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  *
  31  * lnet/lnet/lib-move.c
  32  *
  33  * Data movement routines
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <linux/pagemap.h>
  39
  40 #include <lnet/lib-lnet.h>
  41 #include <linux/nsproxy.h>
  42 #include <lnet/lnet_rdma.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         struct lnet_nid sd_dst_nid;
  59         struct lnet_nid sd_src_nid;
  60         struct lnet_nid 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 nid4, unsigned int threshold)
 194 {
 195         struct lnet_test_peer *tp;
 196         struct list_head *el;
 197         struct list_head *next;
 198         struct lnet_nid nid;
 199         LIST_HEAD(cull);
 200
 201         lnet_nid4_to_nid(nid4, &nid);
 202         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 203         if (threshold != 0) {
 204                 /* Adding a new entry */
 205                 LIBCFS_ALLOC(tp, sizeof(*tp));
 206                 if (tp == NULL)
 207                         return -ENOMEM;
 208
 209                 tp->tp_nid = nid;
 210                 tp->tp_threshold = threshold;
 211
 212                 lnet_net_lock(0);
 213                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 214                 lnet_net_unlock(0);
 215                 return 0;
 216         }
 217
 218         lnet_net_lock(0);
 219
 220         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 221                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 222
 223                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 224                     LNET_NID_IS_ANY(&nid) ||    /* removing all entries */
 225                     nid_same(&tp->tp_nid, &nid)) {      /* matched this one */
 226                         list_move(&tp->tp_list, &cull);
 227                 }
 228         }
 229
 230         lnet_net_unlock(0);
 231
 232         while ((tp = list_first_entry_or_null(&cull,
 233                                               struct lnet_test_peer,
 234                                               tp_list)) != NULL) {
 235                 list_del(&tp->tp_list);
 236                 LIBCFS_FREE(tp, sizeof(*tp));
 237         }
 238         return 0;
 239 }
 240
 241 static int
 242 fail_peer(struct lnet_nid *nid, int outgoing)
 243 {
 244         struct lnet_test_peer *tp;
 245         struct list_head *el;
 246         struct list_head *next;
 247         LIST_HEAD(cull);
 248         int fail = 0;
 249
 250         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 251         lnet_net_lock(0);
 252
 253         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 254                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 255
 256                 if (tp->tp_threshold == 0) {
 257                         /* zombie entry */
 258                         if (outgoing) {
 259                                 /* only cull zombies on outgoing tests,
 260                                  * since we may be at interrupt priority on
 261                                  * incoming messages. */
 262                                 list_move(&tp->tp_list, &cull);
 263                         }
 264                         continue;
 265                 }
 266
 267                 if (LNET_NID_IS_ANY(&tp->tp_nid) ||     /* fail every peer */
 268                     nid_same(nid, &tp->tp_nid)) {       /* fail this peer */
 269                         fail = 1;
 270
 271                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 272                                 tp->tp_threshold--;
 273                                 if (outgoing &&
 274                                     tp->tp_threshold == 0) {
 275                                         /* see above */
 276                                         list_move(&tp->tp_list, &cull);
 277                                 }
 278                         }
 279                         break;
 280                 }
 281         }
 282
 283         lnet_net_unlock(0);
 284
 285         while ((tp = list_first_entry_or_null(&cull,
 286                                               struct lnet_test_peer,
 287                                               tp_list)) != NULL) {
 288                 list_del(&tp->tp_list);
 289                 LIBCFS_FREE(tp, sizeof(*tp));
 290         }
 291
 292         return fail;
 293 }
 294
 295 unsigned int
 296 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 297 {
 298         unsigned int nob = 0;
 299
 300         LASSERT(niov == 0 || iov != NULL);
 301         while (niov-- > 0)
 302                 nob += (iov++)->iov_len;
 303
 304         return (nob);
 305 }
 306 EXPORT_SYMBOL(lnet_iov_nob);
 307
 308 void
 309 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 310                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 311                   unsigned int nob)
 312 {
 313         /* NB diov, siov are READ-ONLY */
 314         unsigned int this_nob;
 315
 316         if (nob == 0)
 317                 return;
 318
 319         /* skip complete frags before 'doffset' */
 320         LASSERT(ndiov > 0);
 321         while (doffset >= diov->iov_len) {
 322                 doffset -= diov->iov_len;
 323                 diov++;
 324                 ndiov--;
 325                 LASSERT(ndiov > 0);
 326         }
 327
 328         /* skip complete frags before 'soffset' */
 329         LASSERT(nsiov > 0);
 330         while (soffset >= siov->iov_len) {
 331                 soffset -= siov->iov_len;
 332                 siov++;
 333                 nsiov--;
 334                 LASSERT(nsiov > 0);
 335         }
 336
 337         do {
 338                 LASSERT(ndiov > 0);
 339                 LASSERT(nsiov > 0);
 340                 this_nob = min3((unsigned int)diov->iov_len - doffset,
 341                                 (unsigned int)siov->iov_len - soffset,
 342                                 nob);
 343
 344                 memcpy((char *)diov->iov_base + doffset,
 345                        (char *)siov->iov_base + soffset, this_nob);
 346                 nob -= this_nob;
 347
 348                 if (diov->iov_len > doffset + this_nob) {
 349                         doffset += this_nob;
 350                 } else {
 351                         diov++;
 352                         ndiov--;
 353                         doffset = 0;
 354                 }
 355
 356                 if (siov->iov_len > soffset + this_nob) {
 357                         soffset += this_nob;
 358                 } else {
 359                         siov++;
 360                         nsiov--;
 361                         soffset = 0;
 362                 }
 363         } while (nob > 0);
 364 }
 365 EXPORT_SYMBOL(lnet_copy_iov2iov);
 366
 367 unsigned int
 368 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 369 {
 370         unsigned int  nob = 0;
 371
 372         LASSERT(niov == 0 || kiov != NULL);
 373         while (niov-- > 0)
 374                 nob += (kiov++)->bv_len;
 375
 376         return (nob);
 377 }
 378 EXPORT_SYMBOL(lnet_kiov_nob);
 379
 380 void
 381 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
 382                     unsigned int doffset,
 383                     unsigned int nsiov, struct bio_vec *siov,
 384                     unsigned int soffset,
 385                     unsigned int nob)
 386 {
 387         /* NB diov, siov are READ-ONLY */
 388         unsigned int    this_nob;
 389         char           *daddr = NULL;
 390         char           *saddr = NULL;
 391
 392         if (nob == 0)
 393                 return;
 394
 395         LASSERT (!in_interrupt ());
 396
 397         LASSERT (ndiov > 0);
 398         while (doffset >= diov->bv_len) {
 399                 doffset -= diov->bv_len;
 400                 diov++;
 401                 ndiov--;
 402                 LASSERT(ndiov > 0);
 403         }
 404
 405         LASSERT(nsiov > 0);
 406         while (soffset >= siov->bv_len) {
 407                 soffset -= siov->bv_len;
 408                 siov++;
 409                 nsiov--;
 410                 LASSERT(nsiov > 0);
 411         }
 412
 413         do {
 414                 LASSERT(ndiov > 0);
 415                 LASSERT(nsiov > 0);
 416                 this_nob = min3(diov->bv_len - doffset,
 417                                 siov->bv_len - soffset,
 418                                 nob);
 419
 420                 if (daddr == NULL)
 421                         daddr = ((char *)kmap(diov->bv_page)) +
 422                                 diov->bv_offset + doffset;
 423                 if (saddr == NULL)
 424                         saddr = ((char *)kmap(siov->bv_page)) +
 425                                 siov->bv_offset + soffset;
 426
 427                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 428                  * However in practice at least one of the kiovs will be mapped
 429                  * kernel pages and the map/unmap will be NOOPs */
 430
 431                 memcpy (daddr, saddr, this_nob);
 432                 nob -= this_nob;
 433
 434                 if (diov->bv_len > doffset + this_nob) {
 435                         daddr += this_nob;
 436                         doffset += this_nob;
 437                 } else {
 438                         kunmap(diov->bv_page);
 439                         daddr = NULL;
 440                         diov++;
 441                         ndiov--;
 442                         doffset = 0;
 443                 }
 444
 445                 if (siov->bv_len > soffset + this_nob) {
 446                         saddr += this_nob;
 447                         soffset += this_nob;
 448                 } else {
 449                         kunmap(siov->bv_page);
 450                         saddr = NULL;
 451                         siov++;
 452                         nsiov--;
 453                         soffset = 0;
 454                 }
 455         } while (nob > 0);
 456
 457         if (daddr != NULL)
 458                 kunmap(diov->bv_page);
 459         if (saddr != NULL)
 460                 kunmap(siov->bv_page);
 461 }
 462 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 463
 464 void
 465 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 466                     unsigned int nkiov, struct bio_vec *kiov,
 467                     unsigned int kiovoffset,
 468                     unsigned int nob)
 469 {
 470         /* NB iov, kiov are READ-ONLY */
 471         unsigned int    this_nob;
 472         char           *addr = NULL;
 473
 474         if (nob == 0)
 475                 return;
 476
 477         LASSERT (!in_interrupt ());
 478
 479         LASSERT (niov > 0);
 480         while (iovoffset >= iov->iov_len) {
 481                 iovoffset -= iov->iov_len;
 482                 iov++;
 483                 niov--;
 484                 LASSERT(niov > 0);
 485         }
 486
 487         LASSERT(nkiov > 0);
 488         while (kiovoffset >= kiov->bv_len) {
 489                 kiovoffset -= kiov->bv_len;
 490                 kiov++;
 491                 nkiov--;
 492                 LASSERT(nkiov > 0);
 493         }
 494
 495         do {
 496                 LASSERT(niov > 0);
 497                 LASSERT(nkiov > 0);
 498                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 499                                 (unsigned int)kiov->bv_len - kiovoffset,
 500                                 nob);
 501
 502                 if (addr == NULL)
 503                         addr = ((char *)kmap(kiov->bv_page)) +
 504                                 kiov->bv_offset + kiovoffset;
 505
 506                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 507                 nob -= this_nob;
 508
 509                 if (iov->iov_len > iovoffset + this_nob) {
 510                         iovoffset += this_nob;
 511                 } else {
 512                         iov++;
 513                         niov--;
 514                         iovoffset = 0;
 515                 }
 516
 517                 if (kiov->bv_len > kiovoffset + this_nob) {
 518                         addr += this_nob;
 519                         kiovoffset += this_nob;
 520                 } else {
 521                         kunmap(kiov->bv_page);
 522                         addr = NULL;
 523                         kiov++;
 524                         nkiov--;
 525                         kiovoffset = 0;
 526                 }
 527
 528         } while (nob > 0);
 529
 530         if (addr != NULL)
 531                 kunmap(kiov->bv_page);
 532 }
 533 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 534
 535 void
 536 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
 537                    unsigned int kiovoffset,
 538                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 539                    unsigned int nob)
 540 {
 541         /* NB kiov, iov are READ-ONLY */
 542         unsigned int    this_nob;
 543         char           *addr = NULL;
 544
 545         if (nob == 0)
 546                 return;
 547
 548         LASSERT (!in_interrupt ());
 549
 550         LASSERT (nkiov > 0);
 551         while (kiovoffset >= kiov->bv_len) {
 552                 kiovoffset -= kiov->bv_len;
 553                 kiov++;
 554                 nkiov--;
 555                 LASSERT(nkiov > 0);
 556         }
 557
 558         LASSERT(niov > 0);
 559         while (iovoffset >= iov->iov_len) {
 560                 iovoffset -= iov->iov_len;
 561                 iov++;
 562                 niov--;
 563                 LASSERT(niov > 0);
 564         }
 565
 566         do {
 567                 LASSERT(nkiov > 0);
 568                 LASSERT(niov > 0);
 569                 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
 570                                 (unsigned int)iov->iov_len - iovoffset,
 571                                 nob);
 572
 573                 if (addr == NULL)
 574                         addr = ((char *)kmap(kiov->bv_page)) +
 575                                 kiov->bv_offset + kiovoffset;
 576
 577                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 578                 nob -= this_nob;
 579
 580                 if (kiov->bv_len > kiovoffset + this_nob) {
 581                         addr += this_nob;
 582                         kiovoffset += this_nob;
 583                 } else {
 584                         kunmap(kiov->bv_page);
 585                         addr = NULL;
 586                         kiov++;
 587                         nkiov--;
 588                         kiovoffset = 0;
 589                 }
 590
 591                 if (iov->iov_len > iovoffset + this_nob) {
 592                         iovoffset += this_nob;
 593                 } else {
 594                         iov++;
 595                         niov--;
 596                         iovoffset = 0;
 597                 }
 598         } while (nob > 0);
 599
 600         if (addr != NULL)
 601                 kunmap(kiov->bv_page);
 602 }
 603 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 604
 605 int
 606 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
 607                   int src_niov, struct bio_vec *src,
 608                   unsigned int offset, unsigned int len)
 609 {
 610         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 611          * for exactly 'len' bytes, and return the number of entries.
 612          * NB not destructive to 'src' */
 613         unsigned int    frag_len;
 614         unsigned int    niov;
 615
 616         if (len == 0)                           /* no data => */
 617                 return (0);                     /* no frags */
 618
 619         LASSERT(src_niov > 0);
 620         while (offset >= src->bv_len) {      /* skip initial frags */
 621                 offset -= src->bv_len;
 622                 src_niov--;
 623                 src++;
 624                 LASSERT(src_niov > 0);
 625         }
 626
 627         niov = 1;
 628         for (;;) {
 629                 LASSERT(src_niov > 0);
 630                 LASSERT((int)niov <= dst_niov);
 631
 632                 frag_len = src->bv_len - offset;
 633                 dst->bv_page = src->bv_page;
 634                 dst->bv_offset = src->bv_offset + offset;
 635
 636                 if (len <= frag_len) {
 637                         dst->bv_len = len;
 638                         LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 639                         return niov;
 640                 }
 641
 642                 dst->bv_len = frag_len;
 643                 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 644
 645                 len -= frag_len;
 646                 dst++;
 647                 src++;
 648                 niov++;
 649                 src_niov--;
 650                 offset = 0;
 651         }
 652 }
 653 EXPORT_SYMBOL(lnet_extract_kiov);
 654
 655 void
 656 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 657              int delayed, unsigned int offset, unsigned int mlen,
 658              unsigned int rlen)
 659 {
 660         unsigned int niov = 0;
 661         struct kvec *iov = NULL;
 662         struct bio_vec  *kiov = NULL;
 663         int rc;
 664
 665         LASSERT (!in_interrupt ());
 666         LASSERT (mlen == 0 || msg != NULL);
 667
 668         if (msg != NULL) {
 669                 LASSERT(msg->msg_receiving);
 670                 LASSERT(!msg->msg_sending);
 671                 LASSERT(rlen == msg->msg_len);
 672                 LASSERT(mlen <= msg->msg_len);
 673                 LASSERT(msg->msg_offset == offset);
 674                 LASSERT(msg->msg_wanted == mlen);
 675
 676                 msg->msg_receiving = 0;
 677
 678                 if (mlen != 0) {
 679                         niov = msg->msg_niov;
 680                         kiov = msg->msg_kiov;
 681
 682                         LASSERT (niov > 0);
 683                         LASSERT ((iov == NULL) != (kiov == NULL));
 684                 }
 685         }
 686
 687         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 688                                              niov, kiov, offset, mlen,
 689                                              rlen);
 690         if (rc < 0)
 691                 lnet_finalize(msg, rc);
 692 }
 693
 694 static void
 695 lnet_setpayloadbuffer(struct lnet_msg *msg)
 696 {
 697         struct lnet_libmd *md = msg->msg_md;
 698
 699         LASSERT(msg->msg_len > 0);
 700         LASSERT(!msg->msg_routing);
 701         LASSERT(md != NULL);
 702         LASSERT(msg->msg_niov == 0);
 703         LASSERT(msg->msg_kiov == NULL);
 704
 705         msg->msg_niov = md->md_niov;
 706         msg->msg_kiov = md->md_kiov;
 707 }
 708
 709 void
 710 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_processid *target,
 711                unsigned int offset, unsigned int len)
 712 {
 713         msg->msg_type = type;
 714         msg->msg_target = *target;
 715         msg->msg_len = len;
 716         msg->msg_offset = offset;
 717
 718         if (len != 0)
 719                 lnet_setpayloadbuffer(msg);
 720
 721         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 722         msg->msg_hdr.type           = type;
 723         /* dest_nid will be overwritten by lnet_select_pathway() */
 724         msg->msg_hdr.dest_nid = target->nid;
 725         msg->msg_hdr.dest_pid = target->pid;
 726         /* src_nid will be set later */
 727         msg->msg_hdr.src_pid        = the_lnet.ln_pid;
 728         msg->msg_hdr.payload_length = len;
 729 }
 730
 731 void
 732 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 733 {
 734         void *priv = msg->msg_private;
 735         int rc;
 736
 737         LASSERT(!in_interrupt());
 738         LASSERT(nid_is_lo0(&ni->ni_nid) ||
 739                 (msg->msg_txcredit && msg->msg_peertxcredit));
 740
 741         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 742         if (rc < 0) {
 743                 msg->msg_no_resend = true;
 744                 lnet_finalize(msg, rc);
 745         }
 746 }
 747
 748 static int
 749 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 750 {
 751         int     rc;
 752
 753         LASSERT(!msg->msg_sending);
 754         LASSERT(msg->msg_receiving);
 755         LASSERT(!msg->msg_rx_ready_delay);
 756         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 757
 758         msg->msg_rx_ready_delay = 1;
 759         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 760                                                   &msg->msg_private);
 761         if (rc != 0) {
 762                 CERROR("recv from %s / send to %s aborted: "
 763                        "eager_recv failed %d\n",
 764                        libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
 765                        libcfs_idstr(&msg->msg_target), rc);
 766                 LASSERT(rc < 0); /* required by my callers */
 767         }
 768
 769         return rc;
 770 }
 771
 772 /* Returns:
 773  *  -ETIMEDOUT if the message deadline has been exceeded
 774  *  -EHOSTUNREACH if the peer is down
 775  *  0 if this message should not be dropped
 776  */
 777 static int
 778 lnet_check_message_drop(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 779                         struct lnet_msg *msg)
 780 {
 781         /* Drop message if we've exceeded the message deadline */
 782         if (ktime_after(ktime_get(), msg->msg_deadline))
 783                 return -ETIMEDOUT;
 784
 785         if (msg->msg_target.pid & LNET_PID_USERFLAG)
 786                 return 0;
 787
 788         if (!lnet_peer_aliveness_enabled(lpni))
 789                 return 0;
 790
 791         /* If we're resending a message, let's attempt to send it even if
 792          * the peer is down to fulfill our resend quota on the message
 793          */
 794         if (msg->msg_retry_count > 0)
 795                 return 0;
 796
 797         /* try and send recovery messages regardless */
 798         if (msg->msg_recovery)
 799                 return 0;
 800
 801         /* always send any responses */
 802         if (lnet_msg_is_response(msg))
 803                 return 0;
 804
 805         /* always send non-routed messages */
 806         if (!msg->msg_routing)
 807                 return 0;
 808
 809         /* assume peer_ni is alive as long as we're within the configured
 810          * peer timeout
 811          */
 812         if (ktime_get_seconds() >=
 813             (lpni->lpni_last_alive +
 814              lpni->lpni_net->net_tunables.lct_peer_timeout))
 815                 return -EHOSTUNREACH;
 816
 817         return 0;
 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         int rc;
 839
 840         /* non-lnet_send() callers have checked before */
 841         LASSERT(!do_send || msg->msg_tx_delayed);
 842         LASSERT(!msg->msg_receiving);
 843         LASSERT(msg->msg_tx_committed);
 844
 845         /* can't get here if we're sending to the loopback interface */
 846         if (the_lnet.ln_loni)
 847                 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
 848
 849         /* NB 'lp' is always the next hop */
 850         rc = lnet_check_message_drop(ni, lp, msg);
 851         if (rc) {
 852                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 853                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 854                         msg->msg_len;
 855                 lnet_net_unlock(cpt);
 856                 if (msg->msg_txpeer)
 857                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 858                                         msg->msg_type,
 859                                         LNET_STATS_TYPE_DROP);
 860                 if (msg->msg_txni)
 861                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 862                                         msg->msg_type,
 863                                         LNET_STATS_TYPE_DROP);
 864
 865                 if (rc == -EHOSTUNREACH) {
 866                         CNETERR("Dropping message for %s: peer not alive\n",
 867                                 libcfs_idstr(&msg->msg_target));
 868                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 869                 } else {
 870                         CNETERR("Dropping message for %s: exceeded message deadline\n",
 871                                 libcfs_idstr(&msg->msg_target));
 872                         msg->msg_health_status =
 873                                 LNET_MSG_STATUS_NETWORK_TIMEOUT;
 874                 }
 875
 876                 if (do_send)
 877                         lnet_finalize(msg, rc);
 878
 879                 lnet_net_lock(cpt);
 880                 return rc;
 881         }
 882
 883         if (msg->msg_md != NULL &&
 884             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 885                 lnet_net_unlock(cpt);
 886
 887                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 888                         "called on the MD/ME.\n",
 889                         libcfs_idstr(&msg->msg_target));
 890                 if (do_send) {
 891                         msg->msg_no_resend = true;
 892                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 893                                msg, libcfs_idstr(&msg->msg_target));
 894                         lnet_finalize(msg, -ECANCELED);
 895                 }
 896
 897                 lnet_net_lock(cpt);
 898                 return -ECANCELED;
 899         }
 900
 901         if (!msg->msg_peertxcredit) {
 902                 spin_lock(&lp->lpni_lock);
 903                 LASSERT((lp->lpni_txcredits < 0) ==
 904                         !list_empty(&lp->lpni_txq));
 905
 906                 msg->msg_peertxcredit = 1;
 907                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr_nid4);
 908                 lp->lpni_txcredits--;
 909
 910                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 911                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 912
 913                 if (lp->lpni_txcredits < 0) {
 914                         msg->msg_tx_delayed = 1;
 915                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 916                         spin_unlock(&lp->lpni_lock);
 917                         return LNET_CREDIT_WAIT;
 918                 }
 919                 spin_unlock(&lp->lpni_lock);
 920         }
 921
 922         if (!msg->msg_txcredit) {
 923                 LASSERT((tq->tq_credits < 0) ==
 924                         !list_empty(&tq->tq_delayed));
 925
 926                 msg->msg_txcredit = 1;
 927                 tq->tq_credits--;
 928                 atomic_dec(&ni->ni_tx_credits);
 929
 930                 if (tq->tq_credits < tq->tq_credits_min)
 931                         tq->tq_credits_min = tq->tq_credits;
 932
 933                 if (tq->tq_credits < 0) {
 934                         msg->msg_tx_delayed = 1;
 935                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 936                         return LNET_CREDIT_WAIT;
 937                 }
 938         }
 939
 940         if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
 941             lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
 942                 msg->msg_tx_delayed = 1;
 943                 return LNET_CREDIT_WAIT;
 944         }
 945
 946         /* unset the tx_delay flag as we're going to send it now */
 947         msg->msg_tx_delayed = 0;
 948
 949         if (do_send) {
 950                 lnet_net_unlock(cpt);
 951                 lnet_ni_send(ni, msg);
 952                 lnet_net_lock(cpt);
 953         }
 954         return LNET_CREDIT_OK;
 955 }
 956
 957
 958 static struct lnet_rtrbufpool *
 959 lnet_msg2bufpool(struct lnet_msg *msg)
 960 {
 961         struct lnet_rtrbufpool  *rbp;
 962         int                     cpt;
 963
 964         LASSERT(msg->msg_rx_committed);
 965
 966         cpt = msg->msg_rx_cpt;
 967         rbp = &the_lnet.ln_rtrpools[cpt][0];
 968
 969         LASSERT(msg->msg_len <= LNET_MTU);
 970         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 971                 rbp++;
 972                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 973         }
 974
 975         return rbp;
 976 }
 977
 978 static int
 979 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 980 {
 981         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 982          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 983          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 984          * received or OK to receive */
 985         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 986         struct lnet_peer *lp;
 987         struct lnet_rtrbufpool *rbp;
 988         struct lnet_rtrbuf *rb;
 989
 990         LASSERT(msg->msg_kiov == NULL);
 991         LASSERT(msg->msg_niov == 0);
 992         LASSERT(msg->msg_routing);
 993         LASSERT(msg->msg_receiving);
 994         LASSERT(!msg->msg_sending);
 995         LASSERT(lpni->lpni_peer_net);
 996         LASSERT(lpni->lpni_peer_net->lpn_peer);
 997
 998         lp = lpni->lpni_peer_net->lpn_peer;
 999
1000         /* non-lnet_parse callers only receive delayed messages */
1001         LASSERT(!do_recv || msg->msg_rx_delayed);
1002
1003         if (!msg->msg_peerrtrcredit) {
1004                 /* lpni_lock protects the credit manipulation */
1005                 spin_lock(&lpni->lpni_lock);
1006
1007                 msg->msg_peerrtrcredit = 1;
1008                 lpni->lpni_rtrcredits--;
1009                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1010                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1011
1012                 if (lpni->lpni_rtrcredits < 0) {
1013                         spin_unlock(&lpni->lpni_lock);
1014                         /* must have checked eager_recv before here */
1015                         LASSERT(msg->msg_rx_ready_delay);
1016                         msg->msg_rx_delayed = 1;
1017                         /* lp_lock protects the lp_rtrq */
1018                         spin_lock(&lp->lp_lock);
1019                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1020                         spin_unlock(&lp->lp_lock);
1021                         return LNET_CREDIT_WAIT;
1022                 }
1023                 spin_unlock(&lpni->lpni_lock);
1024         }
1025
1026         rbp = lnet_msg2bufpool(msg);
1027
1028         if (!msg->msg_rtrcredit) {
1029                 msg->msg_rtrcredit = 1;
1030                 rbp->rbp_credits--;
1031                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1032                         rbp->rbp_mincredits = rbp->rbp_credits;
1033
1034                 if (rbp->rbp_credits < 0) {
1035                         /* must have checked eager_recv before here */
1036                         LASSERT(msg->msg_rx_ready_delay);
1037                         msg->msg_rx_delayed = 1;
1038                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1039                         return LNET_CREDIT_WAIT;
1040                 }
1041         }
1042
1043         LASSERT(!list_empty(&rbp->rbp_bufs));
1044         rb = list_first_entry(&rbp->rbp_bufs, struct lnet_rtrbuf, rb_list);
1045         list_del(&rb->rb_list);
1046
1047         msg->msg_niov = rbp->rbp_npages;
1048         msg->msg_kiov = &rb->rb_kiov[0];
1049
1050         /* unset the msg-rx_delayed flag since we're receiving the message */
1051         msg->msg_rx_delayed = 0;
1052
1053         if (do_recv) {
1054                 int cpt = msg->msg_rx_cpt;
1055
1056                 lnet_net_unlock(cpt);
1057                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1058                              0, msg->msg_len, msg->msg_len);
1059                 lnet_net_lock(cpt);
1060         }
1061         return LNET_CREDIT_OK;
1062 }
1063
1064 void
1065 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1066 {
1067         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1068         struct lnet_ni          *txni = msg->msg_txni;
1069         struct lnet_msg         *msg2;
1070
1071         if (msg->msg_txcredit) {
1072                 struct lnet_ni       *ni = msg->msg_txni;
1073                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1074
1075                 /* give back NI txcredits */
1076                 msg->msg_txcredit = 0;
1077
1078                 LASSERT((tq->tq_credits < 0) ==
1079                         !list_empty(&tq->tq_delayed));
1080
1081                 tq->tq_credits++;
1082                 atomic_inc(&ni->ni_tx_credits);
1083                 if (tq->tq_credits <= 0) {
1084                         msg2 = list_first_entry(&tq->tq_delayed,
1085                                                 struct lnet_msg, msg_list);
1086                         list_del(&msg2->msg_list);
1087
1088                         LASSERT(msg2->msg_txni == ni);
1089                         LASSERT(msg2->msg_tx_delayed);
1090                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1091
1092                         (void) lnet_post_send_locked(msg2, 1);
1093                 }
1094         }
1095
1096         if (msg->msg_peertxcredit) {
1097                 /* give back peer txcredits */
1098                 msg->msg_peertxcredit = 0;
1099
1100                 spin_lock(&txpeer->lpni_lock);
1101                 LASSERT((txpeer->lpni_txcredits < 0) ==
1102                         !list_empty(&txpeer->lpni_txq));
1103
1104                 txpeer->lpni_txqnob -=  msg->msg_len +
1105                                         sizeof(struct lnet_hdr_nid4);
1106                 LASSERT(txpeer->lpni_txqnob >= 0);
1107
1108                 txpeer->lpni_txcredits++;
1109                 if (txpeer->lpni_txcredits <= 0) {
1110                         int msg2_cpt;
1111
1112                         msg2 = list_first_entry(&txpeer->lpni_txq,
1113                                                 struct lnet_msg, msg_list);
1114                         list_del(&msg2->msg_list);
1115                         spin_unlock(&txpeer->lpni_lock);
1116
1117                         LASSERT(msg2->msg_txpeer == txpeer);
1118                         LASSERT(msg2->msg_tx_delayed);
1119
1120                         msg2_cpt = msg2->msg_tx_cpt;
1121
1122                         /*
1123                          * The msg_cpt can be different from the msg2_cpt
1124                          * so we need to make sure we lock the correct cpt
1125                          * for msg2.
