx_ready_delay = 1;
 757         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 758                                                   &msg->msg_private);
 759         if (rc != 0) {
 760                 CERROR("recv from %s / send to %s aborted: "
 761                        "eager_recv failed %d\n",
 762                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 763                        libcfs_id2str(msg->msg_target), rc);
 764                 LASSERT(rc < 0); /* required by my callers */
 765         }
 766
 767         return rc;
 768 }
 769
 770 static bool
 771 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 772 {
 773         time64_t deadline;
 774
 775         deadline = lpni->lpni_last_alive +
 776                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 777
 778         /*
 779          * assume peer_ni is alive as long as we're within the configured
 780          * peer timeout
 781          */
 782         if (deadline > now)
 783                 return false;
 784
 785         return true;
 786 }
 787
 788 /* NB: returns 1 when alive, 0 when dead, negative when error;
 789  *     may drop the lnet_net_lock */
 790 static int
 791 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 792                        struct lnet_msg *msg)
 793 {
 794         time64_t now = ktime_get_seconds();
 795
 796         if (!lnet_peer_aliveness_enabled(lpni))
 797                 return -ENODEV;
 798
 799         /*
 800          * If we're resending a message, let's attempt to send it even if
 801          * the peer is down to fulfill our resend quota on the message
 802          */
 803         if (msg->msg_retry_count > 0)
 804                 return 1;
 805
 806         /* try and send recovery messages irregardless */
 807         if (msg->msg_recovery)
 808                 return 1;
 809
 810         /* always send any responses */
 811         if (lnet_msg_is_response(msg))
 812                 return 1;
 813
 814         if (!lnet_is_peer_deadline_passed(lpni, now))
 815                 return true;
 816
 817         return lnet_is_peer_ni_alive(lpni);
 818 }
 819
 820 /**
 821  * \param msg The message to be sent.
 822  * \param do_send True if lnet_ni_send() should be called in this function.
 823  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 824  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 825  *
 826  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 827  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 828  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 829  * \retval -ECANCELED If the MD of the message has been unlinked.
 830  */
 831 static int
 832 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 833 {
 834         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 835         struct lnet_ni          *ni = msg->msg_txni;
 836         int                     cpt = msg->msg_tx_cpt;
 837         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 838
 839         /* non-lnet_send() callers have checked before */
 840         LASSERT(!do_send || msg->msg_tx_delayed);
 841         LASSERT(!msg->msg_receiving);
 842         LASSERT(msg->msg_tx_committed);
 843         /* can't get here if we're sending to the loopback interface */
 844         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 845
 846         /* NB 'lp' is always the next hop */
 847         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 848             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 849                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 850                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 851                         msg->msg_len;
 852                 lnet_net_unlock(cpt);
 853                 if (msg->msg_txpeer)
 854                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 855                                         msg->msg_type,
 856                                         LNET_STATS_TYPE_DROP);
 857                 if (msg->msg_txni)
 858                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 859                                         msg->msg_type,
 860                                         LNET_STATS_TYPE_DROP);
 861
 862                 CNETERR("Dropping message for %s: peer not alive\n",
 863                         libcfs_id2str(msg->msg_target));
 864                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 865                 if (do_send)
 866                         lnet_finalize(msg, -EHOSTUNREACH);
 867
 868                 lnet_net_lock(cpt);
 869                 return -EHOSTUNREACH;
 870         }
 871
 872         if (msg->msg_md != NULL &&
 873             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 874                 lnet_net_unlock(cpt);
 875
 876                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 877                         "called on the MD/ME.\n",
 878                         libcfs_id2str(msg->msg_target));
 879                 if (do_send) {
 880                         msg->msg_no_resend = true;
 881                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 882                                msg, libcfs_id2str(msg->msg_target));
 883                         lnet_finalize(msg, -ECANCELED);
 884                 }
 885
 886                 lnet_net_lock(cpt);
 887                 return -ECANCELED;
 888         }
 889
 890         if (!msg->msg_peertxcredit) {
 891                 spin_lock(&lp->lpni_lock);
 892                 LASSERT((lp->lpni_txcredits < 0) ==
 893                         !list_empty(&lp->lpni_txq));
 894
 895                 msg->msg_peertxcredit = 1;
 896                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 897                 lp->lpni_txcredits--;
 898
 899                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 900                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 901
 902                 if (lp->lpni_txcredits < 0) {
 903                         msg->msg_tx_delayed = 1;
 904                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 905                         spin_unlock(&lp->lpni_lock);
 906                         return LNET_CREDIT_WAIT;
 907                 }
 908                 spin_unlock(&lp->lpni_lock);
 909         }
 910
 911         if (!msg->msg_txcredit) {
 912                 LASSERT((tq->tq_credits < 0) ==
 913                         !list_empty(&tq->tq_delayed));
 914
 915                 msg->msg_txcredit = 1;
 916                 tq->tq_credits--;
 917                 atomic_dec(&ni->ni_tx_credits);
 918
 919                 if (tq->tq_credits < tq->tq_credits_min)
 920                         tq->tq_credits_min = tq->tq_credits;
 921
 922                 if (tq->tq_credits < 0) {
 923                         msg->msg_tx_delayed = 1;
 924                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 925                         return LNET_CREDIT_WAIT;
 926                 }
 927         }
 928
 929         /* unset the tx_delay flag as we're going to send it now */
 930         msg->msg_tx_delayed = 0;
 931
 932         if (do_send) {
 933                 lnet_net_unlock(cpt);
 934                 lnet_ni_send(ni, msg);
 935                 lnet_net_lock(cpt);
 936         }
 937         return LNET_CREDIT_OK;
 938 }
 939
 940
 941 static struct lnet_rtrbufpool *
 942 lnet_msg2bufpool(struct lnet_msg *msg)
 943 {
 944         struct lnet_rtrbufpool  *rbp;
 945         int                     cpt;
 946
 947         LASSERT(msg->msg_rx_committed);
 948
 949         cpt = msg->msg_rx_cpt;
 950         rbp = &the_lnet.ln_rtrpools[cpt][0];
 951
 952         LASSERT(msg->msg_len <= LNET_MTU);
 953         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 954                 rbp++;
 955                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 956         }
 957
 958         return rbp;
 959 }
 960
 961 static int
 962 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 963 {
 964         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 965          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 966          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 967          * received or OK to receive */
 968         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 969         struct lnet_peer *lp;
 970         struct lnet_rtrbufpool *rbp;
 971         struct lnet_rtrbuf *rb;
 972
 973         LASSERT(msg->msg_kiov == NULL);
 974         LASSERT(msg->msg_niov == 0);
 975         LASSERT(msg->msg_routing);
 976         LASSERT(msg->msg_receiving);
 977         LASSERT(!msg->msg_sending);
 978         LASSERT(lpni->lpni_peer_net);
 979         LASSERT(lpni->lpni_peer_net->lpn_peer);
 980
 981         lp = lpni->lpni_peer_net->lpn_peer;
 982
 983         /* non-lnet_parse callers only receive delayed messages */
 984         LASSERT(!do_recv || msg->msg_rx_delayed);
 985
 986         if (!msg->msg_peerrtrcredit) {
 987                 /* lpni_lock protects the credit manipulation */
 988                 spin_lock(&lpni->lpni_lock);
 989
 990                 msg->msg_peerrtrcredit = 1;
 991                 lpni->lpni_rtrcredits--;
 992                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
 993                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 994
 995                 if (lpni->lpni_rtrcredits < 0) {
 996                         spin_unlock(&lpni->lpni_lock);
 997                         /* must have checked eager_recv before here */
 998                         LASSERT(msg->msg_rx_ready_delay);
 999                         msg->msg_rx_delayed = 1;
1000                         /* lp_lock protects the lp_rtrq */
1001                         spin_lock(&lp->lp_lock);
1002                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1003                         spin_unlock(&lp->lp_lock);
1004                         return LNET_CREDIT_WAIT;
1005                 }
1006                 spin_unlock(&lpni->lpni_lock);
1007         }
1008
1009         rbp = lnet_msg2bufpool(msg);
1010
1011         if (!msg->msg_rtrcredit) {
1012                 msg->msg_rtrcredit = 1;
1013                 rbp->rbp_credits--;
1014                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1015                         rbp->rbp_mincredits = rbp->rbp_credits;
1016
1017                 if (rbp->rbp_credits < 0) {
1018                         /* must have checked eager_recv before here */
1019                         LASSERT(msg->msg_rx_ready_delay);
1020                         msg->msg_rx_delayed = 1;
1021                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1022                         return LNET_CREDIT_WAIT;
1023                 }
1024         }
1025
1026         LASSERT(!list_empty(&rbp->rbp_bufs));
1027         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1028         list_del(&rb->rb_list);
1029
1030         msg->msg_niov = rbp->rbp_npages;
1031         msg->msg_kiov = &rb->rb_kiov[0];
1032
1033         /* unset the msg-rx_delayed flag since we're receiving the message */
1034         msg->msg_rx_delayed = 0;
1035
1036         if (do_recv) {
1037                 int cpt = msg->msg_rx_cpt;
1038
1039                 lnet_net_unlock(cpt);
1040                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1041                              0, msg->msg_len, msg->msg_len);
1042                 lnet_net_lock(cpt);
1043         }
1044         return LNET_CREDIT_OK;
1045 }
1046
1047 void
1048 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1049 {
1050         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1051         struct lnet_ni          *txni = msg->msg_txni;
1052         struct lnet_msg         *msg2;
1053
1054         if (msg->msg_txcredit) {
1055                 struct lnet_ni       *ni = msg->msg_txni;
1056                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1057
1058                 /* give back NI txcredits */
1059                 msg->msg_txcredit = 0;
1060
1061                 LASSERT((tq->tq_credits < 0) ==
1062                         !list_empty(&tq->tq_delayed));
1063
1064                 tq->tq_credits++;
1065                 atomic_inc(&ni->ni_tx_credits);
1066                 if (tq->tq_credits <= 0) {
1067                         msg2 = list_entry(tq->tq_delayed.next,
1068                                           struct lnet_msg, msg_list);
1069                         list_del(&msg2->msg_list);
1070
1071                         LASSERT(msg2->msg_txni == ni);
1072                         LASSERT(msg2->msg_tx_delayed);
1073                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1074
1075                         (void) lnet_post_send_locked(msg2, 1);
1076                 }
1077         }
1078
1079         if (msg->msg_peertxcredit) {
1080                 /* give back peer txcredits */
1081                 msg->msg_peertxcredit = 0;
1082
1083                 spin_lock(&txpeer->lpni_lock);
1084                 LASSERT((txpeer->lpni_txcredits < 0) ==
1085                         !list_empty(&txpeer->lpni_txq));
1086
1087                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1088                 LASSERT(txpeer->lpni_txqnob >= 0);
1089
1090                 txpeer->lpni_txcredits++;
1091                 if (txpeer->lpni_txcredits <= 0) {
1092                         int msg2_cpt;
1093
1094                         msg2 = list_entry(txpeer->lpni_txq.next,
1095                                               struct lnet_msg, msg_list);
1096                         list_del(&msg2->msg_list);
1097                         spin_unlock(&txpeer->lpni_lock);
1098
1099                         LASSERT(msg2->msg_txpeer == txpeer);
1100                         LASSERT(msg2->msg_tx_delayed);
1101
1102                         msg2_cpt = msg2->msg_tx_cpt;
1103
1104                         /*
1105                          * The msg_cpt can be different from the msg2_cpt
1106                          * so we need to make sure we lock the correct cpt
1107                          * for msg2.
1108                          * Once we call lnet_post_send_locked() it is no
1109                          * longer safe to access msg2, since it could've
1110                          * been freed by lnet_finalize(), but we still
1111                          * need to relock the correct cpt, so we cache the
1112                          * msg2_cpt for the purpose of the check that
1113                          * follows the call to lnet_pose_send_locked().
1114                          */
1115                         if (msg2_cpt != msg->msg_tx_cpt) {
1116                                 lnet_net_unlock(msg->msg_tx_cpt);
1117                                 lnet_net_lock(msg2_cpt);
1118                         }
1119                         (void) lnet_post_send_locked(msg2, 1);
1120                         if (msg2_cpt != msg->msg_tx_cpt) {
1121                                 lnet_net_unlock(msg2_cpt);
1122                                 lnet_net_lock(msg->msg_tx_cpt);
1123                         }
1124                 } else {
1125                         spin_unlock(&txpeer->lpni_lock);
1126                 }
1127         }
1128
1129         if (txni != NULL) {
1130                 msg->msg_txni = NULL;
1131                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1132         }
1133
1134         if (txpeer != NULL) {
1135                 msg->msg_txpeer = NULL;
1136                 lnet_peer_ni_decref_locked(txpeer);
1137         }
1138 }
1139
1140 void
1141 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1142 {
1143         struct lnet_msg *msg;
1144
1145         if (list_empty(&rbp->rbp_msgs))
1146                 return;
1147         msg = list_entry(rbp->rbp_msgs.next,
1148                          struct lnet_msg, msg_list);
1149         list_del(&msg->msg_list);
1150
1151         (void)lnet_post_routed_recv_locked(msg, 1);
1152 }
1153
1154 void
1155 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1156 {
1157         struct lnet_msg *msg;
1158         struct lnet_msg *tmp;
1159
1160         lnet_net_unlock(cpt);
1161
1162         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1163                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1164                              0, 0, 0, msg->msg_hdr.payload_length);
1165                 list_del_init(&msg->msg_list);
1166                 msg->msg_no_resend = true;
1167                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1168                 lnet_finalize(msg, -ECANCELED);
1169         }
1170
1171         lnet_net_lock(cpt);
1172 }
1173
1174 void
1175 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1176 {
1177         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1178         struct lnet_peer *lp;
1179         struct lnet_ni *rxni = msg->msg_rxni;
1180         struct lnet_msg *msg2;
1181
1182         if (msg->msg_rtrcredit) {
1183                 /* give back global router credits */
1184                 struct lnet_rtrbuf *rb;
1185                 struct lnet_rtrbufpool *rbp;
1186
1187                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1188                  * there until it gets one allocated, or aborts the wait
1189                  * itself */
1190                 LASSERT(msg->msg_kiov != NULL);
1191
1192                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1193                 rbp = rb->rb_pool;
1194
1195                 msg->msg_kiov = NULL;
1196                 msg->msg_rtrcredit = 0;
1197
1198                 LASSERT(rbp == lnet_msg2bufpool(msg));
1199
1200                 LASSERT((rbp->rbp_credits > 0) ==
1201                         !list_empty(&rbp->rbp_bufs));
1202
1203                 /* If routing is now turned off, we just drop this buffer and
1204                  * don't bother trying to return credits.  */
1205                 if (!the_lnet.ln_routing) {
1206                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1207                         goto routing_off;
1208                 }
1209
1210                 /* It is possible that a user has lowered the desired number of
1211                  * buffers in this pool.  Make sure we never put back
1212                  * more buffers than the stated number. */
1213                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1214                         /* Discard this buffer so we don't have too
1215                          * many. */
1216                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1217                         rbp->rbp_nbuffers--;
1218                 } else {
1219                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1220                         rbp->rbp_credits++;
1221                         if (rbp->rbp_credits <= 0)
1222                                 lnet_schedule_blocked_locked(rbp);
1223                 }
1224         }
1225
1226 routing_off:
1227         if (msg->msg_peerrtrcredit) {
1228                 LASSERT(rxpeerni);
1229                 LASSERT(rxpeerni->lpni_peer_net);
1230                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1231
1232                 /* give back peer router credits */
1233                 msg->msg_peerrtrcredit = 0;
1234
1235                 spin_lock(&rxpeerni->lpni_lock);
1236                 rxpeerni->lpni_rtrcredits++;
1237                 spin_unlock(&rxpeerni->lpni_lock);
1238
1239                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1240                 spin_lock(&lp->lp_lock);
1241
1242                 /* drop all messages which are queued to be routed on that
1243                  * peer. */
1244                 if (!the_lnet.ln_routing) {
1245                         LIST_HEAD(drop);
1246                         list_splice_init(&lp->lp_rtrq, &drop);
1247                         spin_unlock(&lp->lp_lock);
1248                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1249                 } else if (!list_empty(&lp->lp_rtrq)) {
1250                         int msg2_cpt;
1251
1252                         msg2 = list_entry(lp->lp_rtrq.next,
1253                                           struct lnet_msg, msg_list);
1254                         list_del(&msg2->msg_list);
1255                         msg2_cpt = msg2->msg_rx_cpt;
1256                         spin_unlock(&lp->lp_lock);
1257                         /*
1258                          * messages on the lp_rtrq can be from any NID in
1259                          * the peer, which means they might have different
1260                          * cpts. We need to make sure we lock the right
1261                          * one.
1262                          */
1263                         if (msg2_cpt != msg->msg_rx_cpt) {
1264                                 lnet_net_unlock(msg->msg_rx_cpt);
1265                                 lnet_net_lock(msg2_cpt);
1266                         }
1267                         (void) lnet_post_routed_recv_locked(msg2, 1);
1268                         if (msg2_cpt != msg->msg_rx_cpt) {
1269                                 lnet_net_unlock(msg2_cpt);
1270                                 lnet_net_lock(msg->msg_rx_cpt);
1271                         }
1272                 } else {
1273                         spin_unlock(&lp->lp_lock);
1274                 }
1275         }
1276         if (rxni != NULL) {
1277                 msg->msg_rxni = NULL;
1278                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1279         }
1280         if (rxpeerni != NULL) {
1281                 msg->msg_rxpeer = NULL;
1282                 lnet_peer_ni_decref_locked(rxpeerni);
1283         }
1284 }
1285
1286 static int
1287 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1288 {
1289         if (p1->lpni_txqnob < p2->lpni_txqnob)
1290                 return 1;
1291
1292         if (p1->lpni_txqnob > p2->lpni_txqnob)
1293                 return -1;
1294
1295         if (p1->lpni_txcredits > p2->lpni_txcredits)
1296                 return 1;
1297
1298         if (p1->lpni_txcredits < p2->lpni_txcredits)
1299                 return -1;
1300
1301         return 0;
1302 }
1303
1304 static struct lnet_peer_ni *
1305 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1306                     struct lnet_peer *peer,
1307                     struct lnet_peer_ni *best_lpni,
1308                     struct lnet_peer_net *peer_net)
1309 {
1310         /*
1311          * Look at the peer NIs for the destination peer that connect
1312          * to the chosen net. If a peer_ni is preferred when using the
1313          * best_ni to communicate, we use that one. If there is no
1314          * preferred peer_ni, or there are multiple preferred peer_ni,
1315          * the available transmit credits are used. If the transmit
1316          * credits are equal, we round-robin over the peer_ni.
1317          */
1318         struct lnet_peer_ni *lpni = NULL;
1319         int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1320                 INT_MIN;
1321         int best_lpni_healthv = (best_lpni) ?
1322                 atomic_read(&best_lpni->lpni_healthv) : 0;
1323         bool preferred = false;
1324         bool ni_is_pref;
1325         int lpni_healthv;
1326
1327         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1328                 /*
1329                  * if the best_ni we've chosen aleady has this lpni
1330                  * preferred, then let's use it
1331                  */
1332                 if (best_ni) {
1333                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1334                                                                 best_ni->ni_nid);
1335                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1336                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1337                 } else {
1338                         ni_is_pref = false;
1339                 }
1340
1341                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1342
1343                 if (best_lpni)
1344                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1345                                 libcfs_nid2str(lpni->lpni_nid),
1346                                 lpni->lpni_txcredits, best_lpni_credits,
1347                                 lpni->lpni_seq, best_lpni->lpni_seq);
1348
1349                 /* pick the healthiest peer ni */
1350                 if (lpni_healthv < best_lpni_healthv) {
1351                         continue;
1352                 } else if (lpni_healthv > best_lpni_healthv) {
1353                         best_lpni_healthv = lpni_healthv;
1354                 /* if this is a preferred peer use it */
1355                 } else if (!preferred && ni_is_pref) {
1356                         preferred = true;
1357                 } else if (preferred && !ni_is_pref) {
1358                         /*
1359                          * this is not the preferred peer so let's ignore
1360                          * it.
1361                          */
1362                         continue;
1363                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1364                         /*
1365                          * We already have a peer that has more credits
1366                          * available than this one. No need to consider
1367                          * this peer further.
1368                          */
1369                         continue;
1370                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1371                         /*
1372                          * The best peer found so far and the current peer
1373                          * have the same number of available credits let's
1374                          * make sure to select between them using Round
1375                          * Robin
1376                          */
1377                         if (best_lpni) {
1378                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1379                                         continue;
1380                         }
1381                 }
1382
1383                 best_lpni = lpni;
1384                 best_lpni_credits = lpni->lpni_txcredits;
1385         }
1386
1387         /* if we still can't find a peer ni then we can't reach it */
1388         if (!best_lpni) {
1389                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1390                         LNET_NIDNET(dst_nid);
1391                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1392                                 libcfs_net2str(net_id));
1393                 return NULL;
1394         }
1395
1396         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1397                libcfs_nid2str(best_lpni->lpni_nid));
1398
1399         return best_lpni;
1400 }
1401
1402 /*
1403  * Prerequisite: the best_ni should already be set in the sd
1404  * Find the best lpni.
1405  * If the net id is provided then restrict lpni selection on
1406  * that particular net.
1407  * Otherwise find any reachable lpni. When dealing with an MR
1408  * gateway and it has multiple lpnis which we can use
1409  * we want to select the best one from the list of reachable
1410  * ones.
1411  */
1412 static inline struct lnet_peer_ni *
1413 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1414                     struct lnet_peer *peer, __u32 net_id)
1415 {
1416         struct lnet_peer_net *peer_net;
1417         __u32 any_net = LNET_NIDNET(LNET_NID_ANY);
1418
1419         /* find the best_lpni on any local network */
1420         if (net_id == any_net) {
1421                 struct lnet_peer_ni *best_lpni = NULL;
1422                 struct lnet_peer_net *lpn;
1423                 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1424                         /* no net specified find any reachable peer ni */
1425                         if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1426                                 continue;
1427                         best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1428                                                         best_lpni, lpn);
1429                 }
1430
1431                 return best_lpni;
1432         }
1433         /* restrict on the specified net */
1434         peer_net = lnet_peer_get_net_locked(peer, net_id);
1435         if (peer_net)
1436                 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1437
1438         return NULL;
1439 }
1440
1441 /* Compare route priorities and hop counts */
1442 static int
1443 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1444 {
1445         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1446         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1447
1448         if (r1->lr_priority < r2->lr_priority)
1449                 return 1;
1450
1451         if (r1->lr_priority > r2->lr_priority)
1452                 return -1;
1453
1454         if (r1_hops < r2_hops)
1455                 return 1;
1456
1457         if (r1_hops > r2_hops)
1458                 return -1;
1459
1460         return 0;
1461 }
1462
1463 static struct lnet_route *
1464 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1465                        struct lnet_route **prev_route,
1466                        struct lnet_peer_ni **gwni)
1467 {
1468         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1469         struct lnet_route *best_route;
1470         struct lnet_route *last_route;
1471         struct lnet_route *route;
1472         int rc;
1473
1474         CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1475                libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1476
1477         best_route = last_route = NULL;
1478         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1479                 if (!lnet_is_route_alive(route))
1480                         continue;
1481
1482                 /*
1483                  * Restrict the selection of the router NI on the src_net
1484                  * provided. If the src_net is LNET_NID_ANY, then select
1485                  * the best interface available.
1486                  */
1487                 if (!best_route) {
1488                         lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1489                                                    route->lr_gateway,
1490                                                    src_net);
1491                         if (lpni) {
1492                                 best_route = last_route = route;
1493                                 best_gw_ni = lpni;
1494                         } else {
1495                                 CDEBUG(D_NET, "Gateway %s does not have a peer NI on net %s\n",
1496                                        libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1497                                        libcfs_net2str(src_net));
1498                         }
1499
1500                         continue;
1501                 }
1502
1503                 /* no protection on below fields, but it's harmless */
1504                 if (last_route->lr_seq - route->lr_seq < 0)
1505                         last_route = route;
1506
1507                 rc = lnet_compare_routes(route, best_route);
1508                 if (rc == -1)
1509                         continue;
1510
1511                 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1512                                            route->lr_gateway,
1513                                            src_net);
1514                 /* restrict the lpni on the src_net if specified */
1515                 if (!lpni) {
1516                         CDEBUG(D_NET, "Gateway %s does not have a peer NI on net %s\n",
1517                                libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1518                                libcfs_net2str(src_net));
1519                         continue;
1520                 }
1521
1522                 if (rc == 1) {
1523                         best_route = route;
1524                         best_gw_ni = lpni;
1525                         continue;
1526                 }
1527
1528                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1529                 if (rc == -1)
1530                         continue;
1531
1532                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1533                         best_route = route;
1534                         best_gw_ni = lpni;
1535                         continue;
1536                 }
1537         }
1538
1539         *prev_route = last_route;
1540         *gwni = best_gw_ni;
1541
1542         return best_route;
1543 }
1544
1545 static struct lnet_ni *
1546 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1547                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1548                  int md_cpt)
1549 {
1550         struct lnet_ni *ni = NULL;
1551         unsigned int shortest_distance;
1552         int best_credits;
1553         int best_healthv;
1554
1555         /*
1556          * If there is no peer_ni that we can send to on this network,
1557          * then there is no point in looking for a new best_ni here.
1558         */
1559         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1560                 return best_ni;
1561
1562         if (best_ni == NULL) {
1563                 shortest_distance = UINT_MAX;
1564                 best_credits = INT_MIN;
1565                 best_healthv = 0;
1566         } else {
1567                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1568                                                      best_ni->ni_dev_cpt);
1569                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1570                 best_healthv = atomic_read(&best_ni->ni_healthv);
1571         }
1572
1573         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1574                 unsigned int distance;
1575                 int ni_credits;
1576                 int ni_healthv;
1577                 int ni_fatal;
1578
1579                 ni_credits = atomic_read(&ni->ni_tx_credits);
1580                 ni_healthv = atomic_read(&ni->ni_healthv);
1581                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1582
1583                 /*
1584                  * calculate the distance from the CPT on which
1585                  * the message memory is allocated to the CPT of
1586                  * the NI's physical device
1587                  */
1588                 distance = cfs_cpt_distance(lnet_cpt_table(),
1589                                             md_cpt,
1590                                             ni->ni_dev_cpt);
1591
1592                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1593                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1594                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1595                         : "not seleced", best_credits, shortest_distance,
1596                         (best_ni) ? best_ni->ni_seq : 0);
1597
1598                 /*
1599                  * All distances smaller than the NUMA range
1600                  * are treated equally.
1601                  */
1602                 if (distance < lnet_numa_range)
1603                         distance = lnet_numa_range;
1604
1605                 /*
1606                  * Select on health, shorter distance, available
1607                  * credits, then round-robin.
1608                  */
1609                 if (ni_fatal) {
1610                         continue;
1611                 } else if (ni_healthv < best_healthv) {
1612                         continue;
1613                 } else if (ni_healthv > best_healthv) {
1614                         best_healthv = ni_healthv;
1615                         /*
1616                          * If we're going to prefer this ni because it's
1617                          * the healthiest, then we should set the
1618                          * shortest_distance in the algorithm in case
1619                          * there are multiple NIs with the same health but
1620                          * different distances.
1621                          */
1622                         if (distance < shortest_distance)
1623                                 shortest_distance = distance;
1624                 } else if (distance > shortest_distance) {
1625                         continue;
1626                 } else if (distance < shortest_distance) {
1627                         shortest_distance = distance;
1628                 } else if (ni_credits < best_credits) {
1629                         continue;
1630                 } else if (ni_credits == best_credits) {
1631                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1632                                 continue;
1633                 }
1634                 best_ni = ni;
1635                 best_credits = ni_credits;
1636         }
1637
1638         CDEBUG(D_NET, "selected best_ni %s\n",
1639                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1640
1641         return best_ni;
1642 }
1643
1644 /*
1645  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1646  * because such traffic is required to perform discovery. We therefore
1647  * exclude all GET and PUT on that portal. We also exclude all ACK and
1648  * REPLY traffic, but that is because the portal is not tracked in the
1649  * message structure for these message types. We could restrict this
1650  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1651  */
1652 static bool
1653 lnet_msg_discovery(struct lnet_msg *msg)
1654 {
1655         if (msg->msg_type == LNET_MSG_PUT) {
1656                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1657                         return true;
1658         } else if (msg->msg_type == LNET_MSG_GET) {
1659                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1660                         return true;
1661         }
1662         return false;
1663 }
1664
1665 #define SRC_SPEC        0x0001
1666 #define SRC_ANY         0x0002
1667 #define LOCAL_DST       0x0004
1668 #define REMOTE_DST      0x0008
1669 #define MR_DST          0x0010
1670 #define NMR_DST         0x0020
1671 #define SND_RESP        0x0040
1672
1673 /* The following to defines are used for return codes */
1674 #define REPEAT_SEND     0x1000
1675 #define PASS_THROUGH    0x2000
1676
1677 /* The different cases lnet_select pathway needs to handle */
1678 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1679 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1680 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1681 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1682 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1683 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1684 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1685 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1686
1687 static int
1688 lnet_handle_lo_send(struct lnet_send_data *sd)
1689 {
1690         struct lnet_msg *msg = sd->sd_msg;
1691         int cpt = sd->sd_cpt;
1692
1693         /* No send credit hassles with LOLND */
1694         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1695         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1696         if (!msg->msg_routing)
1697                 msg->msg_hdr.src_nid =
1698                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1699         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1700         lnet_msg_commit(msg, cpt);
1701         msg->msg_txni = the_lnet.ln_loni;
1702
1703         return LNET_CREDIT_OK;
1704 }
1705
1706 static int
1707 lnet_handle_send(struct lnet_send_data *sd)
1708 {
1709         struct lnet_ni *best_ni = sd->sd_best_ni;
1710         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1711         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1712         struct lnet_msg *msg = sd->sd_msg;
1713         int cpt2;
1714         __u32 send_case = sd->sd_send_case;
1715         int rc;
1716         __u32 routing = send_case & REMOTE_DST;
1717          struct lnet_rsp_tracker *rspt;
1718
1719         /*
1720          * Increment sequence number of the selected peer so that we
1721          * pick the next one in Round Robin.
1722          */
1723         best_lpni->lpni_seq++;
1724
1725         /*
1726          * grab a reference on the peer_ni so it sticks around even if
1727          * we need to drop and relock the lnet_net_lock below.
1728          */
1729         lnet_peer_ni_addref_locked(best_lpni);
1730
1731         /*
1732          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1733          * message. This ensures that we get a CPT that is correct for
1734          * the NI when the NI has been restricted to a subset of all CPTs.
1735          * If the selected CPT differs from the one currently locked, we
1736          * must unlock and relock the lnet_net_lock(), and then check whether
1737          * the configuration has changed. We don't have a hold on the best_ni
1738          * yet, and it may have vanished.
1739          */
1740         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1741         if (sd->sd_cpt != cpt2) {
1742                 __u32 seq = lnet_get_dlc_seq_locked();
1743                 lnet_net_unlock(sd->sd_cpt);
1744                 sd->sd_cpt = cpt2;
1745                 lnet_net_lock(sd->sd_cpt);
1746                 if (seq != lnet_get_dlc_seq_locked()) {
1747                         lnet_peer_ni_decref_locked(best_lpni);
1748                         return REPEAT_SEND;
1749                 }
1750         }
1751
1752         /*
1753          * store the best_lpni in the message right away to avoid having
1754          * to do the same operation under different conditions
1755          */
1756         msg->msg_txpeer = best_lpni;
1757         msg->msg_txni = best_ni;
1758
1759         /*
1760          * grab a reference for the best_ni since now it's in use in this
1761          * send. The reference will be dropped in lnet_finalize()
1762          */
1763         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1764
1765         /*
1766          * Always set the target.nid to the best peer picked. Either the
1767          * NID will be one of the peer NIDs selected, or the same NID as
1768          * what was originally set in the target or it will be the NID of
1769          * a router if this message should be routed
1770          */
1771         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1772
1773         /*
1774          * lnet_msg_commit assigns the correct cpt to the message, which
1775          * is used to decrement the correct refcount on the ni when it's
1776          * time to return the credits
1777          */
1778         lnet_msg_commit(msg, sd->sd_cpt);
1779
1780         /*
1781          * If we are routing the message then we keep the src_nid that was
1782          * set by the originator. If we are not routing then we are the
1783          * originator and set it here.
1784          */
1785         if (!msg->msg_routing)
1786                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1787
1788         if (routing) {
1789                 msg->msg_target_is_router = 1;
1790                 msg->msg_target.pid = LNET_PID_LUSTRE;
1791                 /*
1792                  * since we're routing we want to ensure that the
1793                  * msg_hdr.dest_nid is set to the final destination. When
1794                  * the router receives this message it knows how to route
1795                  * it.
1796                  *
1797                  * final_dst_lpni is set at the beginning of the
1798                  * lnet_select_pathway() function and is never changed.
1799                  * It's safe to use it here.
1800                  */
1801                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1802         } else {
1803                 /*
1804                  * if we're not routing set the dest_nid to the best peer
1805                  * ni NID that we picked earlier in the algorithm.
1806                  */
1807                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1808         }
1809
1810         /*
1811          * if we have response tracker block update it with the next hop
1812          * nid
1813          */
1814         if (msg->msg_md) {
1815                 rspt = msg->msg_md->md_rspt_ptr;
1816                 if (rspt) {
1817                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1818                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1819                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1820                 }
1821         }
1822
1823         rc = lnet_post_send_locked(msg, 0);
1824
1825         if (!rc)
1826                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1827                        libcfs_nid2str(msg->msg_hdr.src_nid),
1828                        libcfs_nid2str(msg->msg_txni->ni_nid),
1829                        libcfs_nid2str(sd->sd_src_nid),
1830                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1831                        libcfs_nid2str(sd->sd_dst_nid),
1832                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1833                        libcfs_nid2str(sd->sd_rtr_nid),
1834                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1835
1836         return rc;
1837 }
1838
1839 static inline void
1840 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1841                          struct lnet_msg *msg)
1842 {
1843         if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1844             !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1845                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1846                        libcfs_nid2str(lni->ni_nid),
1847                        libcfs_nid2str(lpni->lpni_nid));
1848                 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1849         }
1850 }
1851
1852 /*
1853  * Source Specified
1854  * Local Destination
1855  * non-mr peer
1856  *
1857  * use the source and destination NIDs as the pathway
1858  */
1859 static int
1860 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1861 {
1862         /* the destination lpni is set before we get here. */
1863
1864         /* find local NI */
1865         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1866         if (!sd->sd_best_ni) {
1867                 CERROR("Can't send to %s: src %s is not a "
1868                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1869                                 libcfs_nid2str(sd->sd_src_nid));
1870                 return -EINVAL;
1871         }
1872
1873         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1874
1875         return lnet_handle_send(sd);
1876 }
1877
1878 /*
1879  * Source Specified
1880  * Local Destination
1881  * MR Peer
1882  *
1883  * Don't run the selection algorithm on the peer NIs. By specifying the
1884  * local NID, we're also saying that we should always use the destination NID
1885  * provided. This handles the case where we should be using the same
1886  * destination NID for the all the messages which belong to the same RPC
1887  * request.