1126                          * Once we call lnet_post_send_locked() it is no
1127                          * longer safe to access msg2, since it could've
1128                          * been freed by lnet_finalize(), but we still
1129                          * need to relock the correct cpt, so we cache the
1130                          * msg2_cpt for the purpose of the check that
1131                          * follows the call to lnet_pose_send_locked().
1132                          */
1133                         if (msg2_cpt != msg->msg_tx_cpt) {
1134                                 lnet_net_unlock(msg->msg_tx_cpt);
1135                                 lnet_net_lock(msg2_cpt);
1136                         }
1137                         (void) lnet_post_send_locked(msg2, 1);
1138                         if (msg2_cpt != msg->msg_tx_cpt) {
1139                                 lnet_net_unlock(msg2_cpt);
1140                                 lnet_net_lock(msg->msg_tx_cpt);
1141                         }
1142                 } else {
1143                         spin_unlock(&txpeer->lpni_lock);
1144                 }
1145         }
1146
1147         if (txni != NULL) {
1148                 msg->msg_txni = NULL;
1149                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1150         }
1151
1152         if (txpeer != NULL) {
1153                 msg->msg_txpeer = NULL;
1154                 lnet_peer_ni_decref_locked(txpeer);
1155         }
1156 }
1157
1158 void
1159 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1160 {
1161         struct lnet_msg *msg;
1162
1163         if (list_empty(&rbp->rbp_msgs))
1164                 return;
1165         msg = list_first_entry(&rbp->rbp_msgs,
1166                                struct lnet_msg, msg_list);
1167         list_del(&msg->msg_list);
1168
1169         (void)lnet_post_routed_recv_locked(msg, 1);
1170 }
1171
1172 void
1173 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1174 {
1175         struct lnet_msg *msg;
1176         struct lnet_msg *tmp;
1177
1178         lnet_net_unlock(cpt);
1179
1180         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1181                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1182                              0, 0, 0, msg->msg_hdr.payload_length);
1183                 list_del_init(&msg->msg_list);
1184                 msg->msg_no_resend = true;
1185                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1186                 lnet_finalize(msg, -ECANCELED);
1187         }
1188
1189         lnet_net_lock(cpt);
1190 }
1191
1192 void
1193 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1194 {
1195         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1196         struct lnet_peer *lp;
1197         struct lnet_ni *rxni = msg->msg_rxni;
1198         struct lnet_msg *msg2;
1199
1200         if (msg->msg_rtrcredit) {
1201                 /* give back global router credits */
1202                 struct lnet_rtrbuf *rb;
1203                 struct lnet_rtrbufpool *rbp;
1204
1205                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1206                  * there until it gets one allocated, or aborts the wait
1207                  * itself */
1208                 LASSERT(msg->msg_kiov != NULL);
1209
1210                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1211                 rbp = rb->rb_pool;
1212
1213                 msg->msg_kiov = NULL;
1214                 msg->msg_rtrcredit = 0;
1215
1216                 LASSERT(rbp == lnet_msg2bufpool(msg));
1217
1218                 LASSERT((rbp->rbp_credits > 0) ==
1219                         !list_empty(&rbp->rbp_bufs));
1220
1221                 /* If routing is now turned off, we just drop this buffer and
1222                  * don't bother trying to return credits.  */
1223                 if (!the_lnet.ln_routing) {
1224                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1225                         goto routing_off;
1226                 }
1227
1228                 /* It is possible that a user has lowered the desired number of
1229                  * buffers in this pool.  Make sure we never put back
1230                  * more buffers than the stated number. */
1231                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1232                         /* Discard this buffer so we don't have too
1233                          * many. */
1234                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1235                         rbp->rbp_nbuffers--;
1236                 } else {
1237                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1238                         rbp->rbp_credits++;
1239                         if (rbp->rbp_credits <= 0)
1240                                 lnet_schedule_blocked_locked(rbp);
1241                 }
1242         }
1243
1244 routing_off:
1245         if (msg->msg_peerrtrcredit) {
1246                 LASSERT(rxpeerni);
1247                 LASSERT(rxpeerni->lpni_peer_net);
1248                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1249
1250                 /* give back peer router credits */
1251                 msg->msg_peerrtrcredit = 0;
1252
1253                 spin_lock(&rxpeerni->lpni_lock);
1254                 rxpeerni->lpni_rtrcredits++;
1255                 spin_unlock(&rxpeerni->lpni_lock);
1256
1257                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1258                 spin_lock(&lp->lp_lock);
1259
1260                 /* drop all messages which are queued to be routed on that
1261                  * peer. */
1262                 if (!the_lnet.ln_routing) {
1263                         LIST_HEAD(drop);
1264                         list_splice_init(&lp->lp_rtrq, &drop);
1265                         spin_unlock(&lp->lp_lock);
1266                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1267                 } else if (!list_empty(&lp->lp_rtrq)) {
1268                         int msg2_cpt;
1269
1270                         msg2 = list_first_entry(&lp->lp_rtrq,
1271                                                 struct lnet_msg, msg_list);
1272                         list_del(&msg2->msg_list);
1273                         msg2_cpt = msg2->msg_rx_cpt;
1274                         spin_unlock(&lp->lp_lock);
1275                         /*
1276                          * messages on the lp_rtrq can be from any NID in
1277                          * the peer, which means they might have different
1278                          * cpts. We need to make sure we lock the right
1279                          * one.
1280                          */
1281                         if (msg2_cpt != msg->msg_rx_cpt) {
1282                                 lnet_net_unlock(msg->msg_rx_cpt);
1283                                 lnet_net_lock(msg2_cpt);
1284                         }
1285                         (void) lnet_post_routed_recv_locked(msg2, 1);
1286                         if (msg2_cpt != msg->msg_rx_cpt) {
1287                                 lnet_net_unlock(msg2_cpt);
1288                                 lnet_net_lock(msg->msg_rx_cpt);
1289                         }
1290                 } else {
1291                         spin_unlock(&lp->lp_lock);
1292                 }
1293         }
1294         if (rxni != NULL) {
1295                 msg->msg_rxni = NULL;
1296                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1297         }
1298         if (rxpeerni != NULL) {
1299                 msg->msg_rxpeer = NULL;
1300                 lnet_peer_ni_decref_locked(rxpeerni);
1301         }
1302 }
1303
1304 static struct lnet_peer_ni *
1305 lnet_select_peer_ni(struct lnet_ni *best_ni, struct lnet_nid *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 best_lpni_is_preferred = false;
1324         bool lpni_is_preferred;
1325         int lpni_healthv;
1326         __u32 lpni_sel_prio;
1327         __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1328
1329         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1330                 /*
1331                  * if the best_ni we've chosen aleady has this lpni
1332                  * preferred, then let's use it
1333                  */
1334                 if (best_ni) {
1335                         lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1336                                 lpni, &best_ni->ni_nid);
1337                         CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1338                                libcfs_nidstr(&best_ni->ni_nid),
1339                                lpni_is_preferred);
1340                 } else {
1341                         lpni_is_preferred = false;
1342                 }
1343
1344                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1345                 lpni_sel_prio = lpni->lpni_sel_priority;
1346
1347                 if (best_lpni)
1348                         CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1349                                 libcfs_nidstr(&lpni->lpni_nid),
1350                                 libcfs_nidstr(&best_lpni->lpni_nid),
1351                                 lpni_healthv, best_lpni_healthv,
1352                                 lpni_sel_prio, best_sel_prio,
1353                                 lpni->lpni_txcredits, best_lpni_credits,
1354                                 lpni->lpni_seq, best_lpni->lpni_seq);
1355                 else
1356                         goto select_lpni;
1357
1358                 /* pick the healthiest peer ni */
1359                 if (lpni_healthv < best_lpni_healthv)
1360                         continue;
1361                 else if (lpni_healthv > best_lpni_healthv) {
1362                         if (best_lpni_is_preferred)
1363                                 best_lpni_is_preferred = false;
1364                         goto select_lpni;
1365                 }
1366
1367                 if (lpni_sel_prio > best_sel_prio)
1368                         continue;
1369                 else if (lpni_sel_prio < best_sel_prio) {
1370                         if (best_lpni_is_preferred)
1371                                 best_lpni_is_preferred = false;
1372                         goto select_lpni;
1373                 }
1374
1375                 /* if this is a preferred peer use it */
1376                 if (!best_lpni_is_preferred && lpni_is_preferred) {
1377                         best_lpni_is_preferred = true;
1378                         goto select_lpni;
1379                 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1380                         /* this is not the preferred peer so let's ignore
1381                          * it.
1382                          */
1383                         continue;
1384                 }
1385
1386                 if (lpni->lpni_txcredits < best_lpni_credits)
1387                         /* We already have a peer that has more credits
1388                          * available than this one. No need to consider
1389                          * this peer further.
1390                          */
1391                         continue;
1392                 else if (lpni->lpni_txcredits > best_lpni_credits)
1393                         goto select_lpni;
1394
1395                 /* The best peer found so far and the current peer
1396                  * have the same number of available credits let's
1397                  * make sure to select between them using Round Robin
1398                  */
1399                 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1400                         continue;
1401 select_lpni:
1402                 best_lpni_is_preferred = lpni_is_preferred;
1403                 best_lpni_healthv = lpni_healthv;
1404                 best_sel_prio = lpni_sel_prio;
1405                 best_lpni = lpni;
1406                 best_lpni_credits = lpni->lpni_txcredits;
1407         }
1408
1409         /* if we still can't find a peer ni then we can't reach it */
1410         if (!best_lpni) {
1411                 u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1412                              LNET_NID_NET(dst_nid);
1413                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1414                                 libcfs_net2str(net_id));
1415                 return NULL;
1416         }
1417
1418         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1419                libcfs_nidstr(&best_lpni->lpni_nid));
1420
1421         return best_lpni;
1422 }
1423
1424 /*
1425  * Prerequisite: the best_ni should already be set in the sd
1426  * Find the best lpni.
1427  * If the net id is provided then restrict lpni selection on
1428  * that particular net.
1429  * Otherwise find any reachable lpni. When dealing with an MR
1430  * gateway and it has multiple lpnis which we can use
1431  * we want to select the best one from the list of reachable
1432  * ones.
1433  */
1434 static inline struct lnet_peer_ni *
1435 lnet_find_best_lpni(struct lnet_ni *lni, struct lnet_nid *dst_nid,
1436                     struct lnet_peer *peer, u32 net_id)
1437 {
1438         struct lnet_peer_net *peer_net;
1439
1440         /* find the best_lpni on any local network */
1441         if (net_id == LNET_NET_ANY) {
1442                 struct lnet_peer_ni *best_lpni = NULL;
1443                 struct lnet_peer_net *lpn;
1444                 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1445                         /* no net specified find any reachable peer ni */
1446                         if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1447                                 continue;
1448                         best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1449                                                         best_lpni, lpn);
1450                 }
1451
1452                 return best_lpni;
1453         }
1454         /* restrict on the specified net */
1455         peer_net = lnet_peer_get_net_locked(peer, net_id);
1456         if (peer_net)
1457                 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1458
1459         return NULL;
1460 }
1461
1462 static int
1463 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1464 {
1465         if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1466                 return 1;
1467
1468         if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1469                 return -1;
1470
1471         if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1472                 return 1;
1473
1474         if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1475                 return -1;
1476
1477         return 0;
1478 }
1479
1480 /* Compare route priorities and hop counts */
1481 static int
1482 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1483 {
1484         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1485         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1486
1487         if (r1->lr_priority < r2->lr_priority)
1488                 return 1;
1489
1490         if (r1->lr_priority > r2->lr_priority)
1491                 return -1;
1492
1493         if (r1_hops < r2_hops)
1494                 return 1;
1495
1496         if (r1_hops > r2_hops)
1497                 return -1;
1498
1499         return 0;
1500 }
1501
1502 static struct lnet_route *
1503 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1504                        struct lnet_peer_ni *remote_lpni,
1505                        struct lnet_route **prev_route,
1506                        struct lnet_peer_ni **gwni)
1507 {
1508         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1509         struct lnet_route *best_route;
1510         struct lnet_route *last_route;
1511         struct lnet_route *route;
1512         int rc;
1513         bool best_rte_is_preferred = false;
1514         struct lnet_nid *gw_pnid;
1515
1516         CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1517                libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1518
1519         best_route = last_route = NULL;
1520         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1521                 if (!lnet_is_route_alive(route))
1522                         continue;
1523                 gw_pnid = &route->lr_gateway->lp_primary_nid;
1524
1525                 /* no protection on below fields, but it's harmless */
1526                 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1527                         last_route = route;
1528
1529                 /* if the best route found is in the preferred list then
1530                  * tag it as preferred and use it later on. But if we
1531                  * didn't find any routes which are on the preferred list
1532                  * then just use the best route possible.
1533                  */
1534                 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1535
1536                 if (!best_route || (rc && !best_rte_is_preferred)) {
1537                         /* Restrict the selection of the router NI on the
1538                          * src_net provided. If the src_net is LNET_NID_ANY,
1539                          * then select the best interface available.
1540                          */
1541                         lpni = lnet_find_best_lpni(NULL, NULL,
1542                                                    route->lr_gateway,
1543                                                    src_net);
1544                         if (!lpni) {
1545                                 CDEBUG(D_NET,
1546                                        "Gateway %s does not have a peer NI on net %s\n",
1547                                        libcfs_nidstr(gw_pnid),
1548                                        libcfs_net2str(src_net));
1549                                 continue;
1550                         }
1551                 }
1552
1553                 if (rc && !best_rte_is_preferred) {
1554                         /* This is the first preferred route we found,
1555                          * so it beats any route found previously
1556                          */
1557                         best_route = route;
1558                         if (!last_route)
1559                                 last_route = route;
1560                         best_gw_ni = lpni;
1561                         best_rte_is_preferred = true;
1562                         CDEBUG(D_NET, "preferred gw = %s\n",
1563                                libcfs_nidstr(gw_pnid));
1564                         continue;
1565                 } else if ((!rc) && best_rte_is_preferred)
1566                         /* The best route we found so far is in the preferred
1567                          * list, so it beats any non-preferred route
1568                          */
1569                         continue;
1570
1571                 if (!best_route) {
1572                         best_route = last_route = route;
1573                         best_gw_ni = lpni;
1574                         continue;
1575                 }
1576
1577                 rc = lnet_compare_routes(route, best_route);
1578                 if (rc == -1)
1579                         continue;
1580
1581                 /* Restrict the selection of the router NI on the
1582                  * src_net provided. If the src_net is LNET_NID_ANY,
1583                  * then select the best interface available.
1584                  */
1585                 lpni = lnet_find_best_lpni(NULL, NULL, route->lr_gateway,
1586                                            src_net);
1587                 if (!lpni) {
1588                         CDEBUG(D_NET,
1589                                "Gateway %s does not have a peer NI on net %s\n",
1590                                libcfs_nidstr(gw_pnid),
1591                                libcfs_net2str(src_net));
1592                         continue;
1593                 }
1594
1595                 if (rc == 1) {
1596                         best_route = route;
1597                         best_gw_ni = lpni;
1598                         continue;
1599                 }
1600
1601                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1602                 if (rc == -1)
1603                         continue;
1604
1605                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1606                         best_route = route;
1607                         best_gw_ni = lpni;
1608                         continue;
1609                 }
1610         }
1611
1612         *prev_route = last_route;
1613         *gwni = best_gw_ni;
1614
1615         return best_route;
1616 }
1617
1618 static inline unsigned int
1619 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1620 {
1621         if (dev_idx == UINT_MAX)
1622                 return UINT_MAX;
1623
1624         if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1625             !ni->ni_net->net_lnd->lnd_get_dev_prio)
1626                 return UINT_MAX;
1627
1628         return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1629 }
1630
1631 static struct lnet_ni *
1632 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1633                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1634                  struct lnet_msg *msg, int md_cpt)
1635 {
1636         struct lnet_libmd *md = msg->msg_md;
1637         unsigned int offset = msg->msg_offset;
1638         unsigned int shortest_distance;
1639         struct lnet_ni *ni = NULL;
1640         int best_credits;
1641         int best_healthv;
1642         __u32 best_sel_prio;
1643         unsigned int best_dev_prio;
1644         int best_ni_fatal;
1645         unsigned int dev_idx = UINT_MAX;
1646         bool gpu = md ? (md->md_flags & LNET_MD_FLAG_GPU) : false;
1647
1648         if (gpu) {
1649                 struct page *page = lnet_get_first_page(md, offset);
1650
1651                 dev_idx = lnet_get_dev_idx(page);
1652         }
1653
1654         /*
1655          * If there is no peer_ni that we can send to on this network,
1656          * then there is no point in looking for a new best_ni here.