1888  */
1889 static int
1890 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1891 {
1892         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1893         if (!sd->sd_best_ni) {
1894                 CERROR("Can't send to %s: src %s is not a "
1895                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1896                                 libcfs_nid2str(sd->sd_src_nid));
1897                 return -EINVAL;
1898         }
1899
1900         if (sd->sd_best_lpni &&
1901             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1902                 return lnet_handle_lo_send(sd);
1903         else if (sd->sd_best_lpni)
1904                 return lnet_handle_send(sd);
1905
1906         CERROR("can't send to %s. no NI on %s\n",
1907                libcfs_nid2str(sd->sd_dst_nid),
1908                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1909
1910         return -EHOSTUNREACH;
1911 }
1912
1913 struct lnet_ni *
1914 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1915                               struct lnet_peer *peer,
1916                               struct lnet_peer_net *peer_net,
1917                               int cpt,
1918                               bool incr_seq)
1919 {
1920         struct lnet_net *local_net;
1921         struct lnet_ni *best_ni;
1922
1923         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1924         if (!local_net)
1925                 return NULL;
1926
1927         /*
1928          * Iterate through the NIs in this local Net and select
1929          * the NI to send from. The selection is determined by
1930          * these 3 criterion in the following priority:
1931          *      1. NUMA
1932          *      2. NI available credits
1933          *      3. Round Robin
1934          */
1935         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1936                                    peer, peer_net, cpt);
1937
1938         if (incr_seq && best_ni)
1939                 best_ni->ni_seq++;
1940
1941         return best_ni;
1942 }
1943
1944 static int
1945 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
1946                              int cpt)
1947 {
1948         struct lnet_peer *peer;
1949         int rc;
1950
1951         lnet_peer_ni_addref_locked(lpni);
1952
1953         peer = lpni->lpni_peer_net->lpn_peer;
1954
1955         if (lnet_peer_gw_discovery(peer)) {
1956                 lnet_peer_ni_decref_locked(lpni);
1957                 return 0;
1958         }
1959
1960         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1961                 lnet_peer_ni_decref_locked(lpni);
1962                 return 0;
1963         }
1964
1965         rc = lnet_discover_peer_locked(lpni, cpt, false);
1966         if (rc) {
1967                 lnet_peer_ni_decref_locked(lpni);
1968                 return rc;
1969         }
1970         /* The peer may have changed. */
1971         peer = lpni->lpni_peer_net->lpn_peer;
1972         spin_lock(&peer->lp_lock);
1973         if (lnet_peer_is_uptodate_locked(peer)) {
1974                 spin_unlock(&peer->lp_lock);
1975                 lnet_peer_ni_decref_locked(lpni);
1976                 return 0;
1977         }
1978         /* queue message and return */
1979         msg->msg_sending = 0;
1980         msg->msg_txpeer = NULL;
1981         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1982         spin_unlock(&peer->lp_lock);
1983
1984         lnet_peer_ni_decref_locked(lpni);
1985
1986         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1987                msg, libcfs_nid2str(peer->lp_primary_nid));
1988
1989         return LNET_DC_WAIT;
1990 }
1991
1992 static int
1993 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1994                              lnet_nid_t dst_nid,
1995                              struct lnet_peer_ni **gw_lpni,
1996                              struct lnet_peer **gw_peer)
1997 {
1998         int rc;
1999         __u32 local_lnet;
2000         struct lnet_peer *gw;
2001         struct lnet_peer *lp;
2002         struct lnet_peer_net *lpn;
2003         struct lnet_peer_net *best_lpn = NULL;
2004         struct lnet_remotenet *rnet, *best_rnet = NULL;
2005         struct lnet_route *best_route = NULL;
2006         struct lnet_route *last_route = NULL;
2007         struct lnet_peer_ni *lpni = NULL;
2008         struct lnet_peer_ni *gwni = NULL;
2009         bool route_found = false;
2010         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2011                 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
2012                 LNET_NID_ANY;
2013
2014         CDEBUG(D_NET, "using src nid %s for route restriction\n",
2015                libcfs_nid2str(src_nid));
2016
2017         /* If a router nid was specified then we are replying to a GET or
2018          * sending an ACK. In this case we use the gateway associated with the
2019          * specified router nid.
2020          */
2021         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2022                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2023                 if (gwni) {
2024                         gw = gwni->lpni_peer_net->lpn_peer;
2025                         lnet_peer_ni_decref_locked(gwni);
2026                         if (gw->lp_rtr_refcount) {
2027                                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2028                                 route_found = true;
2029                         }
2030                 } else {
2031                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2032                                libcfs_nid2str(sd->sd_rtr_nid));
2033                 }
2034         }
2035
2036         if (!route_found) {
2037                 /* we've already looked up the initial lpni using dst_nid */
2038                 lpni = sd->sd_best_lpni;
2039                 /* the peer tree must be in existence */
2040                 LASSERT(lpni && lpni->lpni_peer_net &&
2041                         lpni->lpni_peer_net->lpn_peer);
2042                 lp = lpni->lpni_peer_net->lpn_peer;
2043
2044                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2045                         /* is this remote network reachable?  */
2046                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2047                         if (!rnet)
2048                                 continue;
2049
2050                         if (!best_lpn) {
2051                                 best_lpn = lpn;
2052                                 best_rnet = rnet;
2053                         }
2054
2055                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2056                                 continue;
2057
2058                         best_lpn = lpn;
2059                         best_rnet = rnet;
2060                 }
2061
2062                 if (!best_lpn) {
2063                         CERROR("peer %s has no available nets\n",
2064                                libcfs_nid2str(sd->sd_dst_nid));
2065                         return -EHOSTUNREACH;
2066                 }
2067
2068                 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2069                                                        sd->sd_dst_nid,
2070                                                        lp,
2071                                                        best_lpn->lpn_net_id);
2072                 if (!sd->sd_best_lpni) {
2073                         CERROR("peer %s is unreachable\n",
2074                                libcfs_nid2str(sd->sd_dst_nid));
2075                         return -EHOSTUNREACH;
2076                 }
2077
2078                 /*
2079                  * We're attempting to round robin over the remote peer
2080                  * NI's so update the final destination we selected
2081                  */
2082                 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2083
2084                 /*
2085                  * find the best route. Restrict the selection on the net of the
2086                  * local NI if we've already picked the local NI to send from.
2087                  * Otherwise, let's pick any route we can find and then find
2088                  * a local NI we can reach the route's gateway on. Any route we select
2089                  * will be reachable by virtue of the restriction we have when
2090                  * adding a route.
2091                  */
2092                 best_route = lnet_find_route_locked(best_rnet,
2093                                                     LNET_NIDNET(src_nid),
2094                                                     &last_route, &gwni);
2095
2096                 if (!best_route) {
2097                         CERROR("no route to %s from %s\n",
2098                                libcfs_nid2str(dst_nid),
2099                                libcfs_nid2str(src_nid));
2100                         return -EHOSTUNREACH;
2101                 }
2102
2103                 if (!gwni) {
2104                         CERROR("Internal Error. Route expected to %s from %s\n",
2105                                libcfs_nid2str(dst_nid),
2106                                libcfs_nid2str(src_nid));
2107                         return -EFAULT;
2108                 }
2109
2110                 gw = best_route->lr_gateway;
2111                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2112                 local_lnet = best_route->lr_lnet;
2113
2114                 /*
2115                  * Increment the sequence number of the remote lpni so we
2116                  * can round robin over the different interfaces of the
2117                  * remote lpni
2118                  */
2119                 sd->sd_best_lpni->lpni_seq++;
2120         }
2121
2122         /*
2123          * Discover this gateway if it hasn't already been discovered.
2124          * This means we might delay the message until discovery has
2125          * completed
2126          */
2127         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2128         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2129         if (rc)
2130                 return rc;
2131
2132         if (!sd->sd_best_ni)
2133                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2134                                         lnet_peer_get_net_locked(gw,
2135                                                                  local_lnet),
2136                                         sd->sd_md_cpt,
2137                                         true);
2138
2139         if (!sd->sd_best_ni) {
2140                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2141                        libcfs_net2str(local_lnet),
2142                        libcfs_nid2str(sd->sd_src_nid));
2143                 return -EFAULT;
2144         }
2145
2146         *gw_lpni = gwni;
2147         *gw_peer = gw;
2148
2149         /*
2150          * increment the sequence numbers since now we're sure we're
2151          * going to use this path
2152          */
2153         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2154                 LASSERT(best_route && last_route);
2155                 best_route->lr_seq = last_route->lr_seq + 1;
2156                 best_lpn->lpn_seq++;
2157         }
2158
2159         return 0;
2160 }
2161
2162 /*
2163  * Handle two cases:
2164  *
2165  * Case 1:
2166  *  Source specified
2167  *  Remote destination
2168  *  Non-MR destination
2169  *
2170  * Case 2:
2171  *  Source specified
2172  *  Remote destination
2173  *  MR destination
2174  *
2175  * The handling of these two cases is similar. Even though the destination
2176  * can be MR or non-MR, we'll deal directly with the router.
2177  */
2178 static int
2179 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2180 {
2181         int rc;
2182         struct lnet_peer_ni *gw_lpni = NULL;
2183         struct lnet_peer *gw_peer = NULL;
2184
2185         /* find local NI */
2186         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2187         if (!sd->sd_best_ni) {
2188                 CERROR("Can't send to %s: src %s is not a "
2189                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2190                                 libcfs_nid2str(sd->sd_src_nid));
2191                 return -EINVAL;
2192         }
2193
2194         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2195                                      &gw_peer);
2196         if (rc)
2197                 return rc;
2198
2199         if (sd->sd_send_case & NMR_DST)
2200                 /*
2201                  * since the final destination is non-MR let's set its preferred
2202                  * NID before we send
2203                  */
2204                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2205                                          sd->sd_msg);
2206
2207         /*
2208          * We're going to send to the gw found so let's set its
2209          * info
2210          */
2211         sd->sd_peer = gw_peer;
2212         sd->sd_best_lpni = gw_lpni;
2213
2214         return lnet_handle_send(sd);
2215 }
2216
2217 struct lnet_ni *
2218 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2219                                bool discovery)
2220 {
2221         struct lnet_peer_net *peer_net = NULL;
2222         struct lnet_ni *best_ni = NULL;
2223         int lpn_healthv = 0;
2224
2225         /*
2226          * The peer can have multiple interfaces, some of them can be on
2227          * the local network and others on a routed network. We should
2228          * prefer the local network. However if the local network is not
2229          * available then we need to try the routed network
2230          */
2231
2232         /* go through all the peer nets and find the best_ni */
2233         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2234                 /*
2235                  * The peer's list of nets can contain non-local nets. We
2236                  * want to only examine the local ones.
2237                  */
2238                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2239                         continue;
2240
2241                 /* always select the lpn with the best health */
2242                 if (lpn_healthv <= peer_net->lpn_healthv)
2243                         lpn_healthv = peer_net->lpn_healthv;
2244                 else
2245                         continue;
2246
2247                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2248                                                         md_cpt, false);
2249
2250                 /*
2251                  * if this is a discovery message and lp_disc_net_id is
2252                  * specified then use that net to send the discovery on.
2253                  */
2254                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2255                     discovery)
2256                         break;
2257         }
2258
2259         if (best_ni)
2260                 /* increment sequence number so we can round robin */
2261                 best_ni->ni_seq++;
2262
2263         return best_ni;
2264 }
2265
2266 static struct lnet_ni *
2267 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2268 {
2269         struct lnet_ni *best_ni = NULL;
2270         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2271         struct lnet_peer_ni *lpni_entry;
2272
2273         /*
2274          * We must use a consistent source address when sending to a
2275          * non-MR peer. However, a non-MR peer can have multiple NIDs
2276          * on multiple networks, and we may even need to talk to this
2277          * peer on multiple networks -- certain types of
2278          * load-balancing configuration do this.
2279          *
2280          * So we need to pick the NI the peer prefers for this
2281          * particular network.
2282          */
2283         LASSERT(peer_net);
2284         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2285                             lpni_peer_nis) {
2286                 if (lpni_entry->lpni_pref_nnids == 0)
2287                         continue;
2288                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2289                 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2290                 break;
2291         }
2292
2293         return best_ni;
2294 }
2295
2296 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2297 static int
2298 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2299 {
2300         struct lnet_ni *best_ni = NULL;
2301         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2302
2303         /*
2304          * We must use a consistent source address when sending to a
2305          * non-MR peer. However, a non-MR peer can have multiple NIDs
2306          * on multiple networks, and we may even need to talk to this
2307          * peer on multiple networks -- certain types of
2308          * load-balancing configuration do this.
2309          *
2310          * So we need to pick the NI the peer prefers for this
2311          * particular network.
2312          */
2313
2314         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2315                                                        sd->sd_cpt);
2316
2317         /* if best_ni is still not set just pick one */
2318         if (!best_ni) {
2319                 best_ni =
2320                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2321                                                 sd->sd_best_lpni->lpni_peer_net,
2322                                                 sd->sd_md_cpt, true);
2323                 /* If there is no best_ni we don't have a route */
2324                 if (!best_ni) {
2325                         CERROR("no path to %s from net %s\n",
2326                                 libcfs_nid2str(best_lpni->lpni_nid),
2327                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2328                         return -EHOSTUNREACH;
2329                 }
2330         }
2331
2332         sd->sd_best_ni = best_ni;
2333
2334         /* Set preferred NI if necessary. */
2335         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2336
2337         return 0;
2338 }
2339
2340
2341 /*
2342  * Source not specified
2343  * Local destination
2344  * Non-MR Peer
2345  *
2346  * always use the same source NID for NMR peers
2347  * If we've talked to that peer before then we already have a preferred
2348  * source NI associated with it. Otherwise, we select a preferred local NI
2349  * and store it in the peer
2350  */
2351 static int
2352 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2353 {
2354         int rc = 0;
2355
2356         /* sd->sd_best_lpni is already set to the final destination */
2357
2358         /*
2359          * At this point we should've created the peer ni and peer. If we
2360          * can't find it, then something went wrong. Instead of assert
2361          * output a relevant message and fail the send
2362          */
2363         if (!sd->sd_best_lpni) {
2364                 CERROR("Internal fault. Unable to send msg %s to %s. "
2365                        "NID not known\n",
2366                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2367                        libcfs_nid2str(sd->sd_dst_nid));
2368                 return -EFAULT;
2369         }
2370
2371         if (sd->sd_msg->msg_routing) {
2372                 /* If I'm forwarding this message then I can choose any NI
2373                  * on the destination peer net
2374                  */
2375                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2376                                                                sd->sd_peer,
2377                                                                sd->sd_best_lpni->lpni_peer_net,
2378                                                                sd->sd_md_cpt,
2379                                                                true);
2380                 if (!sd->sd_best_ni) {
2381                         CERROR("Unable to forward message to %s. No local NI available\n",
2382                                libcfs_nid2str(sd->sd_dst_nid));
2383                         rc = -EHOSTUNREACH;
2384                 }
2385         } else
2386                 rc = lnet_select_preferred_best_ni(sd);
2387
2388         if (!rc)
2389                 rc = lnet_handle_send(sd);
2390
2391         return rc;
2392 }
2393
2394 static int
2395 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2396 {
2397         /*
2398          * NOTE we've already handled the remote peer case. So we only
2399          * need to worry about the local case here.
2400          *
2401          * if we're sending a response, ACK or reply, we need to send it
2402          * to the destination NID given to us. At this point we already
2403          * have the peer_ni we're suppose to send to, so just find the
2404          * best_ni on the peer net and use that. Since we're sending to an
2405          * MR peer then we can just run the selection algorithm on our
2406          * local NIs and pick the best one.
2407          */
2408         if (sd->sd_send_case & SND_RESP) {
2409                 sd->sd_best_ni =
2410                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2411                                                 sd->sd_best_lpni->lpni_peer_net,
2412                                                 sd->sd_md_cpt, true);
2413
2414                 if (!sd->sd_best_ni) {
2415                         /*
2416                          * We're not going to deal with not able to send
2417                          * a response to the provided final destination
2418                          */
2419                         CERROR("Can't send response to %s. "
2420                                "No local NI available\n",
2421                                 libcfs_nid2str(sd->sd_dst_nid));
2422                         return -EHOSTUNREACH;
2423                 }
2424
2425                 return lnet_handle_send(sd);
2426         }
2427
2428         /*
2429          * If we get here that means we're sending a fresh request, PUT or
2430          * GET, so we need to run our standard selection algorithm.
2431          * First find the best local interface that's on any of the peer's
2432          * networks.
2433          */
2434         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2435                                         sd->sd_md_cpt,
2436                                         lnet_msg_discovery(sd->sd_msg));
2437         if (sd->sd_best_ni) {
2438                 sd->sd_best_lpni =
2439                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2440                                       sd->sd_peer,
2441                                       sd->sd_best_ni->ni_net->net_id);
2442
2443                 /*
2444                  * if we're successful in selecting a peer_ni on the local
2445                  * network, then send to it. Otherwise fall through and
2446                  * try and see if we can reach it over another routed
2447                  * network
2448                  */
2449                 if (sd->sd_best_lpni &&
2450                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2451                         /*
2452                          * in case we initially started with a routed
2453                          * destination, let's reset to local
2454                          */
2455                         sd->sd_send_case &= ~REMOTE_DST;
2456                         sd->sd_send_case |= LOCAL_DST;
2457                         return lnet_handle_lo_send(sd);
2458                 } else if (sd->sd_best_lpni) {
2459                         /*
2460                          * in case we initially started with a routed
2461                          * destination, let's reset to local
2462                          */
2463                         sd->sd_send_case &= ~REMOTE_DST;
2464                         sd->sd_send_case |= LOCAL_DST;
2465                         return lnet_handle_send(sd);
2466                 }
2467
2468                 CERROR("Internal Error. Expected to have a best_lpni: "
2469                        "%s -> %s\n",
2470                        libcfs_nid2str(sd->sd_src_nid),
2471                        libcfs_nid2str(sd->sd_dst_nid));
2472
2473                 return -EFAULT;
2474         }
2475
2476         /*
2477          * Peer doesn't have a local network. Let's see if there is
2478          * a remote network we can reach it on.
2479          */
2480         return PASS_THROUGH;
2481 }
2482
2483 /*
2484  * Case 1:
2485  *      Source NID not specified
2486  *      Local destination
2487  *      MR peer
2488  *
2489  * Case 2:
2490  *      Source NID not speified
2491  *      Remote destination
2492  *      MR peer
2493  *
2494  * In both of these cases if we're sending a response, ACK or REPLY, then
2495  * we need to send to the destination NID provided.
2496  *
2497  * In the remote case let's deal with MR routers.
2498  *
2499  */
2500
2501 static int
2502 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2503 {
2504         int rc = 0;
2505         struct lnet_peer *gw_peer = NULL;
2506         struct lnet_peer_ni *gw_lpni = NULL;
2507
2508         /*
2509          * handle sending a response to a remote peer here so we don't
2510          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2511          */
2512         if (sd->sd_send_case & REMOTE_DST &&
2513             sd->sd_send_case & SND_RESP) {
2514                 struct lnet_peer_ni *gw;
2515                 struct lnet_peer *gw_peer;
2516
2517                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2518                                                   &gw_peer);
2519                 if (rc < 0) {
2520                         CERROR("Can't send response to %s. "
2521                                "No route available\n",
2522                                 libcfs_nid2str(sd->sd_dst_nid));
2523                         return -EHOSTUNREACH;
2524                 } else if (rc > 0) {
2525                         return rc;
2526                 }
2527
2528                 sd->sd_best_lpni = gw;
2529                 sd->sd_peer = gw_peer;
2530
2531                 return lnet_handle_send(sd);
2532         }
2533
2534         /*
2535          * Even though the NID for the peer might not be on a local network,
2536          * since the peer is MR there could be other interfaces on the
2537          * local network. In that case we'd still like to prefer the local
2538          * network over the routed network. If we're unable to do that
2539          * then we select the best router among the different routed networks,
2540          * and if the router is MR then we can deal with it as such.
2541          */
2542         rc = lnet_handle_any_mr_dsta(sd);
2543         if (rc != PASS_THROUGH)
2544                 return rc;
2545
2546         /*
2547          * Now that we must route to the destination, we must consider the
2548          * MR case, where the destination has multiple interfaces, some of
2549          * which we can route to and others we do not. For this reason we
2550          * need to select the destination which we can route to and if
2551          * there are multiple, we need to round robin.
2552          */
2553         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2554                                           &gw_peer);
2555         if (rc)
2556                 return rc;
2557
2558         sd->sd_send_case &= ~LOCAL_DST;
2559         sd->sd_send_case |= REMOTE_DST;
2560
2561         sd->sd_peer = gw_peer;
2562         sd->sd_best_lpni = gw_lpni;
2563
2564         return lnet_handle_send(sd);
2565 }
2566
2567 /*
2568  * Source not specified
2569  * Remote destination
2570  * Non-MR peer
2571  *
2572  * Must send to the specified peer NID using the same source NID that
2573  * we've used before. If it's the first time to talk to that peer then
2574  * find the source NI and assign it as preferred to that peer
2575  */
2576 static int
2577 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2578 {
2579         int rc;
2580         struct lnet_peer_ni *gw_lpni = NULL;
2581         struct lnet_peer *gw_peer = NULL;
2582
2583         /*
2584          * Let's see if we have a preferred NI to talk to this NMR peer
2585          */
2586         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2587                                                               sd->sd_cpt);
2588
2589         /*
2590          * find the router and that'll find the best NI if we didn't find
2591          * it already.
2592          */
2593         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2594                                           &gw_peer);
2595         if (rc)
2596                 return rc;
2597
2598         /*
2599          * set the best_ni we've chosen as the preferred one for
2600          * this peer
2601          */
2602         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2603
2604         /* we'll be sending to the gw */
2605         sd->sd_best_lpni = gw_lpni;
2606         sd->sd_peer = gw_peer;
2607
2608         return lnet_handle_send(sd);
2609 }
2610
2611 static int
2612 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2613 {
2614         /*
2615          * turn off the SND_RESP bit.
2616          * It will be checked in the case handling
2617          */
2618         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2619
2620         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2621                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2622                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2623                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2624                 libcfs_nid2str(sd->sd_dst_nid),
2625                 (send_case & LOCAL_DST) ? "local" : "routed");
2626
2627         switch (send_case) {
2628         /*
2629          * For all cases where the source is specified, we should always
2630          * use the destination NID, whether it's an MR destination or not,
2631          * since we're continuing a series of related messages for the
2632          * same RPC
2633          */
2634         case SRC_SPEC_LOCAL_NMR_DST:
2635                 return lnet_handle_spec_local_nmr_dst(sd);
2636         case SRC_SPEC_LOCAL_MR_DST:
2637                 return lnet_handle_spec_local_mr_dst(sd);
2638         case SRC_SPEC_ROUTER_NMR_DST:
2639         case SRC_SPEC_ROUTER_MR_DST:
2640                 return lnet_handle_spec_router_dst(sd);
2641         case SRC_ANY_LOCAL_NMR_DST:
2642                 return lnet_handle_any_local_nmr_dst(sd);
2643         case SRC_ANY_LOCAL_MR_DST:
2644         case SRC_ANY_ROUTER_MR_DST:
2645                 return lnet_handle_any_mr_dst(sd);
2646         case SRC_ANY_ROUTER_NMR_DST:
2647                 return lnet_handle_any_router_nmr_dst(sd);
2648         default:
2649                 CERROR("Unknown send case\n");
2650                 return -1;
2651         }
2652 }
2653
2654 static int
2655 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2656                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2657 {
2658         struct lnet_peer_ni *lpni;
2659         struct lnet_peer *peer;
2660         struct lnet_send_data send_data;
2661         int cpt, rc;
2662         int md_cpt;
2663         __u32 send_case = 0;
2664         bool final_hop;
2665         bool mr_forwarding_allowed;
2666
2667         memset(&send_data, 0, sizeof(send_data));
2668
2669         /*
2670          * get an initial CPT to use for locking. The idea here is not to
2671          * serialize the calls to select_pathway, so that as many
2672          * operations can run concurrently as possible. To do that we use
2673          * the CPT where this call is being executed. Later on when we
2674          * determine the CPT to use in lnet_message_commit, we switch the
2675          * lock and check if there was any configuration change.  If none,
2676          * then we proceed, if there is, then we restart the operation.
2677          */
2678         cpt = lnet_net_lock_current();
2679
2680         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2681         if (md_cpt == CFS_CPT_ANY)
2682                 md_cpt = cpt;
2683
2684 again:
2685
2686         /*
2687          * If we're being asked to send to the loopback interface, there
2688          * is no need to go through any selection. We can just shortcut
2689          * the entire process and send over lolnd
2690          */
2691         send_data.sd_msg = msg;
2692         send_data.sd_cpt = cpt;
2693         if (dst_nid == LNET_NID_LO_0) {
2694                 rc = lnet_handle_lo_send(&send_data);
2695                 lnet_net_unlock(cpt);
2696                 return rc;
2697         }
2698
2699         /*
2700          * find an existing peer_ni, or create one and mark it as having been
2701          * created due to network traffic. This call will create the
2702          * peer->peer_net->peer_ni tree.
2703          */
2704         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2705         if (IS_ERR(lpni)) {
2706                 lnet_net_unlock(cpt);
2707                 return PTR_ERR(lpni);
2708         }
2709
2710         /*
2711          * Cache the original src_nid and rtr_nid. If we need to resend the
2712          * message then we'll need to know whether the src_nid was originally
2713          * specified for this message. If it was originally specified,
2714          * then we need to keep using the same src_nid since it's
2715          * continuing the same sequence of messages. Similarly, rtr_nid will
2716          * affect our choice of next hop.
2717          */
2718         msg->msg_src_nid_param = src_nid;
2719         msg->msg_rtr_nid_param = rtr_nid;
2720
2721         /*
2722          * If necessary, perform discovery on the peer that owns this peer_ni.
2723          * Note, this can result in the ownership of this peer_ni changing
2724          * to another peer object.
2725          */
2726         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2727         if (rc) {
2728                 lnet_peer_ni_decref_locked(lpni);
2729                 lnet_net_unlock(cpt);
2730                 return rc;
2731         }
2732         lnet_peer_ni_decref_locked(lpni);
2733
2734         peer = lpni->lpni_peer_net->lpn_peer;
2735
2736         /*
2737          * Identify the different send cases
2738          */
2739         if (src_nid == LNET_NID_ANY)
2740                 send_case |= SRC_ANY;
2741         else
2742                 send_case |= SRC_SPEC;
2743
2744         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2745                 send_case |= LOCAL_DST;
2746         else
2747                 send_case |= REMOTE_DST;
2748
2749         final_hop = false;
2750         if (msg->msg_routing && (send_case & LOCAL_DST))
2751                 final_hop = true;
2752
2753         /* Determine whether to allow MR forwarding for this message.
2754          * NB: MR forwarding is allowed if the message originator and the
2755          * destination are both MR capable, and the destination lpni that was
2756          * originally chosen by the originator is unhealthy or down.
2757          * We check the MR capability of the destination further below
2758          */
2759         mr_forwarding_allowed = false;
2760         if (final_hop) {
2761                 struct lnet_peer *src_lp;
2762                 struct lnet_peer_ni *src_lpni;
2763
2764                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2765                                                   LNET_NID_ANY, cpt);
2766                 /* We don't fail the send if we hit any errors here. We'll just
2767                  * try to send it via non-multi-rail criteria
2768                  */
2769                 if (!IS_ERR(src_lpni)) {
2770                         src_lp = lpni->lpni_peer_net->lpn_peer;
2771                         if (lnet_peer_is_multi_rail(src_lp) &&
2772                             !lnet_is_peer_ni_alive(lpni))
2773                                 mr_forwarding_allowed = true;
2774
2775                 }
2776                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
2777                        mr_forwarding_allowed ? "allowed" : "not allowed");
2778         }
2779
2780         /*
2781          * Deal with the peer as NMR in the following cases:
2782          * 1. the peer is NMR
2783          * 2. We're trying to recover a specific peer NI
2784          * 3. I'm a router sending to the final destination and MR forwarding is
2785          *    not allowed for this message (as determined above).
2786          *    In this case the source of the message would've
2787          *    already selected the final destination so my job
2788          *    is to honor the selection.
2789          */
2790         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2791             (final_hop && !mr_forwarding_allowed))
2792                 send_case |= NMR_DST;
2793         else
2794                 send_case |= MR_DST;
2795
2796         if (lnet_msg_is_response(msg))
2797                 send_case |= SND_RESP;
2798
2799         /* assign parameters to the send_data */
2800         send_data.sd_rtr_nid = rtr_nid;
2801         send_data.sd_src_nid = src_nid;
2802         send_data.sd_dst_nid = dst_nid;
2803         send_data.sd_best_lpni = lpni;
2804         /*
2805          * keep a pointer to the final destination in case we're going to
2806          * route, so we'll need to access it later
2807          */
2808         send_data.sd_final_dst_lpni = lpni;
2809         send_data.sd_peer = peer;
2810         send_data.sd_md_cpt = md_cpt;
2811         send_data.sd_send_case = send_case;
2812
2813         rc = lnet_handle_send_case_locked(&send_data);
2814
2815         /*
2816          * Update the local cpt since send_data.sd_cpt might've been
2817          * updated as a result of calling lnet_handle_send_case_locked().
2818          */
2819         cpt = send_data.sd_cpt;
2820
2821         if (rc == REPEAT_SEND)
2822                 goto again;
2823
2824         lnet_net_unlock(cpt);
2825
2826         return rc;
2827 }
2828
2829 int
2830 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2831 {
2832         lnet_nid_t              dst_nid = msg->msg_target.nid;
2833         int                     rc;
2834
2835         /*
2836          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2837          * but we might want to use pre-determined router for ACK/REPLY
2838          * in the future
2839          */
2840         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2841         LASSERT(msg->msg_txpeer == NULL);
2842         LASSERT(msg->msg_txni == NULL);
2843         LASSERT(!msg->msg_sending);
2844         LASSERT(!msg->msg_target_is_router);
2845         LASSERT(!msg->msg_receiving);
2846
2847         msg->msg_sending = 1;
2848
2849         LASSERT(!msg->msg_tx_committed);
2850
2851         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2852         if (rc < 0) {
2853                 if (rc == -EHOSTUNREACH)
2854                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2855                 else
2856                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2857                 return rc;
2858         }
2859
2860         if (rc == LNET_CREDIT_OK)
2861                 lnet_ni_send(msg->msg_txni, msg);
2862
2863         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2864         return 0;
2865 }
2866
2867 enum lnet_mt_event_type {
2868         MT_TYPE_LOCAL_NI = 0,
2869         MT_TYPE_PEER_NI
2870 };
2871
2872 struct lnet_mt_event_info {
2873         enum lnet_mt_event_type mt_type;
2874         lnet_nid_t mt_nid;
2875 };
2876
2877 /* called with res_lock held */
2878 void
2879 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2880 {
2881         struct lnet_rsp_tracker *rspt;
2882
2883         /*
2884          * msg has a refcount on the MD so the MD is not going away.
2885          * The rspt queue for the cpt is protected by
2886          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2887          */
2888         if (!md->md_rspt_ptr)
2889                 return;
2890
2891         rspt = md->md_rspt_ptr;
2892
2893         /* debug code */
2894         LASSERT(rspt->rspt_cpt == cpt);
2895
2896         md->md_rspt_ptr = NULL;
2897
2898         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2899                 /*
2900                  * The monitor thread has invalidated this handle because the
2901                  * response timed out, but it failed to lookup the MD. That
2902                  * means this response tracker is on the zombie list. We can
2903                  * safely remove it under the resource lock (held by caller) and
2904                  * free the response tracker block.
2905                  */
2906                 list_del(&rspt->rspt_on_list);
2907                 lnet_rspt_free(rspt, cpt);
2908         } else {
2909                 /*
2910                  * invalidate the handle to indicate that a response has been
2911                  * received, which will then lead the monitor thread to clean up
2912                  * the rspt block.
2913                  */
2914                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2915         }
2916 }
2917
2918 void
2919 lnet_clean_zombie_rstqs(void)
2920 {
2921         struct lnet_rsp_tracker *rspt, *tmp;
2922         int i;
2923
2924         cfs_cpt_for_each(i, lnet_cpt_table()) {
2925                 list_for_each_entry_safe(rspt, tmp,
2926                                          the_lnet.ln_mt_zombie_rstqs[i],
2927                                          rspt_on_list) {
2928                         list_del(&rspt->rspt_on_list);
2929                         lnet_rspt_free(rspt, i);
2930                 }
2931         }
2932
2933         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2934 }
2935
2936 static void
2937 lnet_finalize_expired_responses(void)
2938 {
2939         struct lnet_libmd *md;
2940         struct lnet_rsp_tracker *rspt, *tmp;
2941         ktime_t now;
2942         int i;
2943
2944         if (the_lnet.ln_mt_rstq == NULL)
2945                 return;
2946
2947         cfs_cpt_for_each(i, lnet_cpt_table()) {
2948                 LIST_HEAD(local_queue);
2949
2950                 lnet_net_lock(i);
2951                 if (!the_lnet.ln_mt_rstq[i]) {
2952                         lnet_net_unlock(i);
2953                         continue;
2954                 }
2955                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2956                 lnet_net_unlock(i);
2957
2958                 now = ktime_get();
2959
2960                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2961                         /*
2962                          * The rspt mdh will be invalidated when a response
2963                          * is received or whenever we want to discard the
2964                          * block the monitor thread will walk the queue
2965                          * and clean up any rsts with an invalid mdh.