1657         */
1658         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1659                 return best_ni;
1660
1661         if (best_ni == NULL) {
1662                 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1663                 shortest_distance = UINT_MAX;
1664                 best_dev_prio = UINT_MAX;
1665                 best_credits = INT_MIN;
1666                 best_healthv = 0;
1667                 best_ni_fatal = true;
1668         } else {
1669                 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1670                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1671                                                      best_ni->ni_dev_cpt);
1672                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1673                 best_healthv = atomic_read(&best_ni->ni_healthv);
1674                 best_sel_prio = best_ni->ni_sel_priority;
1675                 best_ni_fatal = atomic_read(&best_ni->ni_fatal_error_on);
1676         }
1677
1678         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1679                 unsigned int distance;
1680                 int ni_credits;
1681                 int ni_healthv;
1682                 int ni_fatal;
1683                 __u32 ni_sel_prio;
1684                 unsigned int ni_dev_prio;
1685
1686                 ni_credits = atomic_read(&ni->ni_tx_credits);
1687                 ni_healthv = atomic_read(&ni->ni_healthv);
1688                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1689                 ni_sel_prio = ni->ni_sel_priority;
1690
1691                 /*
1692                  * calculate the distance from the CPT on which
1693                  * the message memory is allocated to the CPT of
1694                  * the NI's physical device
1695                  */
1696                 distance = cfs_cpt_distance(lnet_cpt_table(),
1697                                             md_cpt,
1698                                             ni->ni_dev_cpt);
1699
1700                 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1701
1702                 /*
1703                  * All distances smaller than the NUMA range
1704                  * are treated equally.
1705                  */
1706                 if (!gpu && distance < lnet_numa_range)
1707                         distance = lnet_numa_range;
1708
1709                 /*
1710                  * Select on health, selection policy, direct dma prio,
1711                  * shorter distance, available credits, then round-robin.
1712                  */
1713                 if (best_ni)
1714                         CDEBUG(D_NET, "compare ni %s [f:%s, c:%d, d:%d, s:%d, p:%u, g:%u, h:%d] with best_ni %s [f:%s, c:%d, d:%d, s:%d, p:%u, g:%u, h:%d]\n",
1715                                libcfs_nidstr(&ni->ni_nid),
1716                                ni_fatal ? "y" : "n", ni_credits, distance,
1717                                ni->ni_seq, ni_sel_prio, ni_dev_prio, ni_healthv,
1718                                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1719                                : "not selected",
1720                                best_ni_fatal ? "y" : "n", best_credits,
1721                                shortest_distance,
1722                                (best_ni) ? best_ni->ni_seq : 0,
1723                                best_sel_prio, best_dev_prio, best_healthv);
1724                 else
1725                         goto select_ni;
1726
1727                 if (ni_fatal && !best_ni_fatal)
1728                         continue;
1729                 else if (!ni_fatal && best_ni_fatal)
1730                         goto select_ni;
1731
1732                 if (ni_healthv < best_healthv)
1733                         continue;
1734                 else if (ni_healthv > best_healthv)
1735                         goto select_ni;
1736
1737                 if (ni_sel_prio > best_sel_prio)
1738                         continue;
1739                 else if (ni_sel_prio < best_sel_prio)
1740                         goto select_ni;
1741
1742                 if (ni_dev_prio > best_dev_prio)
1743                         continue;
1744                 else if (ni_dev_prio < best_dev_prio)
1745                         goto select_ni;
1746
1747                 if (distance > shortest_distance)
1748                         continue;
1749                 else if (distance < shortest_distance)
1750                         goto select_ni;
1751
1752                 if (ni_credits < best_credits)
1753                         continue;
1754                 else if (ni_credits > best_credits)
1755                         goto select_ni;
1756
1757                 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1758                         continue;
1759
1760 select_ni:
1761                 best_sel_prio = ni_sel_prio;
1762                 best_dev_prio = ni_dev_prio;
1763                 shortest_distance = distance;
1764                 best_healthv = ni_healthv;
1765                 best_ni = ni;
1766                 best_credits = ni_credits;
1767                 best_ni_fatal = ni_fatal;
1768         }
1769
1770         CDEBUG(D_NET, "selected best_ni %s\n",
1771                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1772
1773         return best_ni;
1774 }
1775
1776 static bool
1777 lnet_reserved_msg(struct lnet_msg *msg)
1778 {
1779         if (msg->msg_type == LNET_MSG_PUT) {
1780                 if (msg->msg_hdr.msg.put.ptl_index == LNET_RESERVED_PORTAL)
1781                         return true;
1782         } else if (msg->msg_type == LNET_MSG_GET) {
1783                 if (msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1784                         return true;
1785         }
1786         return false;
1787 }
1788
1789 /* Can the specified message trigger peer discovery?
1790  *
1791  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1792  * because such traffic is required to perform discovery. We therefore
1793  * exclude all GET and PUT on that portal. We also exclude all ACK and
1794  * REPLY traffic, but that is because the portal is not tracked in the
1795  * message structure for these message types. We could restrict this
1796  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1797  */
1798 static bool
1799 lnet_msg_discovery(struct lnet_msg *msg)
1800 {
1801         return !(lnet_reserved_msg(msg) || lnet_msg_is_response(msg));
1802 }
1803
1804 /* Is the specified message an LNet ping?
1805  */
1806 static bool
1807 lnet_msg_is_ping(struct lnet_msg *msg)
1808 {
1809         if (msg->msg_type == LNET_MSG_GET &&
1810             msg->msg_hdr.msg.get.ptl_index == LNET_RESERVED_PORTAL)
1811                 return true;
1812
1813         return false;
1814 }
1815
1816 #define SRC_SPEC        0x0001
1817 #define SRC_ANY         0x0002
1818 #define LOCAL_DST       0x0004
1819 #define REMOTE_DST      0x0008
1820 #define MR_DST          0x0010
1821 #define NMR_DST         0x0020
1822 #define SND_RESP        0x0040
1823
1824 /* The following to defines are used for return codes */
1825 #define REPEAT_SEND     0x1000
1826 #define PASS_THROUGH    0x2000
1827
1828 /* The different cases lnet_select pathway needs to handle */
1829 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1830 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1831 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1832 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1833 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1834 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1835 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1836 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1837
1838 static int
1839 lnet_handle_lo_send(struct lnet_send_data *sd)
1840 {
1841         struct lnet_msg *msg = sd->sd_msg;
1842         int cpt = sd->sd_cpt;
1843
1844         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1845                 return -ESHUTDOWN;
1846
1847         /* No send credit hassles with LOLND */
1848         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1849         msg->msg_hdr.dest_nid = the_lnet.ln_loni->ni_nid;
1850         if (!msg->msg_routing)
1851                 msg->msg_hdr.src_nid = the_lnet.ln_loni->ni_nid;
1852         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1853         lnet_msg_commit(msg, cpt);
1854         msg->msg_txni = the_lnet.ln_loni;
1855
1856         return LNET_CREDIT_OK;
1857 }
1858
1859 static int
1860 lnet_handle_send(struct lnet_send_data *sd)
1861 {
1862         struct lnet_ni *best_ni = sd->sd_best_ni;
1863         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1864         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1865         struct lnet_msg *msg = sd->sd_msg;
1866         int cpt2;
1867         __u32 send_case = sd->sd_send_case;
1868         int rc;
1869         __u32 routing = send_case & REMOTE_DST;
1870         struct lnet_rsp_tracker *rspt;
1871
1872         /* Increment sequence number of the selected peer, peer net,
1873          * local ni and local net so that we pick the next ones
1874          * in Round Robin.
1875          */
1876         best_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1877         best_lpni->lpni_peer_net->lpn_seq =
1878                 best_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1879         best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1880         the_lnet.ln_net_seq++;
1881         best_ni->ni_net->net_seq = the_lnet.ln_net_seq;
1882         best_ni->ni_seq = best_ni->ni_net->net_seq;
1883
1884         CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1885                libcfs_nidstr(&best_ni->ni_nid),
1886                best_ni->ni_seq, best_ni->ni_net->net_seq,
1887                atomic_read(&best_ni->ni_tx_credits),
1888                best_ni->ni_sel_priority,
1889                libcfs_nidstr(&best_lpni->lpni_nid),
1890                best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1891                best_lpni->lpni_txcredits,
1892                best_lpni->lpni_sel_priority);
1893
1894         /*
1895          * grab a reference on the peer_ni so it sticks around even if
1896          * we need to drop and relock the lnet_net_lock below.
1897          */
1898         lnet_peer_ni_addref_locked(best_lpni);
1899
1900         /*
1901          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1902          * message. This ensures that we get a CPT that is correct for
1903          * the NI when the NI has been restricted to a subset of all CPTs.
1904          * If the selected CPT differs from the one currently locked, we
1905          * must unlock and relock the lnet_net_lock(), and then check whether
1906          * the configuration has changed. We don't have a hold on the best_ni
1907          * yet, and it may have vanished.
1908          */
1909         cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1910         if (sd->sd_cpt != cpt2) {
1911                 __u32 seq = lnet_get_dlc_seq_locked();
1912                 lnet_net_unlock(sd->sd_cpt);
1913                 sd->sd_cpt = cpt2;
1914                 lnet_net_lock(sd->sd_cpt);
1915                 if (seq != lnet_get_dlc_seq_locked()) {
1916                         lnet_peer_ni_decref_locked(best_lpni);
1917                         return REPEAT_SEND;
1918                 }
1919         }
1920
1921         /*
1922          * store the best_lpni in the message right away to avoid having
1923          * to do the same operation under different conditions
1924          */
1925         msg->msg_txpeer = best_lpni;
1926         msg->msg_txni = best_ni;
1927
1928         /*
1929          * grab a reference for the best_ni since now it's in use in this
1930          * send. The reference will be dropped in lnet_finalize()
1931          */
1932         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1933
1934         /*
1935          * Always set the target.nid to the best peer picked. Either the
1936          * NID will be one of the peer NIDs selected, or the same NID as
1937          * what was originally set in the target or it will be the NID of
1938          * a router if this message should be routed
1939          */
1940         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1941
1942         /*
1943          * lnet_msg_commit assigns the correct cpt to the message, which
1944          * is used to decrement the correct refcount on the ni when it's
1945          * time to return the credits
1946          */
1947         lnet_msg_commit(msg, sd->sd_cpt);
1948
1949         /*
1950          * If we are routing the message then we keep the src_nid that was
1951          * set by the originator. If we are not routing then we are the
1952          * originator and set it here.
1953          */
1954         if (!msg->msg_routing)
1955                 msg->msg_hdr.src_nid = msg->msg_txni->ni_nid;
1956
1957         if (routing) {
1958                 msg->msg_target_is_router = 1;
1959                 msg->msg_target.pid = LNET_PID_LUSTRE;
1960                 /*
1961                  * since we're routing we want to ensure that the
1962                  * msg_hdr.dest_nid is set to the final destination. When
1963                  * the router receives this message it knows how to route
1964                  * it.
1965                  *
1966                  * final_dst_lpni is set at the beginning of the
1967                  * lnet_select_pathway() function and is never changed.
1968                  * It's safe to use it here.
1969                  */
1970                 final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq++;
1971                 final_dst_lpni->lpni_peer_net->lpn_seq =
1972                         final_dst_lpni->lpni_peer_net->lpn_peer->lp_send_seq;
1973                 final_dst_lpni->lpni_seq =
1974                         final_dst_lpni->lpni_peer_net->lpn_seq;
1975                 msg->msg_hdr.dest_nid = final_dst_lpni->lpni_nid;
1976         } else {
1977                 /*
1978                  * if we're not routing set the dest_nid to the best peer
1979                  * ni NID that we picked earlier in the algorithm.
1980                  */
1981                 msg->msg_hdr.dest_nid = msg->msg_txpeer->lpni_nid;
1982         }
1983
1984         /*
1985          * if we have response tracker block update it with the next hop
1986          * nid
1987          */
1988         if (msg->msg_md) {
1989                 rspt = msg->msg_md->md_rspt_ptr;
1990                 if (rspt) {
1991                         rspt->rspt_next_hop_nid =
1992                                 msg->msg_txpeer->lpni_nid;
1993                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1994                                libcfs_nidstr(&rspt->rspt_next_hop_nid));
1995                 }
1996         }
1997
1998         rc = lnet_post_send_locked(msg, 0);
1999
2000         if (!rc)
2001                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
2002                        libcfs_nidstr(&msg->msg_hdr.src_nid),
2003                        libcfs_nidstr(&msg->msg_txni->ni_nid),
2004                        libcfs_nidstr(&sd->sd_src_nid),
2005                        libcfs_nidstr(&msg->msg_hdr.dest_nid),
2006                        libcfs_nidstr(&sd->sd_dst_nid),
2007                        libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
2008                        libcfs_nidstr(&sd->sd_rtr_nid),
2009                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
2010
2011         return rc;
2012 }
2013
2014 static inline void
2015 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
2016                          struct lnet_msg *msg)
2017 {
2018         if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
2019             !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
2020                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
2021                        libcfs_nidstr(&lni->ni_nid),
2022                        libcfs_nidstr(&lpni->lpni_nid));
2023                 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
2024         }
2025 }
2026
2027 /*
2028  * Source Specified
2029  * Local Destination
2030  * non-mr peer
2031  *
2032  * use the source and destination NIDs as the pathway
2033  */
2034 static int
2035 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2036 {
2037         /* the destination lpni is set before we get here. */
2038
2039         /* find local NI */
2040         sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2041         if (!sd->sd_best_ni) {
2042                 CERROR("Can't send to %s: src %s is not a local nid\n",
2043                        libcfs_nidstr(&sd->sd_dst_nid),
2044                        libcfs_nidstr(&sd->sd_src_nid));
2045                 return -EINVAL;
2046         }
2047
2048         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2049
2050         return lnet_handle_send(sd);
2051 }
2052
2053 /*
2054  * Source Specified
2055  * Local Destination
2056  * MR Peer
2057  *
2058  * Don't run the selection algorithm on the peer NIs. By specifying the
2059  * local NID, we're also saying that we should always use the destination NID
2060  * provided. This handles the case where we should be using the same
2061  * destination NID for the all the messages which belong to the same RPC
2062  * request.
2063  */
2064 static int
2065 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2066 {
2067         sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2068         if (!sd->sd_best_ni) {
2069                 CERROR("Can't send to %s: src %s is not a local nid\n",
2070                        libcfs_nidstr(&sd->sd_dst_nid),
2071                        libcfs_nidstr(&sd->sd_src_nid));
2072                 return -EINVAL;
2073         }
2074
2075         if (sd->sd_best_lpni &&
2076             nid_same(&sd->sd_best_lpni->lpni_nid,
2077                       &the_lnet.ln_loni->ni_nid))
2078                 return lnet_handle_lo_send(sd);
2079         else if (sd->sd_best_lpni)
2080                 return lnet_handle_send(sd);
2081
2082         CERROR("can't send to %s. no NI on %s\n",
2083                libcfs_nidstr(&sd->sd_dst_nid),
2084                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2085
2086         return -EHOSTUNREACH;
2087 }
2088
2089 static struct lnet_ni *
2090 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2091                               struct lnet_peer *peer,
2092                               struct lnet_peer_net *peer_net,
2093                               struct lnet_msg *msg,
2094                               int cpt)
2095 {
2096         struct lnet_net *local_net;
2097         struct lnet_ni *best_ni;
2098
2099         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2100         if (!local_net)
2101                 return NULL;
2102
2103         /*
2104          * Iterate through the NIs in this local Net and select
2105          * the NI to send from. The selection is determined by
2106          * these 3 criterion in the following priority:
2107          *      1. NUMA
2108          *      2. NI available credits
2109          *      3. Round Robin
2110          */
2111         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2112                                    peer, peer_net, msg, cpt);
2113
2114         return best_ni;
2115 }
2116
2117 static int
2118 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2119                              int cpt)
2120 {
2121         struct lnet_peer *peer;
2122         struct lnet_peer_ni *new_lpni;
2123         int rc;
2124
2125         lnet_peer_ni_addref_locked(lpni);
2126
2127         peer = lpni->lpni_peer_net->lpn_peer;
2128
2129         if (lnet_peer_gw_discovery(peer)) {
2130                 lnet_peer_ni_decref_locked(lpni);
2131                 return 0;
2132         }
2133
2134         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2135                 lnet_peer_ni_decref_locked(lpni);
2136                 return 0;
2137         }
2138
2139         rc = lnet_discover_peer_locked(lpni, cpt, false);
2140         if (rc) {
2141                 lnet_peer_ni_decref_locked(lpni);
2142                 return rc;
2143         }
2144
2145         new_lpni = lnet_peer_ni_find_locked(&lpni->lpni_nid);
2146         if (!new_lpni) {
2147                 lnet_peer_ni_decref_locked(lpni);
2148                 return -ENOENT;
2149         }
2150
2151         peer = new_lpni->lpni_peer_net->lpn_peer;
2152         spin_lock(&peer->lp_lock);
2153         if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2154                 /* The peer NI did not change and the peer is up to date.
2155                  * Nothing more to do.
2156                  */
2157                 spin_unlock(&peer->lp_lock);
2158                 lnet_peer_ni_decref_locked(lpni);
2159                 lnet_peer_ni_decref_locked(new_lpni);
2160                 return 0;
2161         }
2162         spin_unlock(&peer->lp_lock);
2163
2164         /* Either the peer NI changed during discovery, or the peer isn't up
2165          * to date. In both cases we want to queue the message on the
2166          * (possibly new) peer's pending queue and queue the peer for discovery
2167          */
2168         msg->msg_sending = 0;
2169         msg->msg_txpeer = NULL;
2170         lnet_net_unlock(cpt);
2171         lnet_peer_queue_message(peer, msg);
2172         lnet_net_lock(cpt);
2173
2174         lnet_peer_ni_decref_locked(lpni);
2175         lnet_peer_ni_decref_locked(new_lpni);
2176
2177         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2178                msg, libcfs_nidstr(&peer->lp_primary_nid));
2179
2180         return LNET_DC_WAIT;
2181 }
2182
2183 static int
2184 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2185                              struct lnet_nid *dst_nid,
2186                              struct lnet_peer_ni **gw_lpni,
2187                              struct lnet_peer **gw_peer)
2188 {
2189         int rc;
2190         struct lnet_peer *gw;
2191         struct lnet_peer *lp;
2192         struct lnet_peer_net *lpn;
2193         struct lnet_peer_net *best_lpn = NULL;
2194         struct lnet_remotenet *rnet, *best_rnet = NULL;
2195         struct lnet_route *best_route = NULL;
2196         struct lnet_route *last_route = NULL;
2197         struct lnet_peer_ni *lpni = NULL;
2198         struct lnet_peer_ni *gwni = NULL;
2199         bool route_found = false;
2200         struct lnet_nid *src_nid =
2201                 !LNET_NID_IS_ANY(&sd->sd_src_nid) || !sd->sd_best_ni
2202                 ? &sd->sd_src_nid
2203                 : &sd->sd_best_ni->ni_nid;
2204         int best_lpn_healthv = 0;
2205         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2206
2207         CDEBUG(D_NET, "%s route (%s) from local NI %s to destination %s\n",
2208                LNET_NID_IS_ANY(&sd->sd_rtr_nid) ? "Lookup" : "Specified",
2209                libcfs_nidstr(&sd->sd_rtr_nid), libcfs_nidstr(src_nid),
2210                libcfs_nidstr(&sd->sd_dst_nid));
2211
2212         /* If a router nid was specified then we are replying to a GET or
2213          * sending an ACK. In this case we use the gateway associated with the
2214          * specified router nid.
2215          */
2216         if (!LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2217                 gwni = lnet_peer_ni_find_locked(&sd->sd_rtr_nid);
2218                 if (gwni) {
2219                         gw = gwni->lpni_peer_net->lpn_peer;
2220                         lnet_peer_ni_decref_locked(gwni);
2221                         if (gw->lp_rtr_refcount)
2222                                 route_found = true;
2223                 } else {
2224                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2225                               libcfs_nidstr(&sd->sd_rtr_nid));
2226                 }
2227         }
2228
2229         if (!route_found) {
2230                 if (sd->sd_msg->msg_routing || !LNET_NID_IS_ANY(src_nid)) {
2231                         /* If I'm routing this message then I need to find the
2232                          * next hop based on the destination NID
2233                          *
2234                          * We also find next hop based on the destination NID
2235                          * if the source NI was specified
2236                          */
2237                         best_rnet = lnet_find_rnet_locked(LNET_NID_NET(&sd->sd_dst_nid));
2238                         if (!best_rnet) {
2239                                 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2240                                        (src_nid && LNET_NID_IS_ANY(src_nid)) ?