2966                          * The monitor thread will walk the queue until
2967                          * the first unexpired rspt block. This means that
2968                          * some rspt blocks which received their
2969                          * corresponding responses will linger in the
2970                          * queue until they are cleaned up eventually.
2971                          */
2972                         lnet_res_lock(i);
2973                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2974                                 lnet_res_unlock(i);
2975                                 list_del(&rspt->rspt_on_list);
2976                                 lnet_rspt_free(rspt, i);
2977                                 continue;
2978                         }
2979
2980                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2981                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2982                                 struct lnet_peer_ni *lpni;
2983                                 lnet_nid_t nid;
2984
2985                                 md = lnet_handle2md(&rspt->rspt_mdh);
2986                                 if (!md) {
2987                                         /* MD has been queued for unlink, but
2988                                          * rspt hasn't been detached (Note we've
2989                                          * checked above that the rspt_mdh is
2990                                          * valid). Since we cannot lookup the MD
2991                                          * we're unable to detach the rspt
2992                                          * ourselves. Thus, move the rspt to the
2993                                          * zombie list where we'll wait for
2994                                          * either:
2995                                          *   1. The remaining operations on the
2996                                          *   MD to complete. In this case the
2997                                          *   final operation will result in
2998                                          *   lnet_msg_detach_md()->
2999                                          *   lnet_detach_rsp_tracker() where
3000                                          *   we will clean up this response
3001                                          *   tracker.
3002                                          *   2. LNet to shutdown. In this case
3003                                          *   we'll wait until after all LND Nets
3004                                          *   have shutdown and then we can
3005                                          *   safely free any remaining response
3006                                          *   tracker blocks on the zombie list.
3007                                          * Note: We need to hold the resource
3008                                          * lock when adding to the zombie list
3009                                          * because we may have concurrent access
3010                                          * with lnet_detach_rsp_tracker().
3011                                          */
3012                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
3013                                         list_move(&rspt->rspt_on_list,
3014                                                   the_lnet.ln_mt_zombie_rstqs[i]);
3015                                         lnet_res_unlock(i);
3016                                         continue;
3017                                 }
3018                                 LASSERT(md->md_rspt_ptr == rspt);
3019                                 md->md_rspt_ptr = NULL;
3020                                 lnet_res_unlock(i);
3021
3022                                 LNetMDUnlink(rspt->rspt_mdh);
3023
3024                                 nid = rspt->rspt_next_hop_nid;
3025
3026                                 list_del(&rspt->rspt_on_list);
3027                                 lnet_rspt_free(rspt, i);
3028
3029                                 /* If we're shutting down we just want to clean
3030                                  * up the rspt blocks
3031                                  */
3032                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3033                                         continue;
3034
3035                                 lnet_net_lock(i);
3036                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3037                                 lnet_net_unlock(i);
3038
3039                                 CDEBUG(D_NET,
3040                                        "Response timeout: md = %p: nid = %s\n",
3041                                        md, libcfs_nid2str(nid));
3042
3043                                 /*
3044                                  * If there is a timeout on the response
3045                                  * from the next hop decrement its health
3046                                  * value so that we don't use it
3047                                  */
3048                                 lnet_net_lock(0);
3049                                 lpni = lnet_find_peer_ni_locked(nid);
3050                                 if (lpni) {
3051                                         lnet_handle_remote_failure_locked(lpni);
3052                                         lnet_peer_ni_decref_locked(lpni);
3053                                 }
3054                                 lnet_net_unlock(0);
3055                         } else {
3056                                 lnet_res_unlock(i);
3057                                 break;
3058                         }
3059                 }
3060
3061                 if (!list_empty(&local_queue)) {
3062                         lnet_net_lock(i);
3063                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3064                         lnet_net_unlock(i);
3065                 }
3066         }
3067 }
3068
3069 static void
3070 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3071 {
3072         struct lnet_msg *msg;
3073
3074         while (!list_empty(resendq)) {
3075                 struct lnet_peer_ni *lpni;
3076
3077                 msg = list_entry(resendq->next, struct lnet_msg,
3078                                  msg_list);
3079
3080                 list_del_init(&msg->msg_list);
3081
3082                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3083                 if (!lpni) {
3084                         lnet_net_unlock(cpt);
3085                         CERROR("Expected that a peer is already created for %s\n",
3086                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3087                         msg->msg_no_resend = true;
3088                         lnet_finalize(msg, -EFAULT);
3089                         lnet_net_lock(cpt);
3090                 } else {
3091                         int rc;
3092
3093                         lnet_peer_ni_decref_locked(lpni);
3094
3095                         lnet_net_unlock(cpt);
3096                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3097                                libcfs_nid2str(msg->msg_src_nid_param),
3098                                libcfs_id2str(msg->msg_target),
3099                                lnet_msgtyp2str(msg->msg_type),
3100                                msg->msg_recovery,
3101                                msg->msg_retry_count);
3102                         rc = lnet_send(msg->msg_src_nid_param, msg,
3103                                        msg->msg_rtr_nid_param);
3104                         if (rc) {
3105                                 CERROR("Error sending %s to %s: %d\n",
3106                                        lnet_msgtyp2str(msg->msg_type),
3107                                        libcfs_id2str(msg->msg_target), rc);
3108                                 msg->msg_no_resend = true;
3109                                 lnet_finalize(msg, rc);
3110                         }
3111                         lnet_net_lock(cpt);
3112                         if (!rc)
3113                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3114                 }
3115         }
3116 }
3117
3118 static void
3119 lnet_resend_pending_msgs(void)
3120 {
3121         int i;
3122
3123         cfs_cpt_for_each(i, lnet_cpt_table()) {
3124                 lnet_net_lock(i);
3125                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3126                 lnet_net_unlock(i);
3127         }
3128 }
3129
3130 /* called with cpt and ni_lock held */
3131 static void
3132 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3133 {
3134         struct lnet_handle_md recovery_mdh;
3135
3136         LNetInvalidateMDHandle(&recovery_mdh);
3137
3138         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3139             force) {
3140                 recovery_mdh = ni->ni_ping_mdh;
3141                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3142         }
3143         lnet_ni_unlock(ni);
3144         lnet_net_unlock(cpt);
3145         if (!LNetMDHandleIsInvalid(recovery_mdh))
3146                 LNetMDUnlink(recovery_mdh);
3147         lnet_net_lock(cpt);
3148         lnet_ni_lock(ni);
3149 }
3150
3151 static void
3152 lnet_recover_local_nis(void)
3153 {
3154         struct lnet_mt_event_info *ev_info;
3155         LIST_HEAD(processed_list);
3156         LIST_HEAD(local_queue);
3157         struct lnet_handle_md mdh;
3158         struct lnet_ni *tmp;
3159         struct lnet_ni *ni;
3160         lnet_nid_t nid;
3161         int healthv;
3162         int rc;
3163
3164         /*
3165          * splice the recovery queue on a local queue. We will iterate
3166          * through the local queue and update it as needed. Once we're
3167          * done with the traversal, we'll splice the local queue back on
3168          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3169          * will be traversed in the next iteration.
3170          */
3171         lnet_net_lock(0);
3172         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3173                          &local_queue);
3174         lnet_net_unlock(0);
3175
3176         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3177                 /*
3178                  * if an NI is being deleted or it is now healthy, there
3179                  * is no need to keep it around in the recovery queue.
3180                  * The monitor thread is the only thread responsible for
3181                  * removing the NI from the recovery queue.
3182                  * Multiple threads can be adding NIs to the recovery
3183                  * queue.
3184                  */
3185                 healthv = atomic_read(&ni->ni_healthv);
3186
3187                 lnet_net_lock(0);
3188                 lnet_ni_lock(ni);
3189                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3190                     healthv == LNET_MAX_HEALTH_VALUE) {
3191                         list_del_init(&ni->ni_recovery);
3192                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3193                         lnet_ni_unlock(ni);
3194                         lnet_ni_decref_locked(ni, 0);
3195                         lnet_net_unlock(0);
3196                         continue;
3197                 }
3198
3199                 /*
3200                  * if the local NI failed recovery we must unlink the md.
3201                  * But we want to keep the local_ni on the recovery queue
3202                  * so we can continue the attempts to recover it.
3203                  */
3204                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3205                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3206                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3207                 }
3208
3209                 lnet_ni_unlock(ni);
3210                 lnet_net_unlock(0);
3211
3212
3213                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3214                        libcfs_nid2str(ni->ni_nid));
3215
3216                 lnet_ni_lock(ni);
3217                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3218                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3219                         lnet_ni_unlock(ni);
3220
3221                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3222                         if (!ev_info) {
3223                                 CERROR("out of memory. Can't recover %s\n",
3224                                        libcfs_nid2str(ni->ni_nid));
3225                                 lnet_ni_lock(ni);
3226                                 ni->ni_recovery_state &=
3227                                   ~LNET_NI_RECOVERY_PENDING;
3228                                 lnet_ni_unlock(ni);
3229                                 continue;
3230                         }
3231
3232                         mdh = ni->ni_ping_mdh;
3233                         /*
3234                          * Invalidate the ni mdh in case it's deleted.
3235                          * We'll unlink the mdh in this case below.
3236                          */
3237                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3238                         nid = ni->ni_nid;
3239
3240                         /*
3241                          * remove the NI from the local queue and drop the
3242                          * reference count to it while we're recovering
3243                          * it. The reason for that, is that the NI could
3244                          * be deleted, and the way the code is structured
3245                          * is if we don't drop the NI, then the deletion
3246                          * code will enter a loop waiting for the
3247                          * reference count to be removed while holding the
3248                          * ln_mutex_lock(). When we look up the peer to
3249                          * send to in lnet_select_pathway() we will try to
3250                          * lock the ln_mutex_lock() as well, leading to
3251                          * a deadlock. By dropping the refcount and
3252                          * removing it from the list, we allow for the NI
3253                          * to be removed, then we use the cached NID to
3254                          * look it up again. If it's gone, then we just
3255                          * continue examining the rest of the queue.
3256                          */
3257                         lnet_net_lock(0);
3258                         list_del_init(&ni->ni_recovery);
3259                         lnet_ni_decref_locked(ni, 0);
3260                         lnet_net_unlock(0);
3261
3262                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3263                         ev_info->mt_nid = nid;
3264                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3265                                             ev_info, the_lnet.ln_mt_handler,
3266                                             true);
3267                         /* lookup the nid again */
3268                         lnet_net_lock(0);
3269                         ni = lnet_nid2ni_locked(nid, 0);
3270                         if (!ni) {
3271                                 /*
3272                                  * the NI has been deleted when we dropped
3273                                  * the ref count
3274                                  */
3275                                 lnet_net_unlock(0);
3276                                 LNetMDUnlink(mdh);
3277                                 continue;
3278                         }
3279                         /*
3280                          * Same note as in lnet_recover_peer_nis(). When
3281                          * we're sending the ping, the NI is free to be
3282                          * deleted or manipulated. By this point it
3283                          * could've been added back on the recovery queue,
3284                          * and a refcount taken on it.
3285                          * So we can't just add it blindly again or we'll
3286                          * corrupt the queue. We must check under lock if
3287                          * it's not on any list and if not then add it
3288                          * to the processed list, which will eventually be
3289                          * spliced back on to the recovery queue.
3290                          */
3291                         ni->ni_ping_mdh = mdh;
3292                         if (list_empty(&ni->ni_recovery)) {
3293                                 list_add_tail(&ni->ni_recovery, &processed_list);
3294                                 lnet_ni_addref_locked(ni, 0);
3295                         }
3296                         lnet_net_unlock(0);
3297
3298                         lnet_ni_lock(ni);
3299                         if (rc)
3300                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3301                 }
3302                 lnet_ni_unlock(ni);
3303         }
3304
3305         /*
3306          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3307          * reexamined in the next iteration.
3308          */
3309         list_splice_init(&processed_list, &local_queue);
3310         lnet_net_lock(0);
3311         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3312         lnet_net_unlock(0);
3313 }
3314
3315 static int
3316 lnet_resendqs_create(void)
3317 {
3318         struct list_head **resendqs;
3319         resendqs = lnet_create_array_of_queues();
3320
3321         if (!resendqs)
3322                 return -ENOMEM;
3323
3324         lnet_net_lock(LNET_LOCK_EX);
3325         the_lnet.ln_mt_resendqs = resendqs;
3326         lnet_net_unlock(LNET_LOCK_EX);
3327
3328         return 0;
3329 }
3330
3331 static void
3332 lnet_clean_local_ni_recoveryq(void)
3333 {
3334         struct lnet_ni *ni;
3335
3336         /* This is only called when the monitor thread has stopped */
3337         lnet_net_lock(0);
3338
3339         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3340                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3341                                 struct lnet_ni, ni_recovery);
3342                 list_del_init(&ni->ni_recovery);
3343                 lnet_ni_lock(ni);
3344                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3345                 lnet_ni_unlock(ni);
3346                 lnet_ni_decref_locked(ni, 0);
3347         }
3348
3349         lnet_net_unlock(0);
3350 }
3351
3352 static void
3353 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3354                                      bool force)
3355 {
3356         struct lnet_handle_md recovery_mdh;
3357
3358         LNetInvalidateMDHandle(&recovery_mdh);
3359
3360         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3361                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3362                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3363         }
3364         spin_unlock(&lpni->lpni_lock);
3365         lnet_net_unlock(cpt);
3366         if (!LNetMDHandleIsInvalid(recovery_mdh))
3367                 LNetMDUnlink(recovery_mdh);
3368         lnet_net_lock(cpt);
3369         spin_lock(&lpni->lpni_lock);
3370 }
3371
3372 static void
3373 lnet_clean_peer_ni_recoveryq(void)
3374 {
3375         struct lnet_peer_ni *lpni, *tmp;
3376
3377         lnet_net_lock(LNET_LOCK_EX);
3378
3379         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3380                                  lpni_recovery) {
3381                 list_del_init(&lpni->lpni_recovery);
3382                 spin_lock(&lpni->lpni_lock);
3383                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3384                 spin_unlock(&lpni->lpni_lock);
3385                 lnet_peer_ni_decref_locked(lpni);
3386         }
3387
3388         lnet_net_unlock(LNET_LOCK_EX);
3389 }
3390
3391 static void
3392 lnet_clean_resendqs(void)
3393 {
3394         struct lnet_msg *msg, *tmp;
3395         LIST_HEAD(msgs);
3396         int i;
3397
3398         cfs_cpt_for_each(i, lnet_cpt_table()) {
3399                 lnet_net_lock(i);
3400                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3401                 lnet_net_unlock(i);
3402                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3403                         list_del_init(&msg->msg_list);
3404                         msg->msg_no_resend = true;
3405                         lnet_finalize(msg, -ESHUTDOWN);
3406                 }
3407         }
3408
3409         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3410 }
3411
3412 static void
3413 lnet_recover_peer_nis(void)
3414 {
3415         struct lnet_mt_event_info *ev_info;
3416         LIST_HEAD(processed_list);
3417         LIST_HEAD(local_queue);
3418         struct lnet_handle_md mdh;
3419         struct lnet_peer_ni *lpni;
3420         struct lnet_peer_ni *tmp;
3421         lnet_nid_t nid;
3422         int healthv;
3423         int rc;
3424
3425         /*
3426          * Always use cpt 0 for locking across all interactions with
3427          * ln_mt_peerNIRecovq
3428          */
3429         lnet_net_lock(0);
3430         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3431                          &local_queue);
3432         lnet_net_unlock(0);
3433
3434         list_for_each_entry_safe(lpni, tmp, &local_queue,
3435                                  lpni_recovery) {
3436                 /*
3437                  * The same protection strategy is used here as is in the
3438                  * local recovery case.
3439                  */
3440                 lnet_net_lock(0);
3441                 healthv = atomic_read(&lpni->lpni_healthv);
3442                 spin_lock(&lpni->lpni_lock);
3443                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3444                     healthv == LNET_MAX_HEALTH_VALUE) {
3445                         list_del_init(&lpni->lpni_recovery);
3446                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3447                         spin_unlock(&lpni->lpni_lock);
3448                         lnet_peer_ni_decref_locked(lpni);
3449                         lnet_net_unlock(0);
3450                         continue;
3451                 }
3452
3453                 /*
3454                  * If the peer NI has failed recovery we must unlink the
3455                  * md. But we want to keep the peer ni on the recovery
3456                  * queue so we can try to continue recovering it
3457                  */
3458                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3459                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3460                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3461                 }
3462
3463                 spin_unlock(&lpni->lpni_lock);
3464                 lnet_net_unlock(0);
3465
3466                 /*
3467                  * NOTE: we're racing with peer deletion from user space.