2241                                                 "any local NI" :
2242                                                 libcfs_nidstr(src_nid),
2243                                        libcfs_nidstr(&sd->sd_dst_nid));
2244                                 return -EHOSTUNREACH;
2245                         }
2246                         CDEBUG(D_NET, "best_rnet %s\n",
2247                                libcfs_net2str(best_rnet->lrn_net));
2248                 } else {
2249                         /* we've already looked up the initial lpni using
2250                          * dst_nid
2251                          */
2252                         lpni = sd->sd_best_lpni;
2253                         /* the peer tree must be in existence */
2254                         LASSERT(lpni && lpni->lpni_peer_net &&
2255                                 lpni->lpni_peer_net->lpn_peer);
2256                         lp = lpni->lpni_peer_net->lpn_peer;
2257
2258                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2259                                 /* is this remote network reachable?  */
2260                                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2261                                 if (!rnet)
2262                                         continue;
2263
2264                                 if (!best_lpn)
2265                                         goto use_lpn;
2266                                 else
2267                                         CDEBUG(D_NET, "n[%s, %s] h[%d, %d], p[%u, %u], s[%d, %d]\n",
2268                                                libcfs_net2str(lpn->lpn_net_id),
2269                                                libcfs_net2str(best_lpn->lpn_net_id),
2270                                                lpn->lpn_healthv,
2271                                                best_lpn->lpn_healthv,
2272                                                lpn->lpn_sel_priority,
2273                                                best_lpn->lpn_sel_priority,
2274                                                lpn->lpn_seq,
2275                                                best_lpn->lpn_seq);
2276
2277                                 /* select the preferred peer net */
2278                                 if (best_lpn_healthv > lpn->lpn_healthv)
2279                                         continue;
2280                                 else if (best_lpn_healthv < lpn->lpn_healthv)
2281                                         goto use_lpn;
2282
2283                                 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2284                                         continue;
2285                                 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2286                                         goto use_lpn;
2287
2288                                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2289                                         continue;
2290 use_lpn:
2291                                 best_lpn_healthv = lpn->lpn_healthv;
2292                                 best_lpn_sel_prio = lpn->lpn_sel_priority;
2293                                 best_lpn = lpn;
2294                                 best_rnet = rnet;
2295                         }
2296
2297                         if (!best_lpn) {
2298                                 CERROR("peer %s has no available nets\n",
2299                                        libcfs_nidstr(&sd->sd_dst_nid));
2300                                 return -EHOSTUNREACH;
2301                         }
2302
2303                         CDEBUG(D_NET, "selected best_lpn %s\n",
2304                                libcfs_net2str(best_lpn->lpn_net_id));
2305
2306                         sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2307                                                                &sd->sd_dst_nid,
2308                                                                lp,
2309                                                                best_lpn->lpn_net_id);
2310                         if (!sd->sd_best_lpni) {
2311                                 CERROR("peer %s is unreachable\n",
2312                                        libcfs_nidstr(&sd->sd_dst_nid));
2313                                 return -EHOSTUNREACH;
2314                         }
2315
2316                         /* We're attempting to round robin over the remote peer
2317                          * NI's so update the final destination we selected
2318                          */
2319                         sd->sd_final_dst_lpni = sd->sd_best_lpni;
2320                 }
2321
2322                 /*
2323                  * find the best route. Restrict the selection on the net of the
2324                  * local NI if we've already picked the local NI to send from.
2325                  * Otherwise, let's pick any route we can find and then find
2326                  * a local NI we can reach the route's gateway on. Any route we
2327                  * select will be reachable by virtue of the restriction we have
2328                  * when adding a route.
2329                  */
2330                 best_route = lnet_find_route_locked(best_rnet,
2331                                                     LNET_NID_NET(src_nid),
2332                                                     sd->sd_best_lpni,
2333                                                     &last_route, &gwni);
2334
2335                 if (!best_route) {
2336                         CERROR("no route to %s from %s\n",
2337                                libcfs_nidstr(dst_nid),
2338                                libcfs_nidstr(src_nid));
2339                         return -EHOSTUNREACH;
2340                 }
2341
2342                 if (!gwni) {
2343                         CERROR("Internal Error. Route expected to %s from %s\n",
2344                                libcfs_nidstr(dst_nid),
2345                                libcfs_nidstr(src_nid));
2346                         return -EFAULT;
2347                 }
2348
2349                 gw = best_route->lr_gateway;
2350                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2351         }
2352
2353         /*
2354          * If the router checker is not active then discover the gateway here.
2355          * This ensures we are able to take advantage of multi-rail routing, but
2356          * if the router checker is active then we do not unecessarily delay
2357          * messages while the gateway is being checked by the dedicated monitor
2358          * thread.
2359          *
2360          * NB: We're only checking the alive_router_check_interval here, rather
2361          * than calling lnet_router_checker_active(), because the other
2362          * conditions that are checked by that function are either
2363          * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2364          * is not empty)
2365          */
2366         if (alive_router_check_interval <= 0) {
2367                 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2368                 if (rc)
2369                         return rc;
2370         }
2371
2372         if (!sd->sd_best_ni) {
2373                 lpn = gwni->lpni_peer_net;
2374                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2375                                                                sd->sd_msg,
2376                                                                sd->sd_md_cpt);
2377                 if (!sd->sd_best_ni) {
2378                         CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2379                                libcfs_net2str(lpn->lpn_net_id),
2380                                libcfs_nidstr(&sd->sd_src_nid));
2381                         return -EFAULT;
2382                 }
2383         }
2384
2385         *gw_lpni = gwni;
2386         *gw_peer = gw;
2387
2388         /*
2389          * increment the sequence number since now we're sure we're
2390          * going to use this route
2391          */
2392         if (LNET_NID_IS_ANY(&sd->sd_rtr_nid)) {
2393                 LASSERT(best_route && last_route);
2394                 best_route->lr_seq = last_route->lr_seq + 1;
2395         }
2396
2397         return 0;
2398 }
2399
2400 /*
2401  * Handle two cases:
2402  *
2403  * Case 1:
2404  *  Source specified
2405  *  Remote destination
2406  *  Non-MR destination
2407  *
2408  * Case 2:
2409  *  Source specified
2410  *  Remote destination
2411  *  MR destination
2412  *
2413  * The handling of these two cases is similar. Even though the destination
2414  * can be MR or non-MR, we'll deal directly with the router.
2415  */
2416 static int
2417 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2418 {
2419         int rc;
2420         struct lnet_peer_ni *gw_lpni = NULL;
2421         struct lnet_peer *gw_peer = NULL;
2422
2423         /* find local NI */
2424         sd->sd_best_ni = lnet_nid_to_ni_locked(&sd->sd_src_nid, sd->sd_cpt);
2425         if (!sd->sd_best_ni) {
2426                 CERROR("Can't send to %s: src %s is not a local nid\n",
2427                        libcfs_nidstr(&sd->sd_dst_nid),
2428                        libcfs_nidstr(&sd->sd_src_nid));
2429                 return -EINVAL;
2430         }
2431
2432         rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2433                                           &gw_lpni, &gw_peer);
2434         if (rc)
2435                 return rc;
2436
2437         if (sd->sd_send_case & NMR_DST)
2438                 /*
2439                  * since the final destination is non-MR let's set its preferred
2440                  * NID before we send
2441                  */
2442                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2443                                          sd->sd_msg);
2444
2445         /*
2446          * We're going to send to the gw found so let's set its
2447          * info
2448          */
2449         sd->sd_peer = gw_peer;
2450         sd->sd_best_lpni = gw_lpni;
2451
2452         return lnet_handle_send(sd);
2453 }
2454
2455 static struct lnet_ni *
2456 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2457                                struct lnet_msg *msg, bool discovery)
2458 {
2459         struct lnet_peer_net *lpn = NULL;
2460         struct lnet_peer_net *best_lpn = NULL;
2461         struct lnet_net *net = NULL;
2462         struct lnet_net *best_net = NULL;
2463         struct lnet_ni *best_ni = NULL;
2464         int best_lpn_healthv = 0;
2465         int best_net_healthv = 0;
2466         int net_healthv;
2467         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2468         __u32 lpn_sel_prio;
2469         __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2470         __u32 net_sel_prio;
2471
2472         /* If lp_disc_net_id is set, this peer is a router undergoing
2473          * discovery, and this message is an LNet ping, then this may be a
2474          * discovery message and we need to select an NI on the peer net
2475          * specified by lp_disc_net_id
2476          */
2477         if (peer->lp_disc_net_id &&
2478             (peer->lp_state & LNET_PEER_RTR_DISCOVERY) &&
2479             lnet_msg_is_ping(msg)) {
2480                 best_lpn = lnet_peer_get_net_locked(peer, peer->lp_disc_net_id);
2481                 if (best_lpn && lnet_get_net_locked(best_lpn->lpn_net_id))
2482                         goto select_best_ni;
2483         }
2484
2485         /*
2486          * The peer can have multiple interfaces, some of them can be on
2487          * the local network and others on a routed network. We should
2488          * prefer the local network. However if the local network is not
2489          * available then we need to try the routed network
2490          */
2491
2492         /* go through all the peer nets and find the best_ni */
2493         list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2494                 /*
2495                  * The peer's list of nets can contain non-local nets. We
2496                  * want to only examine the local ones.
2497                  */
2498                 net = lnet_get_net_locked(lpn->lpn_net_id);
2499                 if (!net)
2500                         continue;
2501
2502                 lpn_sel_prio = lpn->lpn_sel_priority;
2503                 net_healthv = lnet_get_net_healthv_locked(net);
2504                 net_sel_prio = net->net_sel_priority;
2505
2506                 if (!best_lpn)
2507                         goto select_lpn;
2508                 else
2509                         CDEBUG(D_NET,
2510                                "n[%s, %s] ph[%d, %d], pp[%u, %u], nh[%d, %d], np[%u, %u], ps[%u, %u], ns[%u, %u]\n",
2511                                libcfs_net2str(lpn->lpn_net_id),
2512                                libcfs_net2str(best_lpn->lpn_net_id),
2513                                lpn->lpn_healthv,
2514                                best_lpn_healthv,
2515                                lpn_sel_prio,
2516                                best_lpn_sel_prio,
2517                                net_healthv,
2518                                best_net_healthv,
2519                                net_sel_prio,
2520                                best_net_sel_prio,
2521                                lpn->lpn_seq,
2522                                best_lpn->lpn_seq,
2523                                net->net_seq,
2524                                best_net->net_seq);
2525
2526                 /* always select the lpn with the best health */
2527                 if (best_lpn_healthv > lpn->lpn_healthv)
2528                         continue;
2529                 else if (best_lpn_healthv < lpn->lpn_healthv)
2530                         goto select_lpn;
2531
2532                 /* select the preferred peer and local nets */
2533                 if (best_lpn_sel_prio < lpn_sel_prio)
2534                         continue;
2535                 else if (best_lpn_sel_prio > lpn_sel_prio)
2536                         goto select_lpn;
2537
2538                 if (best_net_healthv > net_healthv)
2539                         continue;
2540                 else if (best_net_healthv < net_healthv)
2541                         goto select_lpn;
2542
2543                 if (best_net_sel_prio < net_sel_prio)
2544                         continue;
2545                 else if (best_net_sel_prio > net_sel_prio)
2546                         goto select_lpn;
2547
2548                 if (best_lpn->lpn_seq < lpn->lpn_seq)
2549                         continue;
2550                 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2551                         goto select_lpn;
2552
2553                 /* round robin over the local networks */
2554                 if (best_net->net_seq <= net->net_seq)
2555                         continue;
2556
2557 select_lpn:
2558                 best_net_healthv = net_healthv;
2559                 best_net_sel_prio = net_sel_prio;
2560                 best_lpn_healthv = lpn->lpn_healthv;
2561                 best_lpn_sel_prio = lpn_sel_prio;
2562                 best_lpn = lpn;
2563                 best_net = net;
2564         }
2565
2566         if (best_lpn) {
2567                 /* Select the best NI on the same net as best_lpn chosen
2568                  * above
2569                  */
2570 select_best_ni:
2571                 CDEBUG(D_NET, "selected best_lpn %s\n",
2572                        libcfs_net2str(best_lpn->lpn_net_id));
2573                 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2574                                                         msg, md_cpt);
2575         }
2576
2577         return best_ni;
2578 }
2579
2580 static struct lnet_ni *
2581 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2582 {
2583         struct lnet_ni *best_ni = NULL;
2584         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2585         struct lnet_peer_ni *lpni_entry;
2586
2587         /*
2588          * We must use a consistent source address when sending to a
2589          * non-MR peer. However, a non-MR peer can have multiple NIDs
2590          * on multiple networks, and we may even need to talk to this
2591          * peer on multiple networks -- certain types of
2592          * load-balancing configuration do this.
2593          *
2594          * So we need to pick the NI the peer prefers for this
2595          * particular network.
2596          */
2597         LASSERT(peer_net);
2598         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2599                             lpni_peer_nis) {
2600                 if (lpni_entry->lpni_pref_nnids == 0)
2601                         continue;
2602                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2603                 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2604                                                 cpt);
2605                 break;
2606         }
2607
2608         return best_ni;
2609 }
2610
2611 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2612 static int
2613 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2614 {
2615         struct lnet_ni *best_ni = NULL;
2616
2617         /*
2618          * We must use a consistent source address when sending to a
2619          * non-MR peer. However, a non-MR peer can have multiple NIDs
2620          * on multiple networks, and we may even need to talk to this
2621          * peer on multiple networks -- certain types of
2622          * load-balancing configuration do this.
2623          *
2624          * So we need to pick the NI the peer prefers for this
2625          * particular network.
2626          *
2627          * An exception is traffic on LNET_RESERVED_PORTAL. Internal LNet
2628          * traffic doesn't care which source NI is used, and we don't actually
2629          * want to restrict local recovery pings to a single source NI.
2630          */
2631         if (!lnet_reserved_msg(sd->sd_msg))
2632                 best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2633                                                                sd->sd_cpt);
2634
2635         if (!best_ni)
2636                 best_ni = lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2637                                                 sd->sd_best_lpni->lpni_peer_net,
2638                                                 sd->sd_msg,
2639                                                 sd->sd_md_cpt);
2640
2641         /* If there is no best_ni we don't have a route */
2642         if (!best_ni) {
2643                 CERROR("no path to %s from net %s\n",
2644                         libcfs_nidstr(&sd->sd_best_lpni->lpni_nid),
2645                         libcfs_net2str(sd->sd_best_lpni->lpni_net->net_id));
2646                 return -EHOSTUNREACH;
2647         }
2648
2649         sd->sd_best_ni = best_ni;
2650
2651         /* Set preferred NI if necessary. */
2652         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2653
2654         return 0;
2655 }
2656
2657
2658 /*
2659  * Source not specified
2660  * Local destination
2661  * Non-MR Peer
2662  *
2663  * always use the same source NID for NMR peers
2664  * If we've talked to that peer before then we already have a preferred
2665  * source NI associated with it. Otherwise, we select a preferred local NI
2666  * and store it in the peer
2667  */
2668 static int
2669 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2670 {
2671         int rc = 0;
2672
2673         /* sd->sd_best_lpni is already set to the final destination */
2674
2675         /*
2676          * At this point we should've created the peer ni and peer. If we
2677          * can't find it, then something went wrong. Instead of assert
2678          * output a relevant message and fail the send
2679          */
2680         if (!sd->sd_best_lpni) {
2681                 CERROR("Internal fault. Unable to send msg %s to %s. NID not known\n",
2682                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2683                        libcfs_nidstr(&sd->sd_dst_nid));
2684                 return -EFAULT;
2685         }
2686
2687         if (sd->sd_msg->msg_routing) {
2688                 /* If I'm forwarding this message then I can choose any NI
2689                  * on the destination peer net
2690                  */
2691                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2692                                                                sd->sd_peer,
2693                                                                sd->sd_best_lpni->lpni_peer_net,
2694                                                                sd->sd_msg,
2695                                                                sd->sd_md_cpt);
2696                 if (!sd->sd_best_ni) {
2697                         CERROR("Unable to forward message to %s. No local NI available\n",
2698                                libcfs_nidstr(&sd->sd_dst_nid));
2699                         rc = -EHOSTUNREACH;
2700                 }
2701         } else
2702                 rc = lnet_select_preferred_best_ni(sd);
2703
2704         if (!rc)
2705                 rc = lnet_handle_send(sd);
2706
2707         return rc;
2708 }
2709
2710 static int
2711 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2712 {
2713         /*
2714          * NOTE we've already handled the remote peer case. So we only
2715          * need to worry about the local case here.
2716          *
2717          * if we're sending a response, ACK or reply, we need to send it
2718          * to the destination NID given to us. At this point we already
2719          * have the peer_ni we're suppose to send to, so just find the
2720          * best_ni on the peer net and use that. Since we're sending to an
2721          * MR peer then we can just run the selection algorithm on our
2722          * local NIs and pick the best one.
2723          */
2724         if (sd->sd_send_case & SND_RESP) {
2725                 sd->sd_best_ni =
2726                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2727                                                 sd->sd_best_lpni->lpni_peer_net,
2728                                                 sd->sd_msg,
2729                                                 sd->sd_md_cpt);
2730
2731                 if (!sd->sd_best_ni) {
2732                         /*
2733                          * We're not going to deal with not able to send
2734                          * a response to the provided final destination
2735                          */
2736                         CERROR("Can't send response to %s. No local NI available\n",
2737                                 libcfs_nidstr(&sd->sd_dst_nid));
2738                         return -EHOSTUNREACH;
2739                 }
2740
2741                 return lnet_handle_send(sd);
2742         }
2743
2744         /*
2745          * If we get here that means we're sending a fresh request, PUT or
2746          * GET, so we need to run our standard selection algorithm.
2747          * First find the best local interface that's on any of the peer's
2748          * networks.
2749          */
2750         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2751                                         sd->sd_md_cpt,
2752                                         sd->sd_msg,
2753                                         lnet_msg_discovery(sd->sd_msg));
2754         if (sd->sd_best_ni) {
2755                 sd->sd_best_lpni =
2756                   lnet_find_best_lpni(sd->sd_best_ni, &sd->sd_dst_nid,
2757                                       sd->sd_peer,
2758                                       sd->sd_best_ni->ni_net->net_id);
2759
2760                 /*
2761                  * if we're successful in selecting a peer_ni on the local
2762                  * network, then send to it. Otherwise fall through and
2763                  * try and see if we can reach it over another routed
2764                  * network
2765                  */
2766                 if (sd->sd_best_lpni &&
2767                     nid_same(&sd->sd_best_lpni->lpni_nid,
2768                              &the_lnet.ln_loni->ni_nid)) {
2769                         /*
2770                          * in case we initially started with a routed
2771                          * destination, let's reset to local
2772                          */
2773                         sd->sd_send_case &= ~REMOTE_DST;
2774                         sd->sd_send_case |= LOCAL_DST;
2775                         return lnet_handle_lo_send(sd);
2776                 } else if (sd->sd_best_lpni) {
2777                         /*
2778                          * in case we initially started with a routed
2779                          * destination, let's reset to local
2780                          */
2781                         sd->sd_send_case &= ~REMOTE_DST;
2782                         sd->sd_send_case |= LOCAL_DST;
2783                         return lnet_handle_send(sd);
2784                 }
2785
2786                 CERROR("Internal Error. Expected to have a best_lpni: "
2787                        "%s -> %s\n",
2788                        libcfs_nidstr(&sd->sd_src_nid),
2789                        libcfs_nidstr(&sd->sd_dst_nid));
2790
2791                 return -EFAULT;
2792         }
2793
2794         /*
2795          * Peer doesn't have a local network. Let's see if there is
2796          * a remote network we can reach it on.
2797          */
2798         return PASS_THROUGH;
2799 }
2800
2801 /*
2802  * Case 1:
2803  *      Source NID not specified
2804  *      Local destination
2805  *      MR peer
2806  *
2807  * Case 2:
2808  *      Source NID not speified
2809  *      Remote destination
2810  *      MR peer
2811  *
2812  * In both of these cases if we're sending a response, ACK or REPLY, then
2813  * we need to send to the destination NID provided.
2814  *
2815  * In the remote case let's deal with MR routers.
2816  *
2817  */
2818
2819 static int
2820 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2821 {
2822         int rc = 0;
2823         struct lnet_peer *gw_peer = NULL;
2824         struct lnet_peer_ni *gw_lpni = NULL;
2825
2826         /*
2827          * handle sending a response to a remote peer here so we don't
2828          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2829          */
2830         if (sd->sd_send_case & REMOTE_DST &&
2831             sd->sd_send_case & SND_RESP) {
2832                 struct lnet_peer_ni *gw;
2833                 struct lnet_peer *gw_peer;
2834
2835                 rc = lnet_handle_find_routed_path(
2836                         sd, &sd->sd_dst_nid, &gw, &gw_peer);
2837                 if (rc < 0) {
2838                         CERROR("Can't send response to %s. No route available\n",
2839                                libcfs_nidstr(&sd->sd_dst_nid));
2840                         return -EHOSTUNREACH;
2841                 } else if (rc > 0) {
2842                         return rc;
2843                 }
2844
2845                 sd->sd_best_lpni = gw;
2846                 sd->sd_peer = gw_peer;
2847
2848                 return lnet_handle_send(sd);
2849         }
2850
2851         /*
2852          * Even though the NID for the peer might not be on a local network,
2853          * since the peer is MR there could be other interfaces on the
2854          * local network. In that case we'd still like to prefer the local
2855          * network over the routed network. If we're unable to do that
2856          * then we select the best router among the different routed networks,
2857          * and if the router is MR then we can deal with it as such.
2858          */
2859         rc = lnet_handle_any_mr_dsta(sd);
2860         if (rc != PASS_THROUGH)
2861                 return rc;
2862
2863         /*
2864          * Now that we must route to the destination, we must consider the
2865          * MR case, where the destination has multiple interfaces, some of
2866          * which we can route to and others we do not. For this reason we
2867          * need to select the destination which we can route to and if
2868          * there are multiple, we need to round robin.
2869          */
2870         rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid,
2871                                           &gw_lpni, &gw_peer);
2872         if (rc)
2873                 return rc;
2874
2875         sd->sd_send_case &= ~LOCAL_DST;
2876         sd->sd_send_case |= REMOTE_DST;
2877
2878         sd->sd_peer = gw_peer;
2879         sd->sd_best_lpni = gw_lpni;
2880
2881         return lnet_handle_send(sd);
2882 }
2883
2884 /*
2885  * Source not specified
2886  * Remote destination
2887  * Non-MR peer
2888  *
2889  * Must send to the specified peer NID using the same source NID that
2890  * we've used before. If it's the first time to talk to that peer then
2891  * find the source NI and assign it as preferred to that peer
2892  */
2893 static int
2894 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2895 {
2896         int rc;
2897         struct lnet_peer_ni *gw_lpni = NULL;
2898         struct lnet_peer *gw_peer = NULL;
2899
2900         /*
2901          * Let's see if we have a preferred NI to talk to this NMR peer
2902          */
2903         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2904                                                               sd->sd_cpt);
2905
2906         /*
2907          * find the router and that'll find the best NI if we didn't find
2908          * it already.
2909          */
2910         rc = lnet_handle_find_routed_path(sd, &sd->sd_dst_nid, &gw_lpni,
2911                                           &gw_peer);
2912         if (rc)
2913                 return rc;
2914
2915         /*
2916          * set the best_ni we've chosen as the preferred one for
2917          * this peer
2918          */
2919         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2920
2921         /* we'll be sending to the gw */
2922         sd->sd_best_lpni = gw_lpni;
2923         sd->sd_peer = gw_peer;
2924
2925         return lnet_handle_send(sd);
2926 }
2927
2928 static int
2929 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2930 {
2931         /*
2932          * turn off the SND_RESP bit.
2933          * It will be checked in the case handling
2934          */
2935         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2936
2937         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2938                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2939                 (send_case & SRC_SPEC) ? libcfs_nidstr(&sd->sd_src_nid) : "",
2940                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2941                 libcfs_nidstr(&sd->sd_dst_nid),
2942                 (send_case & LOCAL_DST) ? "local" : "routed");
2943
2944         switch (send_case) {
2945         /*
2946          * For all cases where the source is specified, we should always
2947          * use the destination NID, whether it's an MR destination or not,
2948          * since we're continuing a series of related messages for the
2949          * same RPC
2950          */
2951         case SRC_SPEC_LOCAL_NMR_DST:
2952                 return lnet_handle_spec_local_nmr_dst(sd);
2953         case SRC_SPEC_LOCAL_MR_DST:
2954                 return lnet_handle_spec_local_mr_dst(sd);
2955         case SRC_SPEC_ROUTER_NMR_DST:
2956         case SRC_SPEC_ROUTER_MR_DST:
2957                 return lnet_handle_spec_router_dst(sd);
2958         case SRC_ANY_LOCAL_NMR_DST:
2959                 return lnet_handle_any_local_nmr_dst(sd);
2960         case SRC_ANY_LOCAL_MR_DST:
2961         case SRC_ANY_ROUTER_MR_DST:
2962                 return lnet_handle_any_mr_dst(sd);
2963         case SRC_ANY_ROUTER_NMR_DST:
2964                 return lnet_handle_any_router_nmr_dst(sd);
2965         default:
2966                 CERROR("Unknown send case\n");
2967                 return -1;
2968         }
2969 }
2970
2971 static int
2972 lnet_select_pathway(struct lnet_nid *src_nid,
2973                     struct lnet_nid *dst_nid,
2974                     struct lnet_msg *msg,
2975                     struct lnet_nid *rtr_nid)
2976 {
2977         struct lnet_peer_ni *lpni;
2978         struct lnet_peer *peer;
2979         struct lnet_send_data send_data;
2980         int cpt, rc;
2981         int md_cpt;
2982         __u32 send_case = 0;
2983         bool final_hop;
2984         bool mr_forwarding_allowed;
2985
2986         memset(&send_data, 0, sizeof(send_data));
2987
2988         /*
2989          * get an initial CPT to use for locking. The idea here is not to
2990          * serialize the calls to select_pathway, so that as many
2991          * operations can run concurrently as possible. To do that we use
2992          * the CPT where this call is being executed. Later on when we
2993          * determine the CPT to use in lnet_message_commit, we switch the
2994          * lock and check if there was any configuration change.  If none,
2995          * then we proceed, if there is, then we restart the operation.
2996          */
2997         cpt = lnet_net_lock_current();
2998
2999         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
3000         if (md_cpt == CFS_CPT_ANY)
3001                 md_cpt = cpt;
3002
3003 again:
3004
3005         /*
3006          * If we're being asked to send to the loopback interface, there
3007          * is no need to go through any selection. We can just shortcut
3008          * the entire process and send over lolnd
3009          */
3010         send_data.sd_msg = msg;
3011         send_data.sd_cpt = cpt;
3012         if (nid_is_lo0(dst_nid)) {
3013                 rc = lnet_handle_lo_send(&send_data);
3014                 lnet_net_unlock(cpt);
3015                 return rc;
3016         }
3017
3018         /*
3019          * find an existing peer_ni, or create one and mark it as having been
3020          * created due to network traffic. This call will create the
3021          * peer->peer_net->peer_ni tree.
3022          */
3023         lpni = lnet_peerni_by_nid_locked(dst_nid, NULL, cpt);
3024         if (IS_ERR(lpni)) {
3025                 lnet_net_unlock(cpt);
3026                 return PTR_ERR(lpni);
3027         }
3028
3029         /*
3030          * Cache the original src_nid and rtr_nid. If we need to resend the
3031          * message then we'll need to know whether the src_nid was originally
3032          * specified for this message. If it was originally specified,
3033          * then we need to keep using the same src_nid since it's
3034          * continuing the same sequence of messages. Similarly, rtr_nid will
3035          * affect our choice of next hop.
3036          */
3037         if (src_nid)
3038                 msg->msg_src_nid_param = *src_nid;
3039         else
3040                 msg->msg_src_nid_param = LNET_ANY_NID;
3041         if (rtr_nid)
3042                 msg->msg_rtr_nid_param = *rtr_nid;
3043         else
3044                 msg->msg_rtr_nid_param = LNET_ANY_NID;
3045
3046         /*
3047          * If necessary, perform discovery on the peer that owns this peer_ni.
3048          * Note, this can result in the ownership of this peer_ni changing
3049          * to another peer object.
3050          */
3051         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
3052         if (rc) {
3053                 lnet_peer_ni_decref_locked(lpni);
3054                 lnet_net_unlock(cpt);
3055                 return rc;
3056         }
3057         lnet_peer_ni_decref_locked(lpni);
3058
3059         peer = lpni->lpni_peer_net->lpn_peer;
3060
3061         /*
3062          * Identify the different send cases
3063          */
3064         if (!src_nid || LNET_NID_IS_ANY(src_nid)) {
3065                 send_case |= SRC_ANY;
3066                 if (lnet_get_net_locked(LNET_NID_NET(dst_nid)))
3067                         send_case |= LOCAL_DST;
3068                 else
3069                         send_case |= REMOTE_DST;
3070         } else {
3071                 send_case |= SRC_SPEC;
3072                 if (LNET_NID_NET(src_nid) == LNET_NID_NET(dst_nid))
3073                         send_case |= LOCAL_DST;
3074                 else
3075                         send_case |= REMOTE_DST;
3076         }
3077
3078         final_hop = false;
3079         if (msg->msg_routing && (send_case & LOCAL_DST))
3080                 final_hop = true;
3081
3082         /* Determine whether to allow MR forwarding for this message.
3083          * NB: MR forwarding is allowed if the message originator and the
3084          * destination are both MR capable, and the destination lpni that was
3085          * originally chosen by the originator is unhealthy or down.
3086          * We check the MR capability of the destination further below
3087          */
3088         mr_forwarding_allowed = false;
3089         if (final_hop) {
3090                 struct lnet_peer *src_lp;
3091                 struct lnet_peer_ni *src_lpni;
3092
3093                 src_lpni = lnet_peerni_by_nid_locked(&msg->msg_hdr.src_nid,
3094                                                    NULL, cpt);
3095                 /* We don't fail the send if we hit any errors here. We'll just
3096                  * try to send it via non-multi-rail criteria
3097                  */
3098                 if (!IS_ERR(src_lpni)) {
3099                         /* Drop ref taken by lnet_peerni_by_nid_locked() */
3100                         lnet_peer_ni_decref_locked(src_lpni);
3101                         src_lp = lpni->lpni_peer_net->lpn_peer;
3102                         if (lnet_peer_is_multi_rail(src_lp) &&
3103                             !lnet_is_peer_ni_alive(lpni))
3104                                 mr_forwarding_allowed = true;
3105
3106                 }
3107                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3108                        mr_forwarding_allowed ? "allowed" : "not allowed");
3109         }
3110
3111         /*
3112          * Deal with the peer as NMR in the following cases:
3113          * 1. the peer is NMR
3114          * 2. We're trying to recover a specific peer NI
3115          * 3. I'm a router sending to the final destination and MR forwarding is
3116          *    not allowed for this message (as determined above).
3117          *    In this case the source of the message would've
3118          *    already selected the final destination so my job
3119          *    is to honor the selection.
3120          */
3121         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3122             (final_hop && !mr_forwarding_allowed))
3123                 send_case |= NMR_DST;
3124         else
3125                 send_case |= MR_DST;
3126
3127         if (lnet_msg_is_response(msg))
3128                 send_case |= SND_RESP;
3129
3130         /* assign parameters to the send_data */
3131         if (rtr_nid)
3132                 send_data.sd_rtr_nid = *rtr_nid;
3133         else
3134                 send_data.sd_rtr_nid = LNET_ANY_NID;
3135         if (src_nid)
3136                 send_data.sd_src_nid = *src_nid;
3137         else
3138                 send_data.sd_src_nid = LNET_ANY_NID;
3139         send_data.sd_dst_nid = *dst_nid;
3140         send_data.sd_best_lpni = lpni;
3141         /*
3142          * keep a pointer to the final destination in case we're going to
3143          * route, so we'll need to access it later
3144          */
3145         send_data.sd_final_dst_lpni = lpni;
3146         send_data.sd_peer = peer;
3147         send_data.sd_md_cpt = md_cpt;
3148         send_data.sd_send_case = send_case;
3149
3150         rc = lnet_handle_send_case_locked(&send_data);
3151
3152         /*
3153          * Update the local cpt since send_data.sd_cpt might've been
3154          * updated as a result of calling lnet_handle_send_case_locked().
3155          */
3156         cpt = send_data.sd_cpt;
3157
3158         if (rc == REPEAT_SEND)
3159                 goto again;
3160
3161         lnet_net_unlock(cpt);
3162
3163         return rc;
3164 }
3165
3166 int
3167 lnet_send(struct lnet_nid *src_nid, struct lnet_msg *msg,
3168           struct lnet_nid *rtr_nid)
3169 {
3170         struct lnet_nid *dst_nid = &msg->msg_target.nid;
3171         int rc;
3172
3173         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3174         LASSERT(msg->msg_txpeer == NULL);
3175         LASSERT(msg->msg_txni == NULL);
3176         LASSERT(!msg->msg_sending);
3177         LASSERT(!msg->msg_target_is_router);
3178         LASSERT(!msg->msg_receiving);
3179
3180         msg->msg_sending = 1;
3181
3182         LASSERT(!msg->msg_tx_committed);
3183
3184         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3185         if (rc < 0) {
3186                 if (rc == -EHOSTUNREACH)
3187                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3188                 else
3189                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3190                 return rc;
3191         }
3192
3193         if (rc == LNET_CREDIT_OK)
3194                 lnet_ni_send(msg->msg_txni, msg);
3195
3196         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3197         return 0;
3198 }
3199
3200 enum lnet_mt_event_type {
3201         MT_TYPE_LOCAL_NI = 0,
3202         MT_TYPE_PEER_NI
3203 };
3204
3205 struct lnet_mt_event_info {
3206         enum lnet_mt_event_type mt_type;
3207         struct lnet_nid mt_nid;
3208 };
3209
3210 /* called with res_lock held */
3211 void
3212 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3213 {
3214         struct lnet_rsp_tracker *rspt;
3215
3216         /*
3217          * msg has a refcount on the MD so the MD is not going away.
3218          * The rspt queue for the cpt is protected by
3219          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3220          */
3221         if (!md->md_rspt_ptr)
3222                 return;
3223
3224         rspt = md->md_rspt_ptr;
3225
3226         /* debug code */
3227         LASSERT(rspt->rspt_cpt == cpt);
3228
3229         md->md_rspt_ptr = NULL;
3230
3231         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3232                 /*
3233                  * The monitor thread has invalidated this handle because the
3234                  * response timed out, but it failed to lookup the MD. That
3235                  * means this response tracker is on the zombie list. We can
3236                  * safely remove it under the resource lock (held by caller) and
3237                  * free the response tracker block.
3238                  */
3239                 list_del(&rspt->rspt_on_list);
3240                 lnet_rspt_free(rspt, cpt);
3241         } else {
3242                 /*
3243                  * invalidate the handle to indicate that a response has been
3244                  * received, which will then lead the monitor thread to clean up
3245                  * the rspt block.
3246                  */
3247                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3248         }
3249 }
3250
3251 void
3252 lnet_clean_zombie_rstqs(void)
3253 {
3254         struct lnet_rsp_tracker *rspt, *tmp;
3255         int i;
3256
3257         cfs_cpt_for_each(i, lnet_cpt_table()) {
3258                 list_for_each_entry_safe(rspt, tmp,
3259                                          the_lnet.ln_mt_zombie_rstqs[i],
3260                                          rspt_on_list) {
3261                         list_del(&rspt->rspt_on_list);
3262                         lnet_rspt_free(rspt, i);
3263                 }
3264         }
3265
3266         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3267 }
3268
3269 static void
3270 lnet_finalize_expired_responses(void)
3271 {
3272         struct lnet_libmd *md;
3273         struct lnet_rsp_tracker *rspt, *tmp;
3274         ktime_t now;
3275         int i;
3276
3277         if (the_lnet.ln_mt_rstq == NULL)
3278                 return;
3279
3280         cfs_cpt_for_each(i, lnet_cpt_table()) {
3281                 LIST_HEAD(local_queue);
3282
3283                 lnet_net_lock(i);
3284                 if (!the_lnet.ln_mt_rstq[i]) {
3285                         lnet_net_unlock(i);
3286                         continue;
3287                 }
3288                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3289                 lnet_net_unlock(i);
3290
3291                 now = ktime_get();
3292
3293                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3294                         /*
3295                          * The rspt mdh will be invalidated when a response
3296                          * is received or whenever we want to discard the
3297                          * block the monitor thread will walk the queue
3298                          * and clean up any rsts with an invalid mdh.
3299                          * The monitor thread will walk the queue until
3300                          * the first unexpired rspt block. This means that
3301                          * some rspt blocks which received their
3302                          * corresponding responses will linger in the
3303                          * queue until they are cleaned up eventually.
3304                          */
3305                         lnet_res_lock(i);
3306                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3307                                 lnet_res_unlock(i);
3308                                 list_del(&rspt->rspt_on_list);
3309                                 lnet_rspt_free(rspt, i);
3310                                 continue;
3311                         }
3312
3313                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3314                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3315                                 struct lnet_peer_ni *lpni;
3316                                 struct lnet_nid nid;
3317
3318                                 md = lnet_handle2md(&rspt->rspt_mdh);
3319                                 if (!md) {
3320                                         /* MD has been queued for unlink, but
3321                                          * rspt hasn't been detached (Note we've
3322                                          * checked above that the rspt_mdh is
3323                                          * valid). Since we cannot lookup the MD
3324                                          * we're unable to detach the rspt
3325                                          * ourselves. Thus, move the rspt to the
3326                                          * zombie list where we'll wait for
3327                                          * either:
3328                                          *   1. The remaining operations on the
3329                                          *   MD to complete. In this case the
3330                                          *   final operation will result in
3331                                          *   lnet_msg_detach_md()->
3332                                          *   lnet_detach_rsp_tracker() where
3333                                          *   we will clean up this response
3334                                          *   tracker.
3335                                          *   2. LNet to shutdown. In this case
3336                                          *   we'll wait until after all LND Nets
3337                                          *   have shutdown and then we can
3338                                          *   safely free any remaining response
3339                                          *   tracker blocks on the zombie list.
3340                                          * Note: We need to hold the resource
3341                                          * lock when adding to the zombie list
3342                                          * because we may have concurrent access
3343                                          * with lnet_detach_rsp_tracker().
3344                                          */
3345                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
3346                                         list_move(&rspt->rspt_on_list,
3347                                                   the_lnet.ln_mt_zombie_rstqs[i]);
3348                                         lnet_res_unlock(i);
3349                                         continue;
3350                                 }
3351                                 LASSERT(md->md_rspt_ptr == rspt);
3352                                 md->md_rspt_ptr = NULL;
3353                                 lnet_res_unlock(i);
3354
3355                                 LNetMDUnlink(rspt->rspt_mdh);
3356
3357                                 nid = rspt->rspt_next_hop_nid;
3358
3359                                 list_del(&rspt->rspt_on_list);
3360                                 lnet_rspt_free(rspt, i);
3361
3362                                 /* If we're shutting down we just want to clean
3363                                  * up the rspt blocks
3364                                  */
3365                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3366                                         continue;
3367
3368                                 lnet_net_lock(i);
3369                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3370                                 lnet_net_unlock(i);
3371
3372                                 CDEBUG(D_NET,
3373                                        "Response timeout: md = %p: nid = %s\n",
3374                                        md, libcfs_nidstr(&nid));
3375
3376                                 /*
3377                                  * If there is a timeout on the response
3378                                  * from the next hop decrement its health
3379                                  * value so that we don't use it
3380                                  */
3381                                 lnet_net_lock(0);
3382                                 lpni = lnet_peer_ni_find_locked(&nid);
3383                                 if (lpni) {
3384                                         lnet_handle_remote_failure_locked(lpni);
3385                                         lnet_peer_ni_decref_locked(lpni);
3386                                 }
3387                                 lnet_net_unlock(0);
3388                         } else {
3389                                 lnet_res_unlock(i);
3390                                 break;
3391                         }
3392                 }
3393
3394                 if (!list_empty(&local_queue)) {
3395                         lnet_net_lock(i);
3396                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3397                         lnet_net_unlock(i);
3398                 }
3399         }
3400 }
3401
3402 static void
3403 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3404 {
3405         struct lnet_msg *msg;
3406
3407         while (!list_empty(resendq)) {
3408                 struct lnet_peer_ni *lpni;
3409
3410                 msg = list_first_entry(resendq, struct lnet_msg,
3411                                        msg_list);
3412
3413                 list_del_init(&msg->msg_list);
3414
3415                 lpni = lnet_peer_ni_find_locked(&msg->msg_hdr.dest_nid);
3416                 if (!lpni) {
3417                         lnet_net_unlock(cpt);
3418                         CERROR("Expected that a peer is already created for %s\n",
3419                                libcfs_nidstr(&msg->msg_hdr.dest_nid));
3420                         msg->msg_no_resend = true;
3421                         lnet_finalize(msg, -EFAULT);
3422                         lnet_net_lock(cpt);
3423                 } else {
3424                         int rc;
3425
3426                         lnet_peer_ni_decref_locked(lpni);
3427
3428                         lnet_net_unlock(cpt);
3429                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3430                                libcfs_nidstr(&msg->msg_src_nid_param),
3431                                libcfs_idstr(&msg->msg_target),
3432                                lnet_msgtyp2str(msg->msg_type),
3433                                msg->msg_recovery,
3434                                msg->msg_retry_count);
3435                         rc = lnet_send(&msg->msg_src_nid_param, msg,
3436                                        &msg->msg_rtr_nid_param);
3437                         if (rc) {
3438                                 CERROR("Error sending %s to %s: %d\n",
3439                                        lnet_msgtyp2str(msg->msg_type),
3440                                        libcfs_idstr(&msg->msg_target), rc);
3441                                 msg->msg_no_resend = true;
3442                                 lnet_finalize(msg, rc);
3443                         }
3444                         lnet_net_lock(cpt);
3445                         if (!rc)
3446                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3447                 }
3448         }
3449 }
3450
3451 static void
3452 lnet_resend_pending_msgs(void)
3453 {
3454         int i;
3455
3456         cfs_cpt_for_each(i, lnet_cpt_table()) {
3457                 lnet_net_lock(i);
3458                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3459                 lnet_net_unlock(i);
3460         }
3461 }
3462
3463 /* called with cpt and ni_lock held */
3464 static void
3465 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3466 {
3467         struct lnet_handle_md recovery_mdh;
3468
3469         LNetInvalidateMDHandle(&recovery_mdh);
3470
3471         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3472             force) {
3473                 recovery_mdh = ni->ni_ping_mdh;
3474                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3475         }
3476         lnet_ni_unlock(ni);
3477         lnet_net_unlock(cpt);
3478         if (!LNetMDHandleIsInvalid(recovery_mdh))
3479                 LNetMDUnlink(recovery_mdh);
3480         lnet_net_lock(cpt);
3481         lnet_ni_lock(ni);
3482 }
3483
3484 /* Returns the total number of local NIs in recovery.
3485  * Records up to @arrsz of the associated NIDs in the @nidarr array
3486  */
3487 static int
3488 lnet_recover_local_nis(struct lnet_nid *nidarr, unsigned int arrsz)
3489 {
3490         struct lnet_mt_event_info *ev_info;
3491         LIST_HEAD(processed_list);
3492         LIST_HEAD(local_queue);
3493         struct lnet_handle_md mdh;
3494         struct lnet_ni *tmp;
3495         struct lnet_ni *ni;
3496         struct lnet_nid nid;
3497         int healthv;
3498         int rc;
3499         time64_t now;
3500         unsigned int nnis = 0;
3501
3502         /*
3503          * splice the recovery queue on a local queue. We will iterate
3504          * through the local queue and update it as needed. Once we're
3505          * done with the traversal, we'll splice the local queue back on
3506          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3507          * will be traversed in the next iteration.
3508          */
3509         lnet_net_lock(0);
3510         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3511                          &local_queue);
3512         lnet_net_unlock(0);
3513
3514         now = ktime_get_seconds();
3515
3516         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3517                 /*
3518                  * if an NI is being deleted or it is now healthy, there
3519                  * is no need to keep it around in the recovery queue.
3520                  * The monitor thread is the only thread responsible for
3521                  * removing the NI from the recovery queue.
3522                  * Multiple threads can be adding NIs to the recovery
3523                  * queue.
3524                  */
3525                 healthv = atomic_read(&ni->ni_healthv);
3526
3527                 lnet_net_lock(0);
3528                 lnet_ni_lock(ni);
3529                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3530                     healthv == LNET_MAX_HEALTH_VALUE) {
3531                         list_del_init(&ni->ni_recovery);
3532                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3533                         lnet_ni_unlock(ni);
3534                         lnet_ni_decref_locked(ni, 0);
3535                         lnet_net_unlock(0);
3536                         continue;
3537                 }
3538
3539                 if (nnis < arrsz)
3540                         nidarr[nnis] = ni->ni_nid;
3541                 nnis++;
3542
3543                 /*
3544                  * if the local NI failed recovery we must unlink the md.
3545                  * But we want to keep the local_ni on the recovery queue
3546                  * so we can continue the attempts to recover it.
3547                  */
3548                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3549                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3550                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3551                 }
3552
3553
3554                 lnet_ni_unlock(ni);
3555
3556                 if (now < ni->ni_next_ping) {
3557                         lnet_net_unlock(0);
3558                         continue;
3559                 }
3560
3561                 lnet_net_unlock(0);
3562
3563                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3564                        libcfs_nidstr(&ni->ni_nid));
3565
3566                 lnet_ni_lock(ni);
3567                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3568                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3569                         lnet_ni_unlock(ni);
3570
3571                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3572                         if (!ev_info) {
3573                                 CERROR("out of memory. Can't recover %s\n",
3574                                        libcfs_nidstr(&ni->ni_nid));
3575                                 lnet_ni_lock(ni);
3576                                 ni->ni_recovery_state &=
3577                                   ~LNET_NI_RECOVERY_PENDING;
3578                                 lnet_ni_unlock(ni);
3579                                 continue;
3580                         }
3581
3582                         mdh = ni->ni_ping_mdh;
3583                         /*
3584                          * Invalidate the ni mdh in case it's deleted.
3585                          * We'll unlink the mdh in this case below.