3468                  * It's possible that a peer is deleted after we check its
3469                  * state. In this case the recovery can create a new peer
3470                  */
3471                 spin_lock(&lpni->lpni_lock);
3472                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3473                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3474                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3475                         spin_unlock(&lpni->lpni_lock);
3476
3477                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3478                         if (!ev_info) {
3479                                 CERROR("out of memory. Can't recover %s\n",
3480                                        libcfs_nid2str(lpni->lpni_nid));
3481                                 spin_lock(&lpni->lpni_lock);
3482                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3483                                 spin_unlock(&lpni->lpni_lock);
3484                                 continue;
3485                         }
3486
3487                         /* look at the comments in lnet_recover_local_nis() */
3488                         mdh = lpni->lpni_recovery_ping_mdh;
3489                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3490                         nid = lpni->lpni_nid;
3491                         lnet_net_lock(0);
3492                         list_del_init(&lpni->lpni_recovery);
3493                         lnet_peer_ni_decref_locked(lpni);
3494                         lnet_net_unlock(0);
3495
3496                         ev_info->mt_type = MT_TYPE_PEER_NI;
3497                         ev_info->mt_nid = nid;
3498                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3499                                             ev_info, the_lnet.ln_mt_handler,
3500                                             true);
3501                         lnet_net_lock(0);
3502                         /*
3503                          * lnet_find_peer_ni_locked() grabs a refcount for
3504                          * us. No need to take it explicitly.
3505                          */
3506                         lpni = lnet_find_peer_ni_locked(nid);
3507                         if (!lpni) {
3508                                 lnet_net_unlock(0);
3509                                 LNetMDUnlink(mdh);
3510                                 continue;
3511                         }
3512
3513                         lpni->lpni_recovery_ping_mdh = mdh;
3514                         /*
3515                          * While we're unlocked the lpni could've been
3516                          * readded on the recovery queue. In this case we
3517                          * don't need to add it to the local queue, since
3518                          * it's already on there and the thread that added
3519                          * it would've incremented the refcount on the
3520                          * peer, which means we need to decref the refcount
3521                          * that was implicitly grabbed by find_peer_ni_locked.
3522                          * Otherwise, if the lpni is still not on
3523                          * the recovery queue, then we'll add it to the
3524                          * processed list.
3525                          */
3526                         if (list_empty(&lpni->lpni_recovery))
3527                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3528                         else
3529                                 lnet_peer_ni_decref_locked(lpni);
3530                         lnet_net_unlock(0);
3531
3532                         spin_lock(&lpni->lpni_lock);
3533                         if (rc)
3534                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3535                 }
3536                 spin_unlock(&lpni->lpni_lock);
3537         }
3538
3539         list_splice_init(&processed_list, &local_queue);
3540         lnet_net_lock(0);
3541         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3542         lnet_net_unlock(0);
3543 }
3544
3545 static int
3546 lnet_monitor_thread(void *arg)
3547 {
3548         time64_t recovery_timeout = 0;
3549         time64_t rsp_timeout = 0;
3550         int interval;
3551         time64_t now;
3552
3553         wait_for_completion(&the_lnet.ln_started);
3554         /*
3555          * The monitor thread takes care of the following:
3556          *  1. Checks the aliveness of routers
3557          *  2. Checks if there are messages on the resend queue to resend
3558          *     them.
3559          *  3. Check if there are any NIs on the local recovery queue and
3560          *     pings them
3561          *  4. Checks if there are any NIs on the remote recovery queue
3562          *     and pings them.
3563          */
3564         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3565                 now = ktime_get_real_seconds();
3566
3567                 if (lnet_router_checker_active())
3568                         lnet_check_routers();
3569
3570                 lnet_resend_pending_msgs();
3571
3572                 if (now >= rsp_timeout) {
3573                         lnet_finalize_expired_responses();
3574                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3575                 }
3576
3577                 if (now >= recovery_timeout) {
3578                         lnet_recover_local_nis();
3579                         lnet_recover_peer_nis();
3580                         recovery_timeout = now + lnet_recovery_interval;
3581                 }
3582
3583                 /*
3584                  * TODO do we need to check if we should sleep without
3585                  * timeout?  Technically, an active system will always
3586                  * have messages in flight so this check will always
3587                  * evaluate to false. And on an idle system do we care
3588                  * if we wake up every 1 second? Although, we've seen
3589                  * cases where we get a complaint that an idle thread
3590                  * is waking up unnecessarily.
3591                  *
3592                  * Take into account the current net_count when you wake
3593                  * up for alive router checking, since we need to check
3594                  * possibly as many networks as we have configured.
3595                  */
3596                 interval = min(lnet_recovery_interval,
3597                                min((unsigned int) alive_router_check_interval /
3598                                         lnet_current_net_count,
3599                                    lnet_transaction_timeout / 2));
3600                 wait_for_completion_interruptible_timeout(
3601                         &the_lnet.ln_mt_wait_complete,
3602                         cfs_time_seconds(interval));
3603                 /* Must re-init the completion before testing anything,
3604                  * including ln_mt_state.
3605                  */
3606                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3607         }
3608
3609         /* Shutting down */
3610         lnet_net_lock(LNET_LOCK_EX);
3611         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3612         lnet_net_unlock(LNET_LOCK_EX);
3613
3614         /* signal that the monitor thread is exiting */
3615         up(&the_lnet.ln_mt_signal);
3616
3617         return 0;
3618 }
3619
3620 /*
3621  * lnet_send_ping
3622  * Sends a ping.
3623  * Returns == 0 if success
3624  * Returns > 0 if LNetMDBind or prior fails
3625  * Returns < 0 if LNetGet fails
3626  */
3627 int
3628 lnet_send_ping(lnet_nid_t dest_nid,
3629                struct lnet_handle_md *mdh, int nnis,
3630                void *user_data, lnet_handler_t handler, bool recovery)
3631 {
3632         struct lnet_md md = { NULL };
3633         struct lnet_process_id id;
3634         struct lnet_ping_buffer *pbuf;
3635         int rc;
3636
3637         if (dest_nid == LNET_NID_ANY) {
3638                 rc = -EHOSTUNREACH;
3639                 goto fail_error;
3640         }
3641
3642         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3643         if (!pbuf) {
3644                 rc = ENOMEM;
3645                 goto fail_error;
3646         }
3647
3648         /* initialize md content */
3649         md.start     = &pbuf->pb_info;
3650         md.length    = LNET_PING_INFO_SIZE(nnis);
3651         md.threshold = 2; /* GET/REPLY */
3652         md.max_size  = 0;
3653         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3654         md.user_ptr  = user_data;
3655         md.handler   = handler;
3656
3657         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3658         if (rc) {
3659                 lnet_ping_buffer_decref(pbuf);
3660                 CERROR("Can't bind MD: %d\n", rc);
3661                 rc = -rc; /* change the rc to positive */
3662                 goto fail_error;
3663         }
3664         id.pid = LNET_PID_LUSTRE;
3665         id.nid = dest_nid;
3666
3667         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3668                      LNET_RESERVED_PORTAL,
3669                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3670
3671         if (rc)
3672                 goto fail_unlink_md;
3673
3674         return 0;
3675
3676 fail_unlink_md:
3677         LNetMDUnlink(*mdh);
3678         LNetInvalidateMDHandle(mdh);
3679 fail_error:
3680         return rc;
3681 }
3682
3683 static void
3684 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3685                            int status, bool send, bool unlink_event)
3686 {
3687         lnet_nid_t nid = ev_info->mt_nid;
3688
3689         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3690                 struct lnet_ni *ni;
3691
3692                 lnet_net_lock(0);
3693                 ni = lnet_nid2ni_locked(nid, 0);
3694                 if (!ni) {
3695                         lnet_net_unlock(0);
3696                         return;
3697                 }
3698                 lnet_ni_lock(ni);
3699                 if (!send || (send && status != 0))
3700                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3701                 if (status)
3702                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3703                 lnet_ni_unlock(ni);
3704                 lnet_net_unlock(0);
3705
3706                 if (status != 0) {
3707                         CERROR("local NI (%s) recovery failed with %d\n",
3708                                libcfs_nid2str(nid), status);
3709                         return;
3710                 }
3711                 /*
3712                  * need to increment healthv for the ni here, because in
3713                  * the lnet_finalize() path we don't have access to this
3714                  * NI. And in order to get access to it, we'll need to
3715                  * carry forward too much information.
3716                  * In the peer case, it'll naturally be incremented
3717                  */
3718                 if (!unlink_event)
3719                         lnet_inc_healthv(&ni->ni_healthv,
3720                                          lnet_health_sensitivity);
3721         } else {
3722                 struct lnet_peer_ni *lpni;
3723                 int cpt;
3724
3725                 cpt = lnet_net_lock_current();
3726                 lpni = lnet_find_peer_ni_locked(nid);
3727                 if (!lpni) {
3728                         lnet_net_unlock(cpt);
3729                         return;
3730                 }
3731                 spin_lock(&lpni->lpni_lock);
3732                 if (!send || (send && status != 0))
3733                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3734                 if (status)
3735                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3736                 spin_unlock(&lpni->lpni_lock);
3737                 lnet_peer_ni_decref_locked(lpni);
3738                 lnet_net_unlock(cpt);
3739
3740                 if (status != 0)
3741                         CERROR("peer NI (%s) recovery failed with %d\n",
3742                                libcfs_nid2str(nid), status);
3743         }
3744 }
3745
3746 void
3747 lnet_mt_event_handler(struct lnet_event *event)
3748 {
3749         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3750         struct lnet_ping_buffer *pbuf;
3751
3752         /* TODO: remove assert */
3753         LASSERT(event->type == LNET_EVENT_REPLY ||
3754                 event->type == LNET_EVENT_SEND ||
3755                 event->type == LNET_EVENT_UNLINK);
3756
3757         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3758                event->status);
3759
3760         switch (event->type) {
3761         case LNET_EVENT_UNLINK:
3762                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3763                        libcfs_nid2str(ev_info->mt_nid));
3764                 /* fallthrough */
3765         case LNET_EVENT_REPLY:
3766                 lnet_handle_recovery_reply(ev_info, event->status, false,
3767                                            event->type == LNET_EVENT_UNLINK);
3768                 break;
3769         case LNET_EVENT_SEND:
3770                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3771                                libcfs_nid2str(ev_info->mt_nid),
3772                                (event->status) ? "unsuccessfully" :
3773                                "successfully", event->status);
3774                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3775                 break;
3776         default:
3777                 CERROR("Unexpected event: %d\n", event->type);
3778                 break;
3779         }
3780         if (event->unlinked) {
3781                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3782                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
3783                 lnet_ping_buffer_decref(pbuf);
3784         }
3785 }
3786
3787 static int
3788 lnet_rsp_tracker_create(void)
3789 {
3790         struct list_head **rstqs;
3791         rstqs = lnet_create_array_of_queues();
3792
3793         if (!rstqs)
3794                 return -ENOMEM;
3795
3796         the_lnet.ln_mt_rstq = rstqs;
3797
3798         return 0;
3799 }
3800
3801 static void
3802 lnet_rsp_tracker_clean(void)
3803 {
3804         lnet_finalize_expired_responses();
3805
3806         cfs_percpt_free(the_lnet.ln_mt_rstq);
3807         the_lnet.ln_mt_rstq = NULL;
3808 }
3809
3810 int lnet_monitor_thr_start(void)
3811 {
3812         int rc = 0;
3813         struct task_struct *task;
3814
3815         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3816                 return -EALREADY;
3817
3818         rc = lnet_resendqs_create();
3819         if (rc)
3820                 return rc;
3821
3822         rc = lnet_rsp_tracker_create();
3823         if (rc)
3824                 goto clean_queues;
3825
3826         sema_init(&the_lnet.ln_mt_signal, 0);
3827
3828         lnet_net_lock(LNET_LOCK_EX);
3829         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3830         lnet_net_unlock(LNET_LOCK_EX);
3831         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3832         if (IS_ERR(task)) {
3833                 rc = PTR_ERR(task);
3834                 CERROR("Can't start monitor thread: %d\n", rc);
3835                 goto clean_thread;
3836         }
3837
3838         return 0;
3839
3840 clean_thread:
3841         lnet_net_lock(LNET_LOCK_EX);
3842         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3843         lnet_net_unlock(LNET_LOCK_EX);
3844         /* block until event callback signals exit */
3845         down(&the_lnet.ln_mt_signal);
3846         /* clean up */
3847         lnet_net_lock(LNET_LOCK_EX);
3848         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3849         lnet_net_unlock(LNET_LOCK_EX);
3850         lnet_rsp_tracker_clean();
3851         lnet_clean_local_ni_recoveryq();
3852         lnet_clean_peer_ni_recoveryq();
3853         lnet_clean_resendqs();
3854         the_lnet.ln_mt_handler = NULL;
3855         return rc;
3856 clean_queues:
3857         lnet_rsp_tracker_clean();
3858         lnet_clean_local_ni_recoveryq();
3859         lnet_clean_peer_ni_recoveryq();
3860         lnet_clean_resendqs();
3861         return rc;
3862 }
3863
3864 void lnet_monitor_thr_stop(void)
3865 {
3866         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3867                 return;
3868
3869         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3870         lnet_net_lock(LNET_LOCK_EX);
3871         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3872         lnet_net_unlock(LNET_LOCK_EX);
3873
3874         /* tell the monitor thread that we're shutting down */
3875         complete(&the_lnet.ln_mt_wait_complete);
3876
3877         /* block until monitor thread signals that it's done */
3878         down(&the_lnet.ln_mt_signal);
3879         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3880
3881         /* perform cleanup tasks */
3882         lnet_rsp_tracker_clean();
3883         lnet_clean_local_ni_recoveryq();
3884         lnet_clean_peer_ni_recoveryq();
3885         lnet_clean_resendqs();
3886 }
3887
3888 void
3889 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3890                   __u32 msg_type)
3891 {
3892         lnet_net_lock(cpt);
3893         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3894         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3895         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3896         lnet_net_unlock(cpt);
3897
3898         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3899 }
3900
3901 static void
3902 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3903 {
3904         struct lnet_hdr *hdr = &msg->msg_hdr;
3905
3906         if (msg->msg_wanted != 0)
3907                 lnet_setpayloadbuffer(msg);
3908
3909         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3910
3911         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3912          * it back into the ACK during lnet_finalize() */
3913         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3914                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3915
3916         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3917                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3918 }
3919
3920 static int
3921 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3922 {
3923         struct lnet_hdr         *hdr = &msg->msg_hdr;
3924         struct lnet_match_info  info;
3925         int                     rc;
3926         bool                    ready_delay;
3927
3928         /* Convert put fields to host byte order */
3929         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3930         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3931         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3932
3933         /* Primary peer NID. */
3934         info.mi_id.nid  = msg->msg_initiator;
3935         info.mi_id.pid  = hdr->src_pid;
3936         info.mi_opc     = LNET_MD_OP_PUT;
3937         info.mi_portal  = hdr->msg.put.ptl_index;
3938         info.mi_rlength = hdr->payload_length;
3939         info.mi_roffset = hdr->msg.put.offset;
3940         info.mi_mbits   = hdr->msg.put.match_bits;
3941         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3942
3943         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3944         ready_delay = msg->msg_rx_ready_delay;
3945
3946  again:
3947         rc = lnet_ptl_match_md(&info, msg);
3948         switch (rc) {
3949         default:
3950                 LBUG();
3951
3952         case LNET_MATCHMD_OK:
3953                 lnet_recv_put(ni, msg);
3954                 return 0;
3955
3956         case LNET_MATCHMD_NONE:
3957                 if (ready_delay)
3958                         /* no eager_recv or has already called it, should
3959                          * have been attached on delayed list */
3960                         return 0;
3961
3962                 rc = lnet_ni_eager_recv(ni, msg);
3963                 if (rc == 0) {
3964                         ready_delay = true;
3965                         goto again;
3966                 }
3967                 /* fall through */
3968
3969         case LNET_MATCHMD_DROP:
3970                 CNETERR("Dropping PUT from %s portal %d match %llu"
3971                         " offset %d length %d: %d\n",
3972                         libcfs_id2str(info.mi_id), info.mi_portal,
3973                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3974
3975                 return -ENOENT; /* -ve: OK but no match */
3976         }
3977 }
3978
3979 static int
3980 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3981 {
3982         struct lnet_match_info info;
3983         struct lnet_hdr *hdr = &msg->msg_hdr;
3984         struct lnet_process_id source_id;
3985         struct lnet_handle_wire reply_wmd;
3986         int rc;
3987
3988         /* Convert get fields to host byte order */
3989         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3990         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3991         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3992         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3993
3994         source_id.nid = hdr->src_nid;
3995         source_id.pid = hdr->src_pid;
3996         /* Primary peer NID */
3997         info.mi_id.nid  = msg->msg_initiator;
3998         info.mi_id.pid  = hdr->src_pid;
3999         info.mi_opc     = LNET_MD_OP_GET;
4000         info.mi_portal  = hdr->msg.get.ptl_index;
4001         info.mi_rlength = hdr->msg.get.sink_length;
4002         info.mi_roffset = hdr->msg.get.src_offset;
4003         info.mi_mbits   = hdr->msg.get.match_bits;
4004         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
4005
4006         rc = lnet_ptl_match_md(&info, msg);
4007         if (rc == LNET_MATCHMD_DROP) {
4008                 CNETERR("Dropping GET from %s portal %d match %llu"
4009                         " offset %d length %d\n",
4010                         libcfs_id2str(info.mi_id), info.mi_portal,
4011                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
4012                 return -ENOENT; /* -ve: OK but no match */
4013         }
4014
4015         LASSERT(rc == LNET_MATCHMD_OK);
4016
4017         lnet_build_msg_event(msg, LNET_EVENT_GET);
4018
4019         reply_wmd = hdr->msg.get.return_wmd;
4020
4021         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4022                        msg->msg_offset, msg->msg_wanted);
4023
4024         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4025
4026         if (rdma_get) {
4027                 /* The LND completes the REPLY from her recv procedure */
4028                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4029                              msg->msg_offset, msg->msg_len, msg->msg_len);