3586                          */
3587                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3588                         nid = ni->ni_nid;
3589
3590                         /*
3591                          * remove the NI from the local queue and drop the
3592                          * reference count to it while we're recovering
3593                          * it. The reason for that, is that the NI could
3594                          * be deleted, and the way the code is structured
3595                          * is if we don't drop the NI, then the deletion
3596                          * code will enter a loop waiting for the
3597                          * reference count to be removed while holding the
3598                          * ln_mutex_lock(). When we look up the peer to
3599                          * send to in lnet_select_pathway() we will try to
3600                          * lock the ln_mutex_lock() as well, leading to
3601                          * a deadlock. By dropping the refcount and
3602                          * removing it from the list, we allow for the NI
3603                          * to be removed, then we use the cached NID to
3604                          * look it up again. If it's gone, then we just
3605                          * continue examining the rest of the queue.
3606                          */
3607                         lnet_net_lock(0);
3608                         list_del_init(&ni->ni_recovery);
3609                         lnet_ni_decref_locked(ni, 0);
3610                         lnet_net_unlock(0);
3611
3612                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3613                         ev_info->mt_nid = nid;
3614                         rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3615                                             ev_info, the_lnet.ln_mt_handler,
3616                                             true);
3617                         /* lookup the nid again */
3618                         lnet_net_lock(0);
3619                         ni = lnet_nid_to_ni_locked(&nid, 0);
3620                         if (!ni) {
3621                                 /*
3622                                  * the NI has been deleted when we dropped
3623                                  * the ref count
3624                                  */
3625                                 lnet_net_unlock(0);
3626                                 LNetMDUnlink(mdh);
3627                                 continue;
3628                         }
3629                         ni->ni_ping_count++;
3630
3631                         ni->ni_ping_mdh = mdh;
3632                         lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3633                                                         now);
3634
3635                         if (rc) {
3636                                 lnet_ni_lock(ni);
3637                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3638                                 lnet_ni_unlock(ni);
3639                         }
3640                         lnet_net_unlock(0);
3641                 } else
3642                         lnet_ni_unlock(ni);
3643         }
3644
3645         /*
3646          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3647          * reexamined in the next iteration.
3648          */
3649         list_splice_init(&processed_list, &local_queue);
3650         lnet_net_lock(0);
3651         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3652         lnet_net_unlock(0);
3653
3654         return nnis;
3655 }
3656
3657 static int
3658 lnet_resendqs_create(void)
3659 {
3660         struct list_head **resendqs;
3661         resendqs = lnet_create_array_of_queues();
3662
3663         if (!resendqs)
3664                 return -ENOMEM;
3665
3666         lnet_net_lock(LNET_LOCK_EX);
3667         the_lnet.ln_mt_resendqs = resendqs;
3668         lnet_net_unlock(LNET_LOCK_EX);
3669
3670         return 0;
3671 }
3672
3673 static void
3674 lnet_clean_local_ni_recoveryq(void)
3675 {
3676         struct lnet_ni *ni;
3677
3678         /* This is only called when the monitor thread has stopped */
3679         lnet_net_lock(0);
3680
3681         while ((ni = list_first_entry_or_null(&the_lnet.ln_mt_localNIRecovq,
3682                                               struct lnet_ni,
3683                                               ni_recovery)) != NULL) {
3684                 list_del_init(&ni->ni_recovery);
3685                 lnet_ni_lock(ni);
3686                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3687                 lnet_ni_unlock(ni);
3688                 lnet_ni_decref_locked(ni, 0);
3689         }
3690
3691         lnet_net_unlock(0);
3692 }
3693
3694 static void
3695 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3696                                      bool force)
3697 {
3698         struct lnet_handle_md recovery_mdh;
3699
3700         LNetInvalidateMDHandle(&recovery_mdh);
3701
3702         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3703                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3704                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3705         }
3706         spin_unlock(&lpni->lpni_lock);
3707         lnet_net_unlock(cpt);
3708         if (!LNetMDHandleIsInvalid(recovery_mdh))
3709                 LNetMDUnlink(recovery_mdh);
3710         lnet_net_lock(cpt);
3711         spin_lock(&lpni->lpni_lock);
3712 }
3713
3714 static void
3715 lnet_clean_peer_ni_recoveryq(void)
3716 {
3717         struct lnet_peer_ni *lpni, *tmp;
3718
3719         lnet_net_lock(LNET_LOCK_EX);
3720
3721         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3722                                  lpni_recovery) {
3723                 list_del_init(&lpni->lpni_recovery);
3724                 spin_lock(&lpni->lpni_lock);
3725                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3726                 spin_unlock(&lpni->lpni_lock);
3727                 lnet_peer_ni_decref_locked(lpni);
3728         }
3729
3730         lnet_net_unlock(LNET_LOCK_EX);
3731 }
3732
3733 static void
3734 lnet_clean_resendqs(void)
3735 {
3736         struct lnet_msg *msg, *tmp;
3737         LIST_HEAD(msgs);
3738         int i;
3739
3740         cfs_cpt_for_each(i, lnet_cpt_table()) {
3741                 lnet_net_lock(i);
3742                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3743                 lnet_net_unlock(i);
3744                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3745                         list_del_init(&msg->msg_list);
3746                         msg->msg_no_resend = true;
3747                         lnet_finalize(msg, -ESHUTDOWN);
3748                 }
3749         }
3750
3751         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3752 }
3753
3754 /* Returns the total number of peer NIs in recovery.
3755  * Records up to @arrsz of the associated NIDs in the @nidarr array
3756  */
3757 static unsigned int
3758 lnet_recover_peer_nis(struct lnet_nid *nidarr, unsigned int arrsz)
3759 {
3760         struct lnet_mt_event_info *ev_info;
3761         LIST_HEAD(processed_list);
3762         LIST_HEAD(local_queue);
3763         struct lnet_handle_md mdh;
3764         struct lnet_peer_ni *lpni;
3765         struct lnet_peer_ni *tmp;
3766         struct lnet_nid nid;
3767         int healthv;
3768         int rc;
3769         time64_t now;
3770         unsigned int nlpnis = 0;
3771
3772         /*
3773          * Always use cpt 0 for locking across all interactions with
3774          * ln_mt_peerNIRecovq
3775          */
3776         lnet_net_lock(0);
3777         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3778                          &local_queue);
3779         lnet_net_unlock(0);
3780
3781         now = ktime_get_seconds();
3782
3783         list_for_each_entry_safe(lpni, tmp, &local_queue,
3784                                  lpni_recovery) {
3785                 /*
3786                  * The same protection strategy is used here as is in the
3787                  * local recovery case.
3788                  */
3789                 lnet_net_lock(0);
3790                 healthv = atomic_read(&lpni->lpni_healthv);
3791                 spin_lock(&lpni->lpni_lock);
3792                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3793                     healthv == LNET_MAX_HEALTH_VALUE) {
3794                         list_del_init(&lpni->lpni_recovery);
3795                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3796                         spin_unlock(&lpni->lpni_lock);
3797                         lnet_peer_ni_decref_locked(lpni);
3798                         lnet_net_unlock(0);
3799                         continue;
3800                 }
3801
3802                 if (nlpnis < arrsz)
3803                         nidarr[nlpnis] = lpni->lpni_nid;
3804                 nlpnis++;
3805
3806                 /*
3807                  * If the peer NI has failed recovery we must unlink the
3808                  * md. But we want to keep the peer ni on the recovery
3809                  * queue so we can try to continue recovering it
3810                  */
3811                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3812                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3813                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3814                 }
3815
3816                 spin_unlock(&lpni->lpni_lock);
3817
3818                 if (now < lpni->lpni_next_ping) {
3819                         lnet_net_unlock(0);
3820                         continue;
3821                 }
3822
3823                 lnet_net_unlock(0);
3824
3825                 /*
3826                  * NOTE: we're racing with peer deletion from user space.
3827                  * It's possible that a peer is deleted after we check its
3828                  * state. In this case the recovery can create a new peer
3829                  */
3830                 spin_lock(&lpni->lpni_lock);
3831                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3832                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3833                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3834                         spin_unlock(&lpni->lpni_lock);
3835
3836                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3837                         if (!ev_info) {
3838                                 CERROR("out of memory. Can't recover %s\n",
3839                                        libcfs_nidstr(&lpni->lpni_nid));
3840                                 spin_lock(&lpni->lpni_lock);
3841                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3842                                 spin_unlock(&lpni->lpni_lock);
3843                                 continue;
3844                         }
3845
3846                         /* look at the comments in lnet_recover_local_nis() */
3847                         mdh = lpni->lpni_recovery_ping_mdh;
3848                         nid = lpni->lpni_nid;
3849                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3850                         lnet_net_lock(0);
3851                         list_del_init(&lpni->lpni_recovery);
3852                         lnet_peer_ni_decref_locked(lpni);
3853                         lnet_net_unlock(0);
3854
3855                         ev_info->mt_type = MT_TYPE_PEER_NI;
3856                         ev_info->mt_nid = nid;
3857                         rc = lnet_send_ping(&nid, &mdh, LNET_PING_INFO_MIN_SIZE,
3858                                             ev_info, the_lnet.ln_mt_handler,
3859                                             true);
3860                         lnet_net_lock(0);
3861                         /*
3862                          * lnet_peer_ni_find_locked() grabs a refcount for
3863                          * us. No need to take it explicitly.
3864                          */
3865                         lpni = lnet_peer_ni_find_locked(&nid);
3866                         if (!lpni) {
3867                                 lnet_net_unlock(0);
3868                                 LNetMDUnlink(mdh);
3869                                 continue;
3870                         }
3871
3872                         lpni->lpni_ping_count++;
3873
3874                         lpni->lpni_recovery_ping_mdh = mdh;
3875
3876                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
3877                                                              &processed_list,
3878                                                              now);
3879                         if (rc) {
3880                                 spin_lock(&lpni->lpni_lock);
3881                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3882                                 spin_unlock(&lpni->lpni_lock);
3883                         }
3884
3885                         /* Drop the ref taken by lnet_peer_ni_find_locked() */
3886                         lnet_peer_ni_decref_locked(lpni);
3887                         lnet_net_unlock(0);
3888                 } else {
3889                         spin_unlock(&lpni->lpni_lock);
3890                 }
3891         }
3892
3893         list_splice_init(&processed_list, &local_queue);
3894         lnet_net_lock(0);
3895         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3896         lnet_net_unlock(0);
3897
3898         return nlpnis;
3899 }
3900
3901 #define LNET_MAX_NNIDS 20
3902 /* @nids is array of nids that are in recovery. It has max size of
3903  * LNET_MAX_NNIDS.
3904  * @nnids is the total number of nids that are in recovery. It can be
3905  * larger than LNET_MAX_NNIDS.
3906  * @local tells us whether these are local or peer NIs in recovery.
3907  */
3908 static void
3909 lnet_print_recovery_list(struct lnet_nid *nids, unsigned int nnids,
3910                          bool local)
3911 {
3912         static bool printed;
3913         char *buf = NULL;
3914         char *tmp;
3915         int i;
3916         unsigned int arrsz;
3917         unsigned int bufsz;
3918
3919         if (!nnids)
3920                 return;
3921
3922         arrsz = nnids < LNET_MAX_NNIDS ? nnids : LNET_MAX_NNIDS;
3923
3924         /* Printing arrsz NIDs, each has max size LNET_NIDSTR_SIZE, a comma
3925          * and space for each nid after the first (2 * (arrsz - 1)),
3926          * + 1 for terminating null byte
3927          */
3928         bufsz = (arrsz * LNET_NIDSTR_SIZE) + (2 * (arrsz - 1)) + 1;
3929         LIBCFS_ALLOC(buf, bufsz);
3930         if (!buf) {
3931                 LCONSOLE(D_INFO, "%u %s NIs in recovery\n",
3932                          nnids, local ? "local" : "peer");
3933                 return;
3934         }
3935
3936         tmp = buf;
3937         tmp += sprintf(tmp, "%s", libcfs_nidstr(&nids[0]));
3938         for (i = 1; i < arrsz; i++)
3939                 tmp += sprintf(tmp, ", %s", libcfs_nidstr(&nids[i]));
3940
3941         /* LCONSOLE() used to avoid rate limiting when we have both local
3942          * and peer NIs in recovery
3943          */
3944         LCONSOLE(D_INFO, "%u %s NIs in recovery (showing %u): %s\n",
3945                  nnids, local ? "local" : "peer", arrsz, buf);
3946
3947         LIBCFS_FREE(buf, bufsz);
3948
3949         if (!printed && nnids > LNET_MAX_NNIDS) {
3950                 LCONSOLE(D_INFO, "See full list with 'lnetctl debug recovery -(p|l)'\n");
3951                 printed = true;
3952         }
3953 }
3954
3955 static void
3956 lnet_health_update_console(struct lnet_nid *lnids, unsigned int nnis,
3957                            struct lnet_nid *rnids, unsigned int nlpnis,
3958                            time64_t now)
3959 {
3960         static time64_t next_ni_update;
3961         static time64_t next_lpni_update;
3962         static time64_t next_msg_update;
3963         static unsigned int num_ni_updates;
3964         static unsigned int num_lpni_updates;
3965         static unsigned int num_msg_updates = 1;
3966         int late_count;
3967
3968         if (now >= next_ni_update) {
3969                 if (nnis) {
3970                         lnet_print_recovery_list(lnids, nnis, true);
3971                         if (num_ni_updates < 5)
3972                                 num_ni_updates++;
3973                         next_ni_update = now + (60 * num_ni_updates);
3974                 } else {
3975                         next_ni_update = 0;
3976                         num_ni_updates = 0;
3977                 }
3978         }
3979
3980
3981         if (now >= next_lpni_update) {
3982                 if (nlpnis) {
3983                         lnet_print_recovery_list(rnids, nlpnis, false);
3984                         if (num_lpni_updates < 5)
3985                                 num_lpni_updates++;
3986                         next_lpni_update = now + (60 * num_lpni_updates);
3987                 } else {
3988                         next_lpni_update = 0;
3989                         num_lpni_updates = 0;
3990                 }
3991         }
3992
3993         /* Let late_count accumulate for 60 seconds */
3994         if (unlikely(!next_msg_update))
3995                 next_msg_update = now + 60;
3996
3997         if (now >= next_msg_update) {
3998                 late_count = atomic_read(&the_lnet.ln_late_msg_count);
3999
4000                 if (late_count) {
4001                         s64 avg = atomic64_xchg(&the_lnet.ln_late_msg_nsecs, 0) /
4002                                   atomic_xchg(&the_lnet.ln_late_msg_count, 0);
4003
4004                         if (avg > NSEC_PER_SEC) {
4005                                 unsigned int avg_msec;
4006
4007                                 avg_msec = do_div(avg, NSEC_PER_SEC) /
4008                                                 NSEC_PER_MSEC;
4009                                 LCONSOLE_INFO("%u messages in past %us over their deadline by avg %lld.%03us\n",
4010                                               late_count,
4011                                               (60 * num_msg_updates), avg,
4012                                               avg_msec);
4013
4014                                 if (num_msg_updates < 5)
4015                                         num_msg_updates++;
4016                                 next_msg_update = now + (60 * num_msg_updates);
4017                         }
4018                 } else {
4019                         next_msg_update = now + 60;
4020                         num_msg_updates = 1;
4021                 }
4022         }
4023 }
4024
4025 static int
4026 lnet_monitor_thread(void *arg)
4027 {
4028         time64_t rsp_timeout = 0;
4029         time64_t now;
4030         unsigned int nnis;
4031         unsigned int nlpnis;
4032         struct lnet_nid local_nids[LNET_MAX_NNIDS];
4033         struct lnet_nid peer_nids[LNET_MAX_NNIDS];
4034
4035         wait_for_completion(&the_lnet.ln_started);
4036
4037         /*
4038          * The monitor thread takes care of the following:
4039          *  1. Checks the aliveness of routers
4040          *  2. Checks if there are messages on the resend queue to resend
4041          *     them.
4042          *  3. Check if there are any NIs on the local recovery queue and
4043          *     pings them
4044          *  4. Checks if there are any NIs on the remote recovery queue
4045          *     and pings them.
4046          */
4047         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
4048                 now = ktime_get_real_seconds();
4049
4050                 if (lnet_router_checker_active())
4051                         lnet_check_routers();
4052
4053                 lnet_resend_pending_msgs();
4054
4055                 if (now >= rsp_timeout) {
4056                         lnet_finalize_expired_responses();
4057                         rsp_timeout = now + (lnet_transaction_timeout / 2);
4058                 }
4059
4060                 nnis = lnet_recover_local_nis(local_nids, LNET_MAX_NNIDS);
4061                 nlpnis = lnet_recover_peer_nis(peer_nids, LNET_MAX_NNIDS);
4062                 lnet_health_update_console(local_nids, nnis, peer_nids, nlpnis,
4063                                            now);
4064
4065                 /*
4066                  * TODO do we need to check if we should sleep without
4067                  * timeout?  Technically, an active system will always
4068                  * have messages in flight so this check will always
4069                  * evaluate to false. And on an idle system do we care
4070                  * if we wake up every 1 second? Although, we've seen
4071                  * cases where we get a complaint that an idle thread
4072                  * is waking up unnecessarily.
4073                  */
4074                 wait_for_completion_interruptible_timeout(
4075                         &the_lnet.ln_mt_wait_complete,
4076                         cfs_time_seconds(1));
4077                 /* Must re-init the completion before testing anything,
4078                  * including ln_mt_state.
4079                  */
4080                 reinit_completion(&the_lnet.ln_mt_wait_complete);
4081         }
4082
4083         /* Shutting down */
4084         lnet_net_lock(LNET_LOCK_EX);
4085         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4086         lnet_net_unlock(LNET_LOCK_EX);
4087
4088         /* signal that the monitor thread is exiting */
4089         up(&the_lnet.ln_mt_signal);
4090
4091         return 0;
4092 }
4093
4094 /*
4095  * lnet_send_ping
4096  * Sends a ping.
4097  * Returns == 0 if success
4098  * Returns > 0 if LNetMDBind or prior fails
4099  * Returns < 0 if LNetGet fails
4100  */
4101 int
4102 lnet_send_ping(struct lnet_nid *dest_nid,
4103                struct lnet_handle_md *mdh, int bytes,
4104                void *user_data, lnet_handler_t handler, bool recovery)
4105 {
4106         struct lnet_md md = { NULL };
4107         struct lnet_processid id;
4108         struct lnet_ping_buffer *pbuf;
4109         int rc;
4110
4111         if (LNET_NID_IS_ANY(dest_nid)) {
4112                 rc = -EHOSTUNREACH;
4113                 goto fail_error;
4114         }
4115
4116         pbuf = lnet_ping_buffer_alloc(bytes, GFP_NOFS);
4117         if (!pbuf) {
4118                 rc = ENOMEM;
4119                 goto fail_error;
4120         }
4121
4122         /* initialize md content */
4123         md.start     = &pbuf->pb_info;
4124         md.length    = bytes;
4125         md.threshold = 2; /* GET/REPLY */
4126         md.max_size  = 0;
4127         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
4128         md.user_ptr  = user_data;
4129         md.handler   = handler;
4130
4131         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
4132         if (rc) {
4133                 lnet_ping_buffer_decref(pbuf);
4134                 CERROR("Can't bind MD: %d\n", rc);
4135                 rc = -rc; /* change the rc to positive */
4136                 goto fail_error;
4137         }
4138         id.pid = LNET_PID_LUSTRE;
4139         id.nid = *dest_nid;
4140
4141         rc = LNetGet(NULL, *mdh, &id,
4142                      LNET_RESERVED_PORTAL,
4143                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
4144
4145         if (rc)
4146                 goto fail_unlink_md;
4147
4148         return 0;
4149
4150 fail_unlink_md:
4151         LNetMDUnlink(*mdh);
4152         LNetInvalidateMDHandle(mdh);
4153 fail_error:
4154         return rc;
4155 }
4156
4157 static void
4158 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
4159                            int status, bool send, bool unlink_event)
4160 {
4161         struct lnet_nid *nid = &ev_info->mt_nid;
4162
4163         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
4164                 struct lnet_ni *ni;
4165
4166                 lnet_net_lock(0);
4167                 ni = lnet_nid_to_ni_locked(nid, 0);
4168                 if (!ni) {
4169                         lnet_net_unlock(0);
4170                         return;
4171                 }
4172                 lnet_ni_lock(ni);
4173                 if (!send || (send && status != 0))
4174                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
4175                 if (status)
4176                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
4177                 lnet_ni_unlock(ni);
4178                 lnet_net_unlock(0);
4179
4180                 if (status != 0) {
4181                         CDEBUG(D_NET, "local NI (%s) recovery failed with %d\n",
4182                                libcfs_nidstr(nid), status);
4183                         return;
4184                 }
4185                 /*
4186                  * need to increment healthv for the ni here, because in
4187                  * the lnet_finalize() path we don't have access to this
4188                  * NI. And in order to get access to it, we'll need to
4189                  * carry forward too much information.
4190                  * In the peer case, it'll naturally be incremented
4191                  */
4192                 if (!unlink_event)
4193                         lnet_inc_healthv(&ni->ni_healthv,
4194                                          lnet_health_sensitivity);
4195         } else {
4196                 struct lnet_peer_ni *lpni;
4197                 int cpt;
4198
4199                 cpt = lnet_net_lock_current();
4200                 lpni = lnet_peer_ni_find_locked(nid);
4201                 if (!lpni) {
4202                         lnet_net_unlock(cpt);
4203                         return;
4204                 }
4205                 spin_lock(&lpni->lpni_lock);
4206                 if (!send || (send && status != 0))
4207                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
4208                 if (status)
4209                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
4210                 spin_unlock(&lpni->lpni_lock);
4211                 lnet_peer_ni_decref_locked(lpni);
4212                 lnet_net_unlock(cpt);
4213
4214                 if (status != 0)
4215                         CDEBUG(D_NET, "peer NI (%s) recovery failed with %d\n",
4216                                libcfs_nidstr(nid), status);
4217         }
4218 }
4219
4220 void
4221 lnet_mt_event_handler(struct lnet_event *event)
4222 {
4223         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4224         struct lnet_ping_buffer *pbuf;
4225
4226         /* TODO: remove assert */
4227         LASSERT(event->type == LNET_EVENT_REPLY ||
4228                 event->type == LNET_EVENT_SEND ||
4229                 event->type == LNET_EVENT_UNLINK);
4230
4231         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4232                event->status);
4233
4234         switch (event->type) {
4235         case LNET_EVENT_UNLINK:
4236                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4237                        libcfs_nidstr(&ev_info->mt_nid));
4238                 fallthrough;
4239         case LNET_EVENT_REPLY:
4240                 lnet_handle_recovery_reply(ev_info, event->status, false,
4241                                            event->type == LNET_EVENT_UNLINK);
4242                 break;
4243         case LNET_EVENT_SEND:
4244                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4245                                libcfs_nidstr(&ev_info->mt_nid),
4246                                (event->status) ? "unsuccessfully" :
4247                                "successfully", event->status);
4248                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4249                 break;
4250         default:
4251                 CERROR("Unexpected event: %d\n", event->type);
4252                 break;
4253         }
4254         if (event->unlinked) {
4255                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4256                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4257                 lnet_ping_buffer_decref(pbuf);
4258         }
4259 }
4260
4261 static int
4262 lnet_rsp_tracker_create(void)
4263 {
4264         struct list_head **rstqs;
4265         rstqs = lnet_create_array_of_queues();
4266
4267         if (!rstqs)
4268                 return -ENOMEM;
4269
4270         the_lnet.ln_mt_rstq = rstqs;
4271
4272         return 0;
4273 }
4274
4275 static void
4276 lnet_rsp_tracker_clean(void)
4277 {
4278         lnet_finalize_expired_responses();
4279
4280         cfs_percpt_free(the_lnet.ln_mt_rstq);
4281         the_lnet.ln_mt_rstq = NULL;
4282 }
4283
4284 int lnet_monitor_thr_start(void)
4285 {
4286         int rc = 0;
4287         struct task_struct *task;
4288
4289         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4290                 return -EALREADY;
4291
4292         rc = lnet_resendqs_create();
4293         if (rc)
4294                 return rc;
4295
4296         rc = lnet_rsp_tracker_create();
4297         if (rc)
4298                 goto clean_queues;
4299
4300         sema_init(&the_lnet.ln_mt_signal, 0);
4301
4302         lnet_net_lock(LNET_LOCK_EX);
4303         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4304         lnet_net_unlock(LNET_LOCK_EX);
4305         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4306         if (IS_ERR(task)) {
4307                 rc = PTR_ERR(task);
4308                 CERROR("Can't start monitor thread: %d\n", rc);
4309                 goto clean_thread;
4310         }
4311
4312         return 0;
4313
4314 clean_thread:
4315         lnet_net_lock(LNET_LOCK_EX);
4316         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4317         lnet_net_unlock(LNET_LOCK_EX);
4318         /* block until event callback signals exit */
4319         down(&the_lnet.ln_mt_signal);
4320         /* clean up */
4321         lnet_net_lock(LNET_LOCK_EX);
4322         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4323         lnet_net_unlock(LNET_LOCK_EX);
4324         lnet_rsp_tracker_clean();
4325         lnet_clean_local_ni_recoveryq();
4326         lnet_clean_peer_ni_recoveryq();
4327         lnet_clean_resendqs();
4328         the_lnet.ln_mt_handler = NULL;
4329         return rc;
4330 clean_queues:
4331         lnet_rsp_tracker_clean();
4332         lnet_clean_local_ni_recoveryq();
4333         lnet_clean_peer_ni_recoveryq();
4334         lnet_clean_resendqs();
4335         return rc;
4336 }
4337
4338 void lnet_monitor_thr_stop(void)
4339 {
4340         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4341                 return;
4342
4343         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4344         lnet_net_lock(LNET_LOCK_EX);
4345         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4346         lnet_net_unlock(LNET_LOCK_EX);
4347
4348         /* tell the monitor thread that we're shutting down */
4349         complete(&the_lnet.ln_mt_wait_complete);
4350
4351         /* block until monitor thread signals that it's done */
4352         mutex_unlock(&the_lnet.ln_api_mutex);
4353         down(&the_lnet.ln_mt_signal);
4354         mutex_lock(&the_lnet.ln_api_mutex);
4355         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4356
4357         /* perform cleanup tasks */
4358         lnet_rsp_tracker_clean();
4359         lnet_clean_local_ni_recoveryq();
4360         lnet_clean_peer_ni_recoveryq();
4361         lnet_clean_resendqs();
4362 }
4363
4364 void
4365 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4366                   __u32 msg_type)
4367 {
4368         lnet_net_lock(cpt);
4369         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4370         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4371         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4372         lnet_net_unlock(cpt);
4373
4374         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4375 }
4376
4377 static void
4378 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4379 {
4380         struct lnet_hdr *hdr = &msg->msg_hdr;
4381
4382         if (msg->msg_wanted != 0)
4383                 lnet_setpayloadbuffer(msg);
4384
4385         lnet_build_msg_event(msg, LNET_EVENT_PUT);
4386
4387         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
4388          * it back into the ACK during lnet_finalize() */
4389         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4390                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4391
4392         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4393                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4394 }
4395
4396 static int
4397 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4398 {
4399         struct lnet_hdr         *hdr = &msg->msg_hdr;
4400         struct lnet_match_info  info;
4401         int                     rc;
4402         bool                    ready_delay;
4403
4404         /* Convert put fields to host byte order */
4405         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4406         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
4407         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
4408
4409         /* Primary peer NID. */
4410         info.mi_id.nid = msg->msg_initiator;
4411         info.mi_id.pid  = hdr->src_pid;
4412         info.mi_opc     = LNET_MD_OP_PUT;
4413         info.mi_portal  = hdr->msg.put.ptl_index;
4414         info.mi_rlength = hdr->payload_length;
4415         info.mi_roffset = hdr->msg.put.offset;
4416         info.mi_mbits   = hdr->msg.put.match_bits;
4417         info.mi_cpt     = lnet_nid2cpt(&msg->msg_initiator, ni);
4418
4419         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4420         ready_delay = msg->msg_rx_ready_delay;
4421
4422  again:
4423         rc = lnet_ptl_match_md(&info, msg);
4424         switch (rc) {
4425         default:
4426                 LBUG();
4427
4428         case LNET_MATCHMD_OK:
4429                 lnet_recv_put(ni, msg);
4430                 return 0;
4431
4432         case LNET_MATCHMD_NONE:
4433                 if (ready_delay)
4434                         /* no eager_recv or has already called it, should
4435                          * have been attached on delayed list */
4436                         return 0;
4437
4438                 rc = lnet_ni_eager_recv(ni, msg);
4439                 if (rc == 0) {
4440                         ready_delay = true;
4441                         goto again;
4442                 }
4443                 fallthrough;
4444
4445         case LNET_MATCHMD_DROP:
4446                 CNETERR("Dropping PUT from %s portal %d match %llu"
4447                         " offset %d length %d: %d\n",
4448                         libcfs_idstr(&info.mi_id), info.mi_portal,
4449                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4450
4451                 return -ENOENT; /* -ve: OK but no match */
4452         }
4453 }
4454
4455 static int
4456 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4457 {
4458         struct lnet_match_info info;
4459         struct lnet_hdr *hdr = &msg->msg_hdr;
4460         struct lnet_processid source_id;
4461         struct lnet_handle_wire reply_wmd;
4462         int rc;
4463
4464         /* Convert get fields to host byte order */
4465         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
4466         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
4467         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
4468         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
4469
4470         source_id.nid = hdr->src_nid;
4471         source_id.pid = hdr->src_pid;
4472         /* Primary peer NID */
4473         info.mi_id.nid  = msg->msg_initiator;
4474         info.mi_id.pid  = hdr->src_pid;
4475         info.mi_opc     = LNET_MD_OP_GET;
4476         info.mi_portal  = hdr->msg.get.ptl_index;
4477         info.mi_rlength = hdr->msg.get.sink_length;
4478         info.mi_roffset = hdr->msg.get.src_offset;
4479         info.mi_mbits   = hdr->msg.get.match_bits;
4480         info.mi_cpt     = lnet_nid2cpt(&msg->msg_initiator, ni);
4481
4482         rc = lnet_ptl_match_md(&info, msg);
4483         if (rc == LNET_MATCHMD_DROP) {
4484                 CNETERR("Dropping GET from %s portal %d match %llu"
4485                         " offset %d length %d\n",
4486                         libcfs_idstr(&info.mi_id), info.mi_portal,
4487                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
4488                 return -ENOENT; /* -ve: OK but no match */
4489         }
4490
4491         LASSERT(rc == LNET_MATCHMD_OK);
4492
4493         lnet_build_msg_event(msg, LNET_EVENT_GET);
4494
4495         reply_wmd = hdr->msg.get.return_wmd;
4496
4497         lnet_prep_send(msg, LNET_MSG_REPLY, &source_id,
4498                        msg->msg_offset, msg->msg_wanted);
4499
4500         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4501
4502         if (rdma_get) {
4503                 /* The LND completes the REPLY from her recv procedure */
4504                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4505                              msg->msg_offset, msg->msg_len, msg->msg_len);
4506                 return 0;
4507         }
4508
4509         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4510         msg->msg_receiving = 0;
4511
4512         rc = lnet_send(&ni->ni_nid, msg, &msg->msg_from);
4513         if (rc < 0) {
4514                 /* didn't get as far as lnet_ni_send() */
4515                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4516                        libcfs_nidstr(&ni->ni_nid),
4517                        libcfs_idstr(&info.mi_id), rc);
4518
4519                 lnet_finalize(msg, rc);
4520         }
4521
4522         return 0;
4523 }
4524
4525 static int
4526 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4527 {
4528         void *private = msg->msg_private;
4529         struct lnet_hdr *hdr = &msg->msg_hdr;
4530         struct lnet_processid src = {};
4531         struct lnet_libmd *md;
4532         unsigned int rlength;
4533         unsigned int mlength;
4534         int cpt;
4535
4536         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4537         lnet_res_lock(cpt);
4538
4539         src.nid = hdr->src_nid;
4540         src.pid = hdr->src_pid;
4541
4542         /* NB handles only looked up by creator (no flips) */
4543         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4544         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4545                 CNETERR("%s: Dropping REPLY from %s for %s "
4546                         "MD %#llx.%#llx\n",
4547                         libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4548                         (md == NULL) ? "invalid" : "inactive",
4549                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4550                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4551                 if (md != NULL && md->md_me != NULL)
4552                         CERROR("REPLY MD also attached to portal %d\n",
4553                                md->md_me->me_portal);
4554
4555                 lnet_res_unlock(cpt);
4556                 return -ENOENT; /* -ve: OK but no match */
4557         }
4558
4559         LASSERT(md->md_offset == 0);
4560
4561         rlength = hdr->payload_length;
4562         mlength = min(rlength, md->md_length);
4563
4564         if (mlength < rlength &&
4565             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4566                 CNETERR("%s: Dropping REPLY from %s length %d "
4567                         "for MD %#llx would overflow (%d)\n",
4568                         libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4569                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4570                         mlength);
4571                 lnet_res_unlock(cpt);
4572                 return -ENOENT; /* -ve: OK but no match */
4573         }
4574
4575         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4576                libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4577                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4578
4579         lnet_msg_attach_md(msg, md, 0, mlength);
4580
4581         if (mlength != 0)
4582                 lnet_setpayloadbuffer(msg);
4583
4584         lnet_res_unlock(cpt);
4585
4586         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4587
4588         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4589         return 0;
4590 }
4591
4592 static int
4593 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4594 {
4595         struct lnet_hdr *hdr = &msg->msg_hdr;
4596         struct lnet_processid src = {};
4597         struct lnet_libmd *md;
4598         int cpt;
4599
4600         src.nid = hdr->src_nid;
4601         src.pid = hdr->src_pid;
4602
4603         /* Convert ack fields to host byte order */
4604         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4605         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4606
4607         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4608         lnet_res_lock(cpt);
4609
4610         /* NB handles only looked up by creator (no flips) */
4611         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4612         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4613                 /* Don't moan; this is expected */
4614                 CDEBUG(D_NET,
4615                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4616                        libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4617                        (md == NULL) ? "invalid" : "inactive",
4618                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4619                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4620                 if (md != NULL && md->md_me != NULL)
4621                         CERROR("Source MD also attached to portal %d\n",
4622                                md->md_me->me_portal);
4623
4624                 lnet_res_unlock(cpt);
4625                 return -ENOENT;                  /* -ve! */
4626         }
4627
4628         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4629                libcfs_nidstr(&ni->ni_nid), libcfs_idstr(&src),
4630                hdr->msg.ack.dst_wmd.wh_object_cookie);
4631
4632         lnet_msg_attach_md(msg, md, 0, 0);
4633
4634         lnet_res_unlock(cpt);
4635
4636         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4637
4638         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4639         return 0;
4640 }
4641
4642 /**
4643  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4644  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4645  * \retval -ve                  error code
4646  */
4647 int
4648 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4649 {
4650         int     rc = 0;
4651
4652         if (!the_lnet.ln_routing)
4653                 return -ECANCELED;
4654
4655         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4656             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4657                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4658                         msg->msg_rx_ready_delay = 1;
4659                 } else {
4660                         lnet_net_unlock(msg->msg_rx_cpt);
4661                         rc = lnet_ni_eager_recv(ni, msg);
4662                         lnet_net_lock(msg->msg_rx_cpt);
4663                 }
4664         }
4665
4666         if (rc == 0)
4667                 rc = lnet_post_routed_recv_locked(msg, 0);
4668         return rc;
4669 }
4670
4671 int
4672 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4673 {
4674         int     rc;
4675
4676         switch (msg->msg_type) {
4677         case LNET_MSG_ACK:
4678                 rc = lnet_parse_ack(ni, msg);
4679                 break;
4680         case LNET_MSG_PUT:
4681                 rc = lnet_parse_put(ni, msg);
4682                 break;
4683         case LNET_MSG_GET:
4684                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4685                 break;
4686         case LNET_MSG_REPLY:
4687                 rc = lnet_parse_reply(ni, msg);
4688                 break;
4689         default: /* prevent an unused label if !kernel */
4690                 LASSERT(0);
4691                 return -EPROTO;
4692         }
4693
4694         LASSERT(rc == 0 || rc == -ENOENT);
4695         return rc;
4696 }
4697
4698 char *
4699 lnet_msgtyp2str (int type)
4700 {
4701         switch (type) {
4702         case LNET_MSG_ACK:
4703                 return ("ACK");
4704         case LNET_MSG_PUT:
4705                 return ("PUT");
4706         case LNET_MSG_GET:
4707                 return ("GET");
4708         case LNET_MSG_REPLY:
4709                 return ("REPLY");
4710         case LNET_MSG_HELLO:
4711                 return ("HELLO");
4712         default:
4713                 return ("<UNKNOWN>");
4714         }
4715 }
4716 EXPORT_SYMBOL(lnet_msgtyp2str);
4717
4718 int
4719 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr,
4720            struct lnet_nid *from_nid, void *private, int rdma_req)
4721 {
4722         struct lnet_peer_ni *lpni;
4723         struct lnet_msg *msg;
4724         __u32 payload_length;
4725         lnet_pid_t dest_pid;
4726         struct lnet_nid dest_nid;
4727         struct lnet_nid src_nid;
4728         bool push = false;
4729         int for_me;
4730         __u32 type;
4731         int rc = 0;
4732         int cpt;
4733         time64_t now = ktime_get_seconds();
4734
4735         LASSERT (!in_interrupt ());
4736
4737         type = hdr->type;
4738         src_nid = hdr->src_nid;
4739         dest_nid = hdr->dest_nid;
4740         dest_pid = hdr->dest_pid;
4741         payload_length = hdr->payload_length;
4742
4743         for_me = nid_same(&ni->ni_nid, &dest_nid);
4744         cpt = lnet_nid2cpt(from_nid, ni);
4745
4746         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4747                 libcfs_nidstr(&dest_nid),
4748                 libcfs_nidstr(&ni->ni_nid),
4749                 libcfs_nidstr(&src_nid),
4750                 lnet_msgtyp2str(type),
4751                 (for_me) ? "for me" : "routed");
4752
4753         switch (type) {
4754         case LNET_MSG_ACK:
4755         case LNET_MSG_GET:
4756                 if (payload_length > 0) {
4757                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4758                                libcfs_nidstr(from_nid),
4759                                libcfs_nidstr(&src_nid),
4760                                lnet_msgtyp2str(type), payload_length);
4761                         return -EPROTO;
4762                 }
4763                 break;
4764
4765         case LNET_MSG_PUT:
4766         case LNET_MSG_REPLY:
4767                 if (payload_length >
4768                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4769                         CERROR("%s, src %s: bad %s payload %d "
4770                                "(%d max expected)\n",
4771                                libcfs_nidstr(from_nid),
4772                                libcfs_nidstr(&src_nid),
4773                                lnet_msgtyp2str(type),
4774                                payload_length,
4775                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4776                         return -EPROTO;
4777                 }
4778                 break;
4779
4780         default:
4781                 CERROR("%s, src %s: Bad message type 0x%x\n",
4782                        libcfs_nidstr(from_nid),
4783                        libcfs_nidstr(&src_nid), type);
4784                 return -EPROTO;
4785         }
4786
4787         /* Only update net_last_alive for incoming GETs on the reserved portal
4788          * (i.e. incoming lnet/discovery pings).
4789          * This avoids situations where the router's own traffic results in NI
4790          * status changes
4791          */
4792         if (the_lnet.ln_routing && type == LNET_MSG_GET &&
4793             hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
4794             !lnet_islocalnid(&src_nid) &&
4795             ni->ni_net->net_last_alive != now) {
4796                 lnet_ni_lock(ni);
4797                 spin_lock(&ni->ni_net->net_lock);
4798                 ni->ni_net->net_last_alive = now;
4799                 spin_unlock(&ni->ni_net->net_lock);
4800                 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4801                 lnet_ni_unlock(ni);
4802         }
4803
4804         if (push)
4805                 lnet_push_update_to_peers(1);
4806
4807         /* Regard a bad destination NID as a protocol error.  Senders should
4808          * know what they're doing; if they don't they're misconfigured, buggy
4809          * or malicious so we chop them off at the knees :) */
4810
4811         if (!for_me) {
4812                 if (LNET_NID_NET(&dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4813                         /* should have gone direct */
4814                         CERROR("%s, src %s: Bad dest nid %s "
4815                                "(should have been sent direct)\n",
4816                                 libcfs_nidstr(from_nid),
4817                                 libcfs_nidstr(&src_nid),
4818                                 libcfs_nidstr(&dest_nid));
4819                         return -EPROTO;
4820                 }
4821
4822                 if (lnet_islocalnid(&dest_nid)) {
4823                         /* dest is another local NI; sender should have used
4824                          * this node's NID on its own network */
4825                         CERROR("%s, src %s: Bad dest nid %s "
4826                                "(it's my nid but on a different network)\n",
4827                                 libcfs_nidstr(from_nid),
4828                                 libcfs_nidstr(&src_nid),
4829                                 libcfs_nidstr(&dest_nid));
4830                         return -EPROTO;
4831                 }
4832
4833                 if (rdma_req && type == LNET_MSG_GET) {
4834                         CERROR("%s, src %s: Bad optimized GET for %s "
4835                                "(final destination must be me)\n",
4836                                 libcfs_nidstr(from_nid),
4837                                 libcfs_nidstr(&src_nid),
4838                                 libcfs_nidstr(&dest_nid));
4839                         return -EPROTO;
4840                 }
4841
4842                 if (!the_lnet.ln_routing) {
4843                         CERROR("%s, src %s: Dropping message for %s "
4844                                "(routing not enabled)\n",
4845                                 libcfs_nidstr(from_nid),
4846                                 libcfs_nidstr(&src_nid),
4847                                 libcfs_nidstr(&dest_nid));
4848                         goto drop;
4849                 }
4850         }
4851
4852         /* Message looks OK; we're not going to return an error, so we MUST
4853          * call back lnd_recv() come what may... */
4854
4855         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4856             fail_peer(&src_nid, 0)) {                   /* shall we now? */
4857                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4858                        libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4859                        lnet_msgtyp2str(type));
4860                 goto drop;
4861         }
4862
4863         if (!list_empty(&the_lnet.ln_drop_rules) &&
4864             lnet_drop_rule_match(hdr, &ni->ni_nid, NULL)) {
4865                 CDEBUG(D_NET,
4866                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4867                        libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4868                        libcfs_nidstr(&dest_nid), lnet_msgtyp2str(type));
4869                 goto drop;
4870         }
4871
4872         msg = lnet_msg_alloc();
4873         if (msg == NULL) {
4874                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4875                        libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4876                        lnet_msgtyp2str(type));
4877                 goto drop;
4878         }
4879
4880         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4881          * pointers NULL etc */
4882
4883         msg->msg_type = type;
4884         msg->msg_private = private;
4885         msg->msg_receiving = 1;
4886         msg->msg_rdma_get = rdma_req;
4887         msg->msg_len = msg->msg_wanted = payload_length;
4888         msg->msg_offset = 0;
4889         msg->msg_hdr = *hdr;
4890         /* for building message event */
4891         msg->msg_from = *from_nid;
4892         if (!for_me) {
4893                 msg->msg_target.pid = dest_pid;
4894                 msg->msg_target.nid = dest_nid;
4895                 msg->msg_routing = 1;
4896         }
4897
4898         lnet_net_lock(cpt);
4899         lpni = lnet_peerni_by_nid_locked(from_nid, &ni->ni_nid, cpt);
4900         if (IS_ERR(lpni)) {
4901                 lnet_net_unlock(cpt);
4902                 rc = PTR_ERR(lpni);
4903                 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4904                        libcfs_nidstr(from_nid), libcfs_nidstr(&src_nid),
4905                        lnet_msgtyp2str(type), rc);
4906                 lnet_msg_free(msg);
4907                 if (rc == -ESHUTDOWN)
4908                         /* We are shutting down.  Don't do anything more */
4909                         return rc;
4910                 goto drop;
4911         }
4912
4913         /* If this message was forwarded to us from a router then we may need
4914          * to update router aliveness or check for an asymmetrical route
4915          * (or both)
4916          */
4917         if (((lnet_drop_asym_route && for_me) ||
4918              !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4919             LNET_NID_NET(&src_nid) != LNET_NID_NET(from_nid)) {
4920                 __u32 src_net_id = LNET_NID_NET(&src_nid);
4921                 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4922                 struct lnet_route *route;
4923                 bool found = false;
4924
4925                 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4926                         if (route->lr_net == src_net_id) {
4927                                 found = true;
4928                                 /* If we're transitioning the gateway from
4929                                  * dead -> alive, and discovery is disabled
4930                                  * locally or on the gateway, then we need to
4931                                  * update the cached route aliveness for each
4932                                  * route to the src_nid's net.