4030                 return 0;
4031         }
4032
4033         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4034         msg->msg_receiving = 0;
4035
4036         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4037         if (rc < 0) {
4038                 /* didn't get as far as lnet_ni_send() */
4039                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4040                        libcfs_nid2str(ni->ni_nid),
4041                        libcfs_id2str(info.mi_id), rc);
4042
4043                 lnet_finalize(msg, rc);
4044         }
4045
4046         return 0;
4047 }
4048
4049 static int
4050 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4051 {
4052         void *private = msg->msg_private;
4053         struct lnet_hdr *hdr = &msg->msg_hdr;
4054         struct lnet_process_id src = {0};
4055         struct lnet_libmd *md;
4056         unsigned int rlength;
4057         unsigned int mlength;
4058         int cpt;
4059
4060         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4061         lnet_res_lock(cpt);
4062
4063         src.nid = hdr->src_nid;
4064         src.pid = hdr->src_pid;
4065
4066         /* NB handles only looked up by creator (no flips) */
4067         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4068         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4069                 CNETERR("%s: Dropping REPLY from %s for %s "
4070                         "MD %#llx.%#llx\n",
4071                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4072                         (md == NULL) ? "invalid" : "inactive",
4073                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4074                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4075                 if (md != NULL && md->md_me != NULL)
4076                         CERROR("REPLY MD also attached to portal %d\n",
4077                                md->md_me->me_portal);
4078
4079                 lnet_res_unlock(cpt);
4080                 return -ENOENT; /* -ve: OK but no match */
4081         }
4082
4083         LASSERT(md->md_offset == 0);
4084
4085         rlength = hdr->payload_length;
4086         mlength = min(rlength, md->md_length);
4087
4088         if (mlength < rlength &&
4089             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4090                 CNETERR("%s: Dropping REPLY from %s length %d "
4091                         "for MD %#llx would overflow (%d)\n",
4092                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4093                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4094                         mlength);
4095                 lnet_res_unlock(cpt);
4096                 return -ENOENT; /* -ve: OK but no match */
4097         }
4098
4099         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4100                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4101                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4102
4103         lnet_msg_attach_md(msg, md, 0, mlength);
4104
4105         if (mlength != 0)
4106                 lnet_setpayloadbuffer(msg);
4107
4108         lnet_res_unlock(cpt);
4109
4110         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4111
4112         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4113         return 0;
4114 }
4115
4116 static int
4117 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4118 {
4119         struct lnet_hdr *hdr = &msg->msg_hdr;
4120         struct lnet_process_id src = {0};
4121         struct lnet_libmd *md;
4122         int cpt;
4123
4124         src.nid = hdr->src_nid;
4125         src.pid = hdr->src_pid;
4126
4127         /* Convert ack fields to host byte order */
4128         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4129         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4130
4131         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4132         lnet_res_lock(cpt);
4133
4134         /* NB handles only looked up by creator (no flips) */
4135         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4136         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4137                 /* Don't moan; this is expected */
4138                 CDEBUG(D_NET,
4139                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4140                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4141                        (md == NULL) ? "invalid" : "inactive",
4142                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4143                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4144                 if (md != NULL && md->md_me != NULL)
4145                         CERROR("Source MD also attached to portal %d\n",
4146                                md->md_me->me_portal);
4147
4148                 lnet_res_unlock(cpt);
4149                 return -ENOENT;                  /* -ve! */
4150         }
4151
4152         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4153                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4154                hdr->msg.ack.dst_wmd.wh_object_cookie);
4155
4156         lnet_msg_attach_md(msg, md, 0, 0);
4157
4158         lnet_res_unlock(cpt);
4159
4160         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4161
4162         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4163         return 0;
4164 }
4165
4166 /**
4167  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4168  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4169  * \retval -ve                  error code
4170  */
4171 int
4172 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4173 {
4174         int     rc = 0;
4175
4176         if (!the_lnet.ln_routing)
4177                 return -ECANCELED;
4178
4179         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4180             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4181                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4182                         msg->msg_rx_ready_delay = 1;
4183                 } else {
4184                         lnet_net_unlock(msg->msg_rx_cpt);
4185                         rc = lnet_ni_eager_recv(ni, msg);
4186                         lnet_net_lock(msg->msg_rx_cpt);
4187                 }
4188         }
4189
4190         if (rc == 0)
4191                 rc = lnet_post_routed_recv_locked(msg, 0);
4192         return rc;
4193 }
4194
4195 int
4196 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4197 {
4198         int     rc;
4199
4200         switch (msg->msg_type) {
4201         case LNET_MSG_ACK:
4202                 rc = lnet_parse_ack(ni, msg);
4203                 break;
4204         case LNET_MSG_PUT:
4205                 rc = lnet_parse_put(ni, msg);
4206                 break;
4207         case LNET_MSG_GET:
4208                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4209                 break;
4210         case LNET_MSG_REPLY:
4211                 rc = lnet_parse_reply(ni, msg);
4212                 break;
4213         default: /* prevent an unused label if !kernel */
4214                 LASSERT(0);
4215                 return -EPROTO;
4216         }
4217
4218         LASSERT(rc == 0 || rc == -ENOENT);
4219         return rc;
4220 }
4221
4222 char *
4223 lnet_msgtyp2str (int type)
4224 {
4225         switch (type) {
4226         case LNET_MSG_ACK:
4227                 return ("ACK");
4228         case LNET_MSG_PUT:
4229                 return ("PUT");
4230         case LNET_MSG_GET:
4231                 return ("GET");
4232         case LNET_MSG_REPLY:
4233                 return ("REPLY");
4234         case LNET_MSG_HELLO:
4235                 return ("HELLO");
4236         default:
4237                 return ("<UNKNOWN>");
4238         }
4239 }
4240
4241 int
4242 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4243            void *private, int rdma_req)
4244 {
4245         struct lnet_peer_ni *lpni;
4246         struct lnet_msg *msg;
4247         __u32 payload_length;
4248         lnet_pid_t dest_pid;
4249         lnet_nid_t dest_nid;
4250         lnet_nid_t src_nid;
4251         bool push = false;
4252         int for_me;
4253         __u32 type;
4254         int rc = 0;
4255         int cpt;
4256
4257         LASSERT (!in_interrupt ());
4258
4259         type = le32_to_cpu(hdr->type);
4260         src_nid = le64_to_cpu(hdr->src_nid);
4261         dest_nid = le64_to_cpu(hdr->dest_nid);
4262         dest_pid = le32_to_cpu(hdr->dest_pid);
4263         payload_length = le32_to_cpu(hdr->payload_length);
4264
4265         for_me = (ni->ni_nid == dest_nid);
4266         cpt = lnet_cpt_of_nid(from_nid, ni);
4267
4268         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4269                 libcfs_nid2str(dest_nid),
4270                 libcfs_nid2str(ni->ni_nid),
4271                 libcfs_nid2str(src_nid),
4272                 lnet_msgtyp2str(type),
4273                 (for_me) ? "for me" : "routed");
4274
4275         switch (type) {
4276         case LNET_MSG_ACK:
4277         case LNET_MSG_GET:
4278                 if (payload_length > 0) {
4279                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4280                                libcfs_nid2str(from_nid),
4281                                libcfs_nid2str(src_nid),
4282                                lnet_msgtyp2str(type), payload_length);
4283                         return -EPROTO;
4284                 }
4285                 break;
4286
4287         case LNET_MSG_PUT:
4288         case LNET_MSG_REPLY:
4289                 if (payload_length >
4290                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4291                         CERROR("%s, src %s: bad %s payload %d "
4292                                "(%d max expected)\n",
4293                                libcfs_nid2str(from_nid),
4294                                libcfs_nid2str(src_nid),
4295                                lnet_msgtyp2str(type),
4296                                payload_length,
4297                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4298                         return -EPROTO;
4299                 }
4300                 break;
4301
4302         default:
4303                 CERROR("%s, src %s: Bad message type 0x%x\n",
4304                        libcfs_nid2str(from_nid),
4305                        libcfs_nid2str(src_nid), type);
4306                 return -EPROTO;
4307         }
4308
4309         if (the_lnet.ln_routing &&
4310             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4311                 lnet_ni_lock(ni);
4312                 spin_lock(&ni->ni_net->net_lock);
4313                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4314                 spin_unlock(&ni->ni_net->net_lock);
4315                 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4316                 lnet_ni_unlock(ni);
4317         }
4318
4319         if (push)
4320                 lnet_push_update_to_peers(1);
4321
4322         /* Regard a bad destination NID as a protocol error.  Senders should
4323          * know what they're doing; if they don't they're misconfigured, buggy
4324          * or malicious so we chop them off at the knees :) */
4325
4326         if (!for_me) {
4327                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4328                         /* should have gone direct */
4329                         CERROR("%s, src %s: Bad dest nid %s "
4330                                "(should have been sent direct)\n",
4331                                 libcfs_nid2str(from_nid),
4332                                 libcfs_nid2str(src_nid),
4333                                 libcfs_nid2str(dest_nid));
4334                         return -EPROTO;
4335                 }
4336
4337                 if (lnet_islocalnid(dest_nid)) {
4338                         /* dest is another local NI; sender should have used
4339                          * this node's NID on its own network */
4340                         CERROR("%s, src %s: Bad dest nid %s "
4341                                "(it's my nid but on a different network)\n",
4342                                 libcfs_nid2str(from_nid),
4343                                 libcfs_nid2str(src_nid),
4344                                 libcfs_nid2str(dest_nid));
4345                         return -EPROTO;
4346                 }
4347
4348                 if (rdma_req && type == LNET_MSG_GET) {
4349                         CERROR("%s, src %s: Bad optimized GET for %s "
4350                                "(final destination must be me)\n",
4351                                 libcfs_nid2str(from_nid),
4352                                 libcfs_nid2str(src_nid),
4353                                 libcfs_nid2str(dest_nid));
4354                         return -EPROTO;
4355                 }
4356
4357                 if (!the_lnet.ln_routing) {
4358                         CERROR("%s, src %s: Dropping message for %s "
4359                                "(routing not enabled)\n",
4360                                 libcfs_nid2str(from_nid),
4361                                 libcfs_nid2str(src_nid),
4362                                 libcfs_nid2str(dest_nid));
4363                         goto drop;
4364                 }
4365         }
4366
4367         /* Message looks OK; we're not going to return an error, so we MUST
4368          * call back lnd_recv() come what may... */
4369
4370         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4371             fail_peer(src_nid, 0)) {                    /* shall we now? */
4372                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4373                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4374                        lnet_msgtyp2str(type));
4375                 goto drop;
4376         }
4377
4378         if (!list_empty(&the_lnet.ln_drop_rules) &&
4379             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4380                 CDEBUG(D_NET,
4381                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4382                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4383                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4384                 goto drop;
4385         }
4386
4387         if (lnet_drop_asym_route && for_me &&
4388             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4389                 struct lnet_net *net;
4390                 struct lnet_remotenet *rnet;
4391                 bool found = true;
4392
4393                 /* we are dealing with a routed message,
4394                  * so see if route to reach src_nid goes through from_nid
4395                  */
4396                 lnet_net_lock(cpt);
4397                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4398                 if (!net) {
4399                         lnet_net_unlock(cpt);
4400                         CERROR("net %s not found\n",
4401                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4402                         return -EPROTO;
4403                 }
4404
4405                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4406                 if (rnet) {
4407                         struct lnet_peer *gw = NULL;
4408                         struct lnet_peer_ni *lpni = NULL;
4409                         struct lnet_route *route;
4410
4411                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4412                                 found = false;
4413                                 gw = route->lr_gateway;
4414                                 if (route->lr_lnet != net->net_id)
4415                                         continue;
4416                                 /*
4417                                  * if the nid is one of the gateway's NIDs
4418                                  * then this is a valid gateway
4419                                  */
4420                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4421                                                 NULL, lpni)) != NULL) {
4422                                         if (lpni->lpni_nid == from_nid) {
4423                                                 found = true;
4424                                                 break;
4425                                         }
4426                                 }
4427                         }
4428                 }
4429                 lnet_net_unlock(cpt);
4430                 if (!found) {
4431                         /* we would not use from_nid to route a message to
4432                          * src_nid
4433                          * => asymmetric routing detected but forbidden
4434                          */
4435                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4436                                libcfs_nid2str(from_nid),
4437                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4438                         goto drop;
4439                 }
4440         }
4441
4442         msg = lnet_msg_alloc();
4443         if (msg == NULL) {
4444                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4445                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4446                        lnet_msgtyp2str(type));
4447                 goto drop;
4448         }
4449
4450         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4451          * pointers NULL etc */
4452
4453         msg->msg_type = type;
4454         msg->msg_private = private;
4455         msg->msg_receiving = 1;
4456         msg->msg_rdma_get = rdma_req;
4457         msg->msg_len = msg->msg_wanted = payload_length;
4458         msg->msg_offset = 0;
4459         msg->msg_hdr = *hdr;
4460         /* for building message event */
4461         msg->msg_from = from_nid;
4462         if (!for_me) {
4463                 msg->msg_target.pid     = dest_pid;
4464                 msg->msg_target.nid     = dest_nid;
4465                 msg->msg_routing        = 1;
4466
4467         } else {
4468                 /* convert common msg->hdr fields to host byteorder */
4469                 msg->msg_hdr.type       = type;
4470                 msg->msg_hdr.src_nid    = src_nid;
4471                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4472                 msg->msg_hdr.dest_nid   = dest_nid;
4473                 msg->msg_hdr.dest_pid   = dest_pid;
4474                 msg->msg_hdr.payload_length = payload_length;
4475         }
4476
4477         lnet_net_lock(cpt);
4478         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4479         if (IS_ERR(lpni)) {
4480                 lnet_net_unlock(cpt);
4481                 CERROR("%s, src %s: Dropping %s "
4482                        "(error %ld looking up sender)\n",
4483                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4484                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4485                 lnet_msg_free(msg);
4486                 if (rc == -ESHUTDOWN)
4487                         /* We are shutting down.  Don't do anything more */
4488                         return 0;
4489                 goto drop;
4490         }
4491
4492         if (the_lnet.ln_routing)
4493                 lpni->lpni_last_alive = ktime_get_seconds();
4494
4495         msg->msg_rxpeer = lpni;
4496         msg->msg_rxni = ni;
4497         lnet_ni_addref_locked(ni, cpt);
4498         /* Multi-Rail: Primary NID of source. */
4499         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4500
4501         /*
4502          * mark the status of this lpni as UP since we received a message
4503          * from it. The ping response reports back the ns_status which is
4504          * marked on the remote as up or down and we cache it here.
4505          */
4506         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4507
4508         lnet_msg_commit(msg, cpt);
4509
4510         /* message delay simulation */
4511         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4512                      lnet_delay_rule_match_locked(hdr, msg))) {
4513                 lnet_net_unlock(cpt);
4514                 return 0;
4515         }
4516
4517         if (!for_me) {
4518                 rc = lnet_parse_forward_locked(ni, msg);
4519                 lnet_net_unlock(cpt);
4520
4521                 if (rc < 0)
4522                         goto free_drop;
4523
4524                 if (rc == LNET_CREDIT_OK) {
4525                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4526                                      0, payload_length, payload_length);
4527                 }
4528                 return 0;
4529         }
4530
4531         lnet_net_unlock(cpt);
4532
4533         rc = lnet_parse_local(ni, msg);
4534         if (rc != 0)
4535                 goto free_drop;
4536         return 0;
4537
4538  free_drop:
4539         LASSERT(msg->msg_md == NULL);
4540         lnet_finalize(msg, rc);
4541
4542  drop:
4543         lnet_drop_message(ni, cpt, private, payload_length, type);
4544         return 0;
4545 }
4546 EXPORT_SYMBOL(lnet_parse);
4547
4548 void
4549 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4550 {
4551         while (!list_empty(head)) {
4552                 struct lnet_process_id id = {0};
4553                 struct lnet_msg *msg;
4554
4555                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4556                 list_del(&msg->msg_list);
4557
4558                 id.nid = msg->msg_hdr.src_nid;
4559                 id.pid = msg->msg_hdr.src_pid;
4560
4561                 LASSERT(msg->msg_md == NULL);
4562                 LASSERT(msg->msg_rx_delayed);
4563                 LASSERT(msg->msg_rxpeer != NULL);
4564                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4565
4566                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4567                       " offset %d length %d: %s\n",
4568                       libcfs_id2str(id),
4569                       msg->msg_hdr.msg.put.ptl_index,
4570                       msg->msg_hdr.msg.put.match_bits,
4571                       msg->msg_hdr.msg.put.offset,
4572                       msg->msg_hdr.payload_length, reason);
4573
4574                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4575                  * called lnet_drop_message(), so I just hang onto msg as well
4576                  * until that's done */
4577
4578                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4579                                   msg->msg_private, msg->msg_len,
4580                                   msg->msg_type);
4581
4582                 msg->msg_no_resend = true;
4583                 /*
4584                  * NB: message will not generate event because w/o attached MD,
4585                  * but we still should give error code so lnet_msg_decommit()
4586                  * can skip counters operations and other checks.