4933                                  *
4934                                  * Otherwise, we're only checking for
4935                                  * symmetrical route, and we can break the
4936                                  * loop
4937                                  */
4938                                 if (!gw->lp_alive &&
4939                                     lnet_is_discovery_disabled(gw))
4940                                         lnet_set_route_aliveness(route, true);
4941                                 else
4942                                         break;
4943                         }
4944                 }
4945                 if (lnet_drop_asym_route && for_me && !found) {
4946                         /* Drop ref taken by lnet_nid2peerni_locked() */
4947                         lnet_peer_ni_decref_locked(lpni);
4948                         lnet_net_unlock(cpt);
4949                         /* we would not use from_nid to route a message to
4950                          * src_nid
4951                          * => asymmetric routing detected but forbidden
4952                          */
4953                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4954                                libcfs_nidstr(from_nid),
4955                                libcfs_nidstr(&src_nid), lnet_msgtyp2str(type));
4956                         lnet_msg_free(msg);
4957                         goto drop;
4958                 }
4959                 if (!gw->lp_alive) {
4960                         struct lnet_peer_net *lpn;
4961                         struct lnet_peer_ni *lpni2;
4962
4963                         gw->lp_alive = true;
4964                         /* Mark all remote NIs on src_nid's net UP */
4965                         lpn = lnet_peer_get_net_locked(gw, src_net_id);
4966                         if (lpn)
4967                                 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4968                                                     lpni_peer_nis)
4969                                         lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4970                 }
4971         }
4972
4973         lpni->lpni_last_alive = now;
4974
4975         msg->msg_rxpeer = lpni;
4976         msg->msg_rxni = ni;
4977         lnet_ni_addref_locked(ni, cpt);
4978         /* Multi-Rail: Primary NID of source. */
4979         lnet_peer_primary_nid_locked(&src_nid, &msg->msg_initiator);
4980
4981         /*
4982          * mark the status of this lpni as UP since we received a message
4983          * from it. The ping response reports back the ns_status which is
4984          * marked on the remote as up or down and we cache it here.
4985          */
4986         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4987
4988         lnet_msg_commit(msg, cpt);
4989
4990         /* message delay simulation */
4991         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4992                      lnet_delay_rule_match_locked(hdr, msg))) {
4993                 lnet_net_unlock(cpt);
4994                 return 0;
4995         }
4996
4997         if (!for_me) {
4998                 rc = lnet_parse_forward_locked(ni, msg);
4999                 lnet_net_unlock(cpt);
5000
5001                 if (rc < 0)
5002                         goto free_drop;
5003
5004                 if (rc == LNET_CREDIT_OK) {
5005                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
5006                                      0, payload_length, payload_length);
5007                 }
5008                 return 0;
5009         }
5010
5011         lnet_net_unlock(cpt);
5012
5013         rc = lnet_parse_local(ni, msg);
5014         if (rc != 0)
5015                 goto free_drop;
5016         return 0;
5017
5018  free_drop:
5019         LASSERT(msg->msg_md == NULL);
5020         lnet_finalize(msg, rc);
5021
5022  drop:
5023         lnet_drop_message(ni, cpt, private, payload_length, type);
5024         return 0;
5025 }
5026 EXPORT_SYMBOL(lnet_parse);
5027
5028 void
5029 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
5030 {
5031         struct lnet_msg *msg;
5032
5033         while ((msg = list_first_entry_or_null(head, struct lnet_msg,
5034                                                msg_list)) != NULL) {
5035                 struct lnet_processid id = {};
5036
5037                 list_del(&msg->msg_list);
5038
5039                 id.nid = msg->msg_hdr.src_nid;
5040                 id.pid = msg->msg_hdr.src_pid;
5041
5042                 LASSERT(msg->msg_md == NULL);
5043                 LASSERT(msg->msg_rx_delayed);
5044                 LASSERT(msg->msg_rxpeer != NULL);
5045                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
5046
5047                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
5048                       " offset %d length %d: %s\n",
5049                       libcfs_idstr(&id),
5050                       msg->msg_hdr.msg.put.ptl_index,
5051                       msg->msg_hdr.msg.put.match_bits,
5052                       msg->msg_hdr.msg.put.offset,
5053                       msg->msg_hdr.payload_length, reason);
5054
5055                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
5056                  * called lnet_drop_message(), so I just hang onto msg as well
5057                  * until that's done */
5058
5059                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
5060                                   msg->msg_private, msg->msg_len,
5061                                   msg->msg_type);
5062
5063                 msg->msg_no_resend = true;
5064                 /*
5065                  * NB: message will not generate event because w/o attached MD,
5066                  * but we still should give error code so lnet_msg_decommit()
5067                  * can skip counters operations and other checks.
5068                  */
5069                 lnet_finalize(msg, -ENOENT);
5070         }
5071 }
5072
5073 void
5074 lnet_recv_delayed_msg_list(struct list_head *head)
5075 {
5076         struct lnet_msg *msg;
5077
5078         while ((msg = list_first_entry_or_null(head, struct lnet_msg,
5079                                                msg_list)) != NULL) {
5080                 struct lnet_processid id;
5081
5082                 list_del(&msg->msg_list);
5083
5084                 /* md won't disappear under me, since each msg
5085                  * holds a ref on it */
5086
5087                 id.nid = msg->msg_hdr.src_nid;
5088                 id.pid = msg->msg_hdr.src_pid;
5089
5090                 LASSERT(msg->msg_rx_delayed);
5091                 LASSERT(msg->msg_md != NULL);
5092                 LASSERT(msg->msg_rxpeer != NULL);
5093                 LASSERT(msg->msg_rxni != NULL);
5094                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
5095
5096                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
5097                        "match %llu offset %d length %d.\n",
5098                         libcfs_idstr(&id), msg->msg_hdr.msg.put.ptl_index,
5099                         msg->msg_hdr.msg.put.match_bits,
5100                         msg->msg_hdr.msg.put.offset,
5101                         msg->msg_hdr.payload_length);
5102
5103                 lnet_recv_put(msg->msg_rxni, msg);
5104         }
5105 }
5106
5107 static void
5108 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
5109                         struct lnet_libmd *md, struct lnet_handle_md mdh)
5110 {
5111         s64 timeout_ns;
5112         struct lnet_rsp_tracker *local_rspt;
5113
5114         /*
5115          * MD has a refcount taken by message so it's not going away.
5116          * The MD however can be looked up. We need to secure the access
5117          * to the md_rspt_ptr by taking the res_lock.
5118          * The rspt can be accessed without protection up to when it gets
5119          * added to the list.
5120          */
5121
5122         lnet_res_lock(cpt);
5123         local_rspt = md->md_rspt_ptr;
5124         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
5125         if (local_rspt != NULL) {
5126                 /*
5127                  * we already have an rspt attached to the md, so we'll
5128                  * update the deadline on that one.
5129                  */
5130                 lnet_rspt_free(rspt, cpt);
5131         } else {
5132                 /* new md */
5133                 rspt->rspt_mdh = mdh;
5134                 rspt->rspt_cpt = cpt;
5135                 /* store the rspt so we can access it when we get the REPLY */
5136                 md->md_rspt_ptr = rspt;
5137                 local_rspt = rspt;
5138         }
5139         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
5140
5141         /*
5142          * add to the list of tracked responses. It's added to tail of the
5143          * list in order to expire all the older entries first.
5144          */
5145         lnet_net_lock(cpt);
5146         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
5147         lnet_net_unlock(cpt);
5148         lnet_res_unlock(cpt);
5149 }
5150
5151 /**
5152  * Initiate an asynchronous PUT operation.
5153  *
5154  * There are several events associated with a PUT: completion of the send on
5155  * the initiator node (LNET_EVENT_SEND), and when the send completes
5156  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
5157  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
5158  * used at the target node to indicate the completion of incoming data
5159  * delivery.
5160  *
5161  * The local events will be logged in the EQ associated with the MD pointed to
5162  * by \a mdh handle. Using a MD without an associated EQ results in these
5163  * events being discarded. In this case, the caller must have another
5164  * mechanism (e.g., a higher level protocol) for determining when it is safe
5165  * to modify the memory region associated with the MD.
5166  *
5167  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
5168  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
5169  *
5170  * \param self Indicates the NID of a local interface through which to send
5171  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
5172  * \param mdh A handle for the MD that describes the memory to be sent. The MD
5173  * must be "free floating" (See LNetMDBind()).
5174  * \param ack Controls whether an acknowledgment is requested.
5175  * Acknowledgments are only sent when they are requested by the initiating
5176  * process and the target MD enables them.
5177  * \param target A process identifier for the target process.
5178  * \param portal The index in the \a target's portal table.
5179  * \param match_bits The match bits to use for MD selection at the target
5180  * process.
5181  * \param offset The offset into the target MD (only used when the target
5182  * MD has the LNET_MD_MANAGE_REMOTE option set).
5183  * \param hdr_data 64 bits of user data that can be included in the message
5184  * header. This data is written to an event queue entry at the target if an
5185  * EQ is present on the matching MD.
5186  *
5187  * \retval  0      Success, and only in this case events will be generated
5188  * and logged to EQ (if it exists).
5189  * \retval -EIO    Simulated failure.
5190  * \retval -ENOMEM Memory allocation failure.
5191  * \retval -ENOENT Invalid MD object.
5192  *
5193  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
5194  */
5195 int
5196 LNetPut(struct lnet_nid *self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
5197         struct lnet_processid *target, unsigned int portal,
5198         __u64 match_bits, unsigned int offset,
5199         __u64 hdr_data)
5200 {
5201         struct lnet_msg *msg;
5202         struct lnet_libmd *md;
5203         int cpt;
5204         int rc;
5205         struct lnet_rsp_tracker *rspt = NULL;
5206
5207         LASSERT(the_lnet.ln_refcount > 0);
5208
5209         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
5210             fail_peer(&target->nid, 1)) {               /* shall we now? */
5211                 CERROR("Dropping PUT to %s: simulated failure\n",
5212                        libcfs_idstr(target));
5213                 return -EIO;
5214         }
5215
5216         msg = lnet_msg_alloc();
5217         if (msg == NULL) {
5218                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5219                        libcfs_idstr(target));
5220                 return -ENOMEM;
5221         }
5222         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5223
5224         cpt = lnet_cpt_of_cookie(mdh.cookie);
5225
5226         if (ack == LNET_ACK_REQ) {
5227                 rspt = lnet_rspt_alloc(cpt);
5228                 if (!rspt) {
5229                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5230                                 libcfs_idstr(target));
5231                         return -ENOMEM;
5232                 }
5233                 INIT_LIST_HEAD(&rspt->rspt_on_list);
5234         }
5235
5236         lnet_res_lock(cpt);
5237
5238         md = lnet_handle2md(&mdh);
5239         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5240                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5241                        match_bits, portal, libcfs_idstr(target),
5242                        md == NULL ? -1 : md->md_threshold);
5243                 if (md != NULL && md->md_me != NULL)
5244                         CERROR("Source MD also attached to portal %d\n",
5245                                md->md_me->me_portal);
5246                 lnet_res_unlock(cpt);
5247
5248                 if (rspt)
5249                         lnet_rspt_free(rspt, cpt);
5250
5251                 lnet_msg_free(msg);
5252                 return -ENOENT;
5253         }
5254
5255         CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5256
5257         lnet_msg_attach_md(msg, md, 0, 0);
5258
5259         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5260
5261         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5262         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5263         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5264         msg->msg_hdr.msg.put.hdr_data = hdr_data;
5265
5266         /* NB handles only looked up by creator (no flips) */
5267         if (ack == LNET_ACK_REQ) {
5268                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5269                         the_lnet.ln_interface_cookie;
5270                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5271                         md->md_lh.lh_cookie;
5272         } else {
5273                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5274                         LNET_WIRE_HANDLE_COOKIE_NONE;
5275                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5276                         LNET_WIRE_HANDLE_COOKIE_NONE;
5277         }
5278
5279         lnet_res_unlock(cpt);
5280
5281         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5282
5283         if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5284                                                    md->md_options))
5285                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5286         else if (rspt)
5287                 lnet_rspt_free(rspt, cpt);
5288
5289         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5290                                  CFS_FAIL_ONCE))
5291                 rc = -EIO;
5292         else
5293                 rc = lnet_send(self, msg, NULL);
5294
5295         if (rc != 0) {
5296                 CNETERR("Error sending PUT to %s: %d\n",
5297                         libcfs_idstr(target), rc);
5298                 msg->msg_no_resend = true;
5299                 lnet_finalize(msg, rc);
5300         }
5301
5302         /* completion will be signalled by an event */
5303         return 0;
5304 }
5305 EXPORT_SYMBOL(LNetPut);
5306
5307 /*
5308  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
5309  * returns a msg for the LND to pass to lnet_finalize() when the sink
5310  * data has been received.
5311  *
5312  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5313  * lnet_finalize() is called on it, so the LND must call this first
5314  */
5315 struct lnet_msg *
5316 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5317 {
5318         struct lnet_msg *msg = lnet_msg_alloc();
5319         struct lnet_libmd *getmd = getmsg->msg_md;
5320         struct lnet_processid *peer_id = &getmsg->msg_target;
5321         int cpt;
5322
5323         LASSERT(!getmsg->msg_target_is_router);
5324         LASSERT(!getmsg->msg_routing);
5325
5326         if (msg == NULL) {
5327                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5328                        libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5329                 goto drop;
5330         }
5331
5332         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5333         lnet_res_lock(cpt);
5334
5335         LASSERT(getmd->md_refcount > 0);
5336
5337         if (getmd->md_threshold == 0) {
5338                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5339                         libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5340                         getmd);
5341                 lnet_res_unlock(cpt);
5342                 goto drop;
5343         }
5344
5345         LASSERT(getmd->md_offset == 0);
5346
5347         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5348                libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5349
5350         /* setup information for lnet_build_msg_event */
5351         msg->msg_initiator =
5352                 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5353         msg->msg_from = peer_id->nid;
5354         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5355         msg->msg_hdr.src_nid = peer_id->nid;
5356         msg->msg_hdr.payload_length = getmd->md_length;
5357         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5358
5359         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5360         lnet_res_unlock(cpt);
5361
5362         cpt = lnet_nid2cpt(&peer_id->nid, ni);
5363
5364         lnet_net_lock(cpt);
5365         lnet_msg_commit(msg, cpt);
5366         lnet_net_unlock(cpt);
5367
5368         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5369
5370         return msg;
5371
5372  drop:
5373         cpt = lnet_nid2cpt(&peer_id->nid, ni);
5374
5375         lnet_net_lock(cpt);
5376         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5377         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5378         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5379                 getmd->md_length;
5380         lnet_net_unlock(cpt);
5381
5382         if (msg != NULL)
5383                 lnet_msg_free(msg);
5384
5385         return NULL;
5386 }
5387 EXPORT_SYMBOL(lnet_create_reply_msg);
5388
5389 void
5390 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5391                        unsigned int len)
5392 {
5393         /* Set the REPLY length, now the RDMA that elides the REPLY message has
5394          * completed and I know it. */
5395         LASSERT(reply != NULL);
5396         LASSERT(reply->msg_type == LNET_MSG_GET);
5397         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5398
5399         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
5400          * the end of my buffer, I might as well be dead. */
5401         LASSERT(len <= reply->msg_ev.mlength);
5402
5403         reply->msg_ev.mlength = len;
5404 }
5405 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5406
5407 /**
5408  * Initiate an asynchronous GET operation.
5409  *
5410  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5411  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5412  * the target node in the REPLY has been written to local MD.
5413  *
5414  * On the target node, an LNET_EVENT_GET is logged when the GET request
5415  * arrives and is accepted into a MD.
5416  *
5417  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5418  * \param mdh A handle for the MD that describes the memory into which the
5419  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5420  *
5421  * \retval  0      Success, and only in this case events will be generated
5422  * and logged to EQ (if it exists) of the MD.
5423  * \retval -EIO    Simulated failure.
5424  * \retval -ENOMEM Memory allocation failure.
5425  * \retval -ENOENT Invalid MD object.
5426  */
5427 int
5428 LNetGet(struct lnet_nid *self, struct lnet_handle_md mdh,
5429         struct lnet_processid *target, unsigned int portal,
5430         __u64 match_bits, unsigned int offset, bool recovery)
5431 {
5432         struct lnet_msg *msg;
5433         struct lnet_libmd *md;
5434         struct lnet_rsp_tracker *rspt;
5435         int cpt;
5436         int rc;
5437
5438         LASSERT(the_lnet.ln_refcount > 0);
5439
5440         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
5441             fail_peer(&target->nid, 1))         /* shall we now? */
5442         {
5443                 CERROR("Dropping GET to %s: simulated failure\n",
5444                        libcfs_idstr(target));
5445                 return -EIO;
5446         }
5447
5448         msg = lnet_msg_alloc();
5449         if (!msg) {
5450                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5451                        libcfs_idstr(target));
5452                 return -ENOMEM;
5453         }
5454
5455         cpt = lnet_cpt_of_cookie(mdh.cookie);
5456
5457         rspt = lnet_rspt_alloc(cpt);
5458         if (!rspt) {
5459                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5460                        libcfs_idstr(target));
5461                 return -ENOMEM;
5462         }
5463         INIT_LIST_HEAD(&rspt->rspt_on_list);
5464
5465         msg->msg_recovery = recovery;
5466
5467         lnet_res_lock(cpt);
5468
5469         md = lnet_handle2md(&mdh);
5470         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5471                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5472                        match_bits, portal, libcfs_idstr(target),
5473                        md == NULL ? -1 : md->md_threshold);
5474                 if (md != NULL && md->md_me != NULL)
5475                         CERROR("REPLY MD also attached to portal %d\n",
5476                                md->md_me->me_portal);
5477
5478                 lnet_res_unlock(cpt);
5479
5480                 lnet_msg_free(msg);
5481                 lnet_rspt_free(rspt, cpt);
5482                 return -ENOENT;
5483         }
5484
5485         CDEBUG(D_NET, "%s -> %s\n", __func__, libcfs_idstr(target));
5486
5487         lnet_msg_attach_md(msg, md, 0, 0);
5488
5489         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5490
5491         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5492         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5493         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5494         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5495
5496         /* NB handles only looked up by creator (no flips) */
5497         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5498                 the_lnet.ln_interface_cookie;
5499         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5500                 md->md_lh.lh_cookie;
5501
5502         lnet_res_unlock(cpt);
5503
5504         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5505
5506         if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5507                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5508         else
5509                 lnet_rspt_free(rspt, cpt);
5510
5511         rc = lnet_send(self, msg, NULL);
5512         if (rc < 0) {
5513                 CNETERR("Error sending GET to %s: %d\n",
5514                         libcfs_idstr(target), rc);
5515                 msg->msg_no_resend = true;
5516                 lnet_finalize(msg, rc);
5517         }
5518
5519         /* completion will be signalled by an event */
5520         return 0;
5521 }
5522 EXPORT_SYMBOL(LNetGet);
5523
5524 /**
5525  * Calculate distance to node at \a dstnid.
5526  *
5527  * \param dstnid Target NID.
5528  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5529  * is saved here.
5530  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5531  * here.
5532  *
5533  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5534  * local_nid_dist_zero is set, which is the default.
5535  * \retval positives Distance to target NID, i.e. number of hops plus one.
5536  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5537  */
5538 int
5539 LNetDist(struct lnet_nid *dstnid, struct lnet_nid *srcnid, __u32 *orderp)
5540 {
5541         struct lnet_ni *ni = NULL;
5542         struct lnet_remotenet *rnet;
5543         __u32 dstnet = LNET_NID_NET(dstnid);
5544         int hops;
5545         int cpt;
5546         __u32 order = 2;
5547         struct list_head *rn_list;
5548         struct lnet_ni *matched_dstnet = NULL;
5549
5550         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5551          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5552          * keep order 0 free for 0@lo and order 1 free for a local NID
5553          * match
5554          * WARNING: dstnid and srcnid might point to same place.
5555          * Don't set *srcnid until late.
5556          */
5557
5558         LASSERT(the_lnet.ln_refcount > 0);
5559
5560         cpt = lnet_net_lock_current();
5561
5562         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5563                 if (nid_same(&ni->ni_nid, dstnid)) {
5564                         if (orderp != NULL) {
5565                                 if (nid_is_lo0(dstnid))
5566                                         *orderp = 0;
5567                                 else
5568                                         *orderp = 1;
5569                         }
5570                         if (srcnid)
5571                                 *srcnid = *dstnid;
5572                         lnet_net_unlock(cpt);
5573
5574                         return local_nid_dist_zero ? 0 : 1;
5575                 }
5576
5577                 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5578                         matched_dstnet = ni;
5579                         /* We matched the destination net, but we may have
5580                          * additional local NIs to inspect.
5581                          *
5582                          * We record the order as appropriate, but
5583                          * they may be overwritten if we match local NI above.
5584                          */
5585
5586                         if (orderp) {
5587                                 /* Check if ni was originally created in
5588                                  * current net namespace.
5589                                  * If not, assign order above 0xffff0000,
5590                                  * to make this ni not a priority.
5591                                  */
5592                                 if (current->nsproxy &&
5593                                     !net_eq(ni->ni_net_ns,
5594                                             current->nsproxy->net_ns))
5595                                         *orderp = order + 0xffff0000;
5596                                 else
5597                                         *orderp = order;
5598                         }
5599                 }
5600
5601                 order++;
5602         }
5603
5604         if (matched_dstnet) {
5605                 if (srcnid)
5606                         *srcnid = matched_dstnet->ni_nid;
5607                 lnet_net_unlock(cpt);
5608                 return 1;
5609         }
5610
5611         rn_list = lnet_net2rnethash(dstnet);
5612         list_for_each_entry(rnet, rn_list, lrn_list) {
5613                 if (rnet->lrn_net == dstnet) {
5614                         struct lnet_route *route;
5615                         struct lnet_route *shortest = NULL;
5616                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5617                         __u32 route_hops;
5618
5619                         LASSERT(!list_empty(&rnet->lrn_routes));
5620
5621                         list_for_each_entry(route, &rnet->lrn_routes,
5622                                             lr_list) {
5623                                 route_hops = route->lr_hops;
5624                                 if (route_hops == LNET_UNDEFINED_HOPS)
5625                                         route_hops = 1;
5626                                 if (shortest == NULL ||
5627                                     route_hops < shortest_hops) {
5628                                         shortest = route;
5629                                         shortest_hops = route_hops;
5630                                 }
5631                         }
5632
5633                         LASSERT(shortest != NULL);
5634                         hops = shortest_hops;
5635                         if (srcnid) {
5636                                 struct lnet_net *net;
5637                                 net = lnet_get_net_locked(shortest->lr_lnet);
5638                                 LASSERT(net);
5639                                 ni = lnet_get_next_ni_locked(net, NULL);
5640                                 *srcnid = ni->ni_nid;
5641                         }
5642                         if (orderp != NULL)
5643                                 *orderp = order;
5644                         lnet_net_unlock(cpt);
5645                         return hops + 1;
5646                 }
5647                 order++;
5648         }
5649
5650         lnet_net_unlock(cpt);
5651         return -EHOSTUNREACH;
5652 }
5653 EXPORT_SYMBOL(LNetDist);