4587                  */
4588                 lnet_finalize(msg, -ENOENT);
4589         }
4590 }
4591
4592 void
4593 lnet_recv_delayed_msg_list(struct list_head *head)
4594 {
4595         while (!list_empty(head)) {
4596                 struct lnet_msg *msg;
4597                 struct lnet_process_id id;
4598
4599                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4600                 list_del(&msg->msg_list);
4601
4602                 /* md won't disappear under me, since each msg
4603                  * holds a ref on it */
4604
4605                 id.nid = msg->msg_hdr.src_nid;
4606                 id.pid = msg->msg_hdr.src_pid;
4607
4608                 LASSERT(msg->msg_rx_delayed);
4609                 LASSERT(msg->msg_md != NULL);
4610                 LASSERT(msg->msg_rxpeer != NULL);
4611                 LASSERT(msg->msg_rxni != NULL);
4612                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4613
4614                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4615                        "match %llu offset %d length %d.\n",
4616                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4617                         msg->msg_hdr.msg.put.match_bits,
4618                         msg->msg_hdr.msg.put.offset,
4619                         msg->msg_hdr.payload_length);
4620
4621                 lnet_recv_put(msg->msg_rxni, msg);
4622         }
4623 }
4624
4625 static void
4626 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4627                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4628 {
4629         s64 timeout_ns;
4630         struct lnet_rsp_tracker *local_rspt;
4631
4632         /*
4633          * MD has a refcount taken by message so it's not going away.
4634          * The MD however can be looked up. We need to secure the access
4635          * to the md_rspt_ptr by taking the res_lock.
4636          * The rspt can be accessed without protection up to when it gets
4637          * added to the list.
4638          */
4639
4640         lnet_res_lock(cpt);
4641         local_rspt = md->md_rspt_ptr;
4642         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4643         if (local_rspt != NULL) {
4644                 /*
4645                  * we already have an rspt attached to the md, so we'll
4646                  * update the deadline on that one.
4647                  */
4648                 lnet_rspt_free(rspt, cpt);
4649         } else {
4650                 /* new md */
4651                 rspt->rspt_mdh = mdh;
4652                 rspt->rspt_cpt = cpt;
4653                 /* store the rspt so we can access it when we get the REPLY */
4654                 md->md_rspt_ptr = rspt;
4655                 local_rspt = rspt;
4656         }
4657         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4658
4659         /*
4660          * add to the list of tracked responses. It's added to tail of the
4661          * list in order to expire all the older entries first.
4662          */
4663         lnet_net_lock(cpt);
4664         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4665         lnet_net_unlock(cpt);
4666         lnet_res_unlock(cpt);
4667 }
4668
4669 /**
4670  * Initiate an asynchronous PUT operation.
4671  *
4672  * There are several events associated with a PUT: completion of the send on
4673  * the initiator node (LNET_EVENT_SEND), and when the send completes
4674  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4675  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4676  * used at the target node to indicate the completion of incoming data
4677  * delivery.
4678  *
4679  * The local events will be logged in the EQ associated with the MD pointed to
4680  * by \a mdh handle. Using a MD without an associated EQ results in these
4681  * events being discarded. In this case, the caller must have another
4682  * mechanism (e.g., a higher level protocol) for determining when it is safe
4683  * to modify the memory region associated with the MD.
4684  *
4685  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4686  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4687  *
4688  * \param self Indicates the NID of a local interface through which to send
4689  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4690  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4691  * must be "free floating" (See LNetMDBind()).
4692  * \param ack Controls whether an acknowledgment is requested.
4693  * Acknowledgments are only sent when they are requested by the initiating
4694  * process and the target MD enables them.
4695  * \param target A process identifier for the target process.
4696  * \param portal The index in the \a target's portal table.
4697  * \param match_bits The match bits to use for MD selection at the target
4698  * process.
4699  * \param offset The offset into the target MD (only used when the target
4700  * MD has the LNET_MD_MANAGE_REMOTE option set).
4701  * \param hdr_data 64 bits of user data that can be included in the message
4702  * header. This data is written to an event queue entry at the target if an
4703  * EQ is present on the matching MD.
4704  *
4705  * \retval  0      Success, and only in this case events will be generated
4706  * and logged to EQ (if it exists).
4707  * \retval -EIO    Simulated failure.
4708  * \retval -ENOMEM Memory allocation failure.
4709  * \retval -ENOENT Invalid MD object.
4710  *
4711  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4712  */
4713 int
4714 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4715         struct lnet_process_id target, unsigned int portal,
4716         __u64 match_bits, unsigned int offset,
4717         __u64 hdr_data)
4718 {
4719         struct lnet_msg *msg;
4720         struct lnet_libmd *md;
4721         int cpt;
4722         int rc;
4723         struct lnet_rsp_tracker *rspt = NULL;
4724
4725         LASSERT(the_lnet.ln_refcount > 0);
4726
4727         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4728             fail_peer(target.nid, 1)) {                 /* shall we now? */
4729                 CERROR("Dropping PUT to %s: simulated failure\n",
4730                        libcfs_id2str(target));
4731                 return -EIO;
4732         }
4733
4734         msg = lnet_msg_alloc();
4735         if (msg == NULL) {
4736                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4737                        libcfs_id2str(target));
4738                 return -ENOMEM;
4739         }
4740         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4741
4742         cpt = lnet_cpt_of_cookie(mdh.cookie);
4743
4744         if (ack == LNET_ACK_REQ) {
4745                 rspt = lnet_rspt_alloc(cpt);
4746                 if (!rspt) {
4747                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4748                                 libcfs_id2str(target));
4749                         return -ENOMEM;
4750                 }
4751                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4752         }
4753
4754         lnet_res_lock(cpt);
4755
4756         md = lnet_handle2md(&mdh);
4757         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4758                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4759                        match_bits, portal, libcfs_id2str(target),
4760                        md == NULL ? -1 : md->md_threshold);
4761                 if (md != NULL && md->md_me != NULL)
4762                         CERROR("Source MD also attached to portal %d\n",
4763                                md->md_me->me_portal);
4764                 lnet_res_unlock(cpt);
4765
4766                 if (rspt)
4767                         lnet_rspt_free(rspt, cpt);
4768
4769                 lnet_msg_free(msg);
4770                 return -ENOENT;
4771         }
4772
4773         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4774
4775         lnet_msg_attach_md(msg, md, 0, 0);
4776
4777         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4778
4779         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4780         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4781         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4782         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4783
4784         /* NB handles only looked up by creator (no flips) */
4785         if (ack == LNET_ACK_REQ) {
4786                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4787                         the_lnet.ln_interface_cookie;
4788                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4789                         md->md_lh.lh_cookie;
4790         } else {
4791                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4792                         LNET_WIRE_HANDLE_COOKIE_NONE;
4793                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4794                         LNET_WIRE_HANDLE_COOKIE_NONE;
4795         }
4796
4797         lnet_res_unlock(cpt);
4798
4799         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4800
4801         if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
4802                                                    md->md_options))
4803                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4804         else if (rspt)
4805                 lnet_rspt_free(rspt, cpt);
4806
4807         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4808                                  CFS_FAIL_ONCE))
4809                 rc = -EIO;
4810         else
4811                 rc = lnet_send(self, msg, LNET_NID_ANY);
4812
4813         if (rc != 0) {
4814                 CNETERR("Error sending PUT to %s: %d\n",
4815                         libcfs_id2str(target), rc);
4816                 msg->msg_no_resend = true;
4817                 lnet_finalize(msg, rc);
4818         }
4819
4820         /* completion will be signalled by an event */
4821         return 0;
4822 }
4823 EXPORT_SYMBOL(LNetPut);
4824
4825 /*
4826  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4827  * returns a msg for the LND to pass to lnet_finalize() when the sink
4828  * data has been received.
4829  *
4830  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4831  * lnet_finalize() is called on it, so the LND must call this first
4832  */
4833 struct lnet_msg *
4834 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4835 {
4836         struct lnet_msg *msg = lnet_msg_alloc();
4837         struct lnet_libmd *getmd = getmsg->msg_md;
4838         struct lnet_process_id peer_id = getmsg->msg_target;
4839         int cpt;
4840
4841         LASSERT(!getmsg->msg_target_is_router);
4842         LASSERT(!getmsg->msg_routing);
4843
4844         if (msg == NULL) {
4845                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4846                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4847                 goto drop;
4848         }
4849
4850         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4851         lnet_res_lock(cpt);
4852
4853         LASSERT(getmd->md_refcount > 0);
4854
4855         if (getmd->md_threshold == 0) {
4856                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4857                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4858                         getmd);
4859                 lnet_res_unlock(cpt);
4860                 goto drop;
4861         }
4862
4863         LASSERT(getmd->md_offset == 0);
4864
4865         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4866                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4867
4868         /* setup information for lnet_build_msg_event */
4869         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4870         msg->msg_from = peer_id.nid;
4871         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4872         msg->msg_hdr.src_nid = peer_id.nid;
4873         msg->msg_hdr.payload_length = getmd->md_length;
4874         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4875
4876         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4877         lnet_res_unlock(cpt);
4878
4879         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4880
4881         lnet_net_lock(cpt);
4882         lnet_msg_commit(msg, cpt);
4883         lnet_net_unlock(cpt);
4884
4885         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4886
4887         return msg;
4888
4889  drop:
4890         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4891
4892         lnet_net_lock(cpt);
4893         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4894         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4895         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4896                 getmd->md_length;
4897         lnet_net_unlock(cpt);
4898
4899         if (msg != NULL)
4900                 lnet_msg_free(msg);
4901
4902         return NULL;
4903 }
4904 EXPORT_SYMBOL(lnet_create_reply_msg);
4905
4906 void
4907 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4908                        unsigned int len)
4909 {
4910         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4911          * completed and I know it. */
4912         LASSERT(reply != NULL);
4913         LASSERT(reply->msg_type == LNET_MSG_GET);
4914         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4915
4916         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4917          * the end of my buffer, I might as well be dead. */
4918         LASSERT(len <= reply->msg_ev.mlength);
4919
4920         reply->msg_ev.mlength = len;
4921 }
4922 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4923
4924 /**
4925  * Initiate an asynchronous GET operation.
4926  *
4927  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4928  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4929  * the target node in the REPLY has been written to local MD.
4930  *
4931  * On the target node, an LNET_EVENT_GET is logged when the GET request
4932  * arrives and is accepted into a MD.
4933  *
4934  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4935  * \param mdh A handle for the MD that describes the memory into which the
4936  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4937  *
4938  * \retval  0      Success, and only in this case events will be generated
4939  * and logged to EQ (if it exists) of the MD.
4940  * \retval -EIO    Simulated failure.
4941  * \retval -ENOMEM Memory allocation failure.
4942  * \retval -ENOENT Invalid MD object.
4943  */
4944 int
4945 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4946         struct lnet_process_id target, unsigned int portal,
4947         __u64 match_bits, unsigned int offset, bool recovery)
4948 {
4949         struct lnet_msg *msg;
4950         struct lnet_libmd *md;
4951         struct lnet_rsp_tracker *rspt;
4952         int cpt;
4953         int rc;
4954
4955         LASSERT(the_lnet.ln_refcount > 0);
4956
4957         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4958             fail_peer(target.nid, 1))                   /* shall we now? */
4959         {
4960                 CERROR("Dropping GET to %s: simulated failure\n",
4961                        libcfs_id2str(target));
4962                 return -EIO;
4963         }
4964
4965         msg = lnet_msg_alloc();
4966         if (!msg) {
4967                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4968                        libcfs_id2str(target));
4969                 return -ENOMEM;
4970         }
4971
4972         cpt = lnet_cpt_of_cookie(mdh.cookie);
4973
4974         rspt = lnet_rspt_alloc(cpt);
4975         if (!rspt) {
4976                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4977                        libcfs_id2str(target));
4978                 return -ENOMEM;
4979         }
4980         INIT_LIST_HEAD(&rspt->rspt_on_list);
4981
4982         msg->msg_recovery = recovery;
4983
4984         lnet_res_lock(cpt);
4985
4986         md = lnet_handle2md(&mdh);
4987         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4988                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4989                        match_bits, portal, libcfs_id2str(target),
4990                        md == NULL ? -1 : md->md_threshold);
4991                 if (md != NULL && md->md_me != NULL)
4992                         CERROR("REPLY MD also attached to portal %d\n",
4993                                md->md_me->me_portal);
4994
4995                 lnet_res_unlock(cpt);
4996
4997                 lnet_msg_free(msg);
4998                 lnet_rspt_free(rspt, cpt);
4999                 return -ENOENT;
5000         }
5001
5002         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5003
5004         lnet_msg_attach_md(msg, md, 0, 0);
5005
5006         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5007
5008         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5009         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5010         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5011         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5012
5013         /* NB handles only looked up by creator (no flips) */
5014         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5015                 the_lnet.ln_interface_cookie;
5016         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5017                 md->md_lh.lh_cookie;
5018
5019         lnet_res_unlock(cpt);
5020
5021         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5022
5023         if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5024                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5025         else
5026                 lnet_rspt_free(rspt, cpt);
5027
5028         rc = lnet_send(self, msg, LNET_NID_ANY);
5029         if (rc < 0) {
5030                 CNETERR("Error sending GET to %s: %d\n",
5031                         libcfs_id2str(target), rc);
5032                 msg->msg_no_resend = true;
5033                 lnet_finalize(msg, rc);
5034         }
5035
5036         /* completion will be signalled by an event */
5037         return 0;
5038 }
5039 EXPORT_SYMBOL(LNetGet);
5040
5041 /**
5042  * Calculate distance to node at \a dstnid.
5043  *
5044  * \param dstnid Target NID.
5045  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5046  * is saved here.
5047  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5048  * here.
5049  *
5050  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5051  * local_nid_dist_zero is set, which is the default.
5052  * \retval positives Distance to target NID, i.e. number of hops plus one.
5053  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5054  */
5055 int
5056 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5057 {
5058         struct list_head *e;
5059         struct lnet_ni *ni = NULL;
5060         struct lnet_remotenet *rnet;
5061         __u32 dstnet = LNET_NIDNET(dstnid);
5062         int hops;
5063         int cpt;
5064         __u32 order = 2;
5065         struct list_head *rn_list;
5066
5067         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5068          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5069          * keep order 0 free for 0@lo and order 1 free for a local NID
5070          * match */
5071
5072         LASSERT(the_lnet.ln_refcount > 0);
5073
5074         cpt = lnet_net_lock_current();
5075
5076         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5077                 if (ni->ni_nid == dstnid) {
5078                         if (srcnidp != NULL)
5079                                 *srcnidp = dstnid;
5080                         if (orderp != NULL) {
5081                                 if (dstnid == LNET_NID_LO_0)
5082                                         *orderp = 0;
5083                                 else
5084                                         *orderp = 1;
5085                         }
5086                         lnet_net_unlock(cpt);
5087
5088                         return local_nid_dist_zero ? 0 : 1;
5089                 }
5090
5091                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5092                         /* Check if ni was originally created in
5093                          * current net namespace.
5094                          * If not, assign order above 0xffff0000,
5095                          * to make this ni not a priority. */
5096                         if (current->nsproxy &&
5097                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5098                                         order += 0xffff0000;
5099                         if (srcnidp != NULL)
5100                                 *srcnidp = ni->ni_nid;
5101                         if (orderp != NULL)
5102                                 *orderp = order;
5103                         lnet_net_unlock(cpt);
5104                         return 1;
5105                 }
5106
5107                 order++;
5108         }
5109
5110         rn_list = lnet_net2rnethash(dstnet);
5111         list_for_each(e, rn_list) {
5112                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5113
5114                 if (rnet->lrn_net == dstnet) {
5115                         struct lnet_route *route;
5116                         struct lnet_route *shortest = NULL;
5117                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5118                         __u32 route_hops;
5119
5120                         LASSERT(!list_empty(&rnet->lrn_routes));
5121
5122                         list_for_each_entry(route, &rnet->lrn_routes,
5123                                             lr_list) {
5124                                 route_hops = route->lr_hops;
5125                                 if (route_hops == LNET_UNDEFINED_HOPS)
5126                                         route_hops = 1;
5127                                 if (shortest == NULL ||
5128                                     route_hops < shortest_hops) {
5129                                         shortest = route;
5130                                         shortest_hops = route_hops;
5131                                 }
5132                         }
5133
5134                         LASSERT(shortest != NULL);
5135                         hops = shortest_hops;
5136                         if (srcnidp != NULL) {
5137                                 struct lnet_net *net;
5138                                 net = lnet_get_net_locked(shortest->lr_lnet);
5139                                 LASSERT(net);
5140                                 ni = lnet_get_next_ni_locked(net, NULL);
5141                                 *srcnidp = ni->ni_nid;
5142                         }
5143                         if (orderp != NULL)
5144                                 *orderp = order;
5145                         lnet_net_unlock(cpt);
5146                         return hops + 1;
5147                 }
5148                 order++;
5149         }
5150
5151         lnet_net_unlock(cpt);
5152         return -EHOSTUNREACH;
5153 }
5154 EXPORT_SYMBOL(LNetDist);