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         lnet_nid_t sd_dst_nid;
  59         lnet_nid_t sd_src_nid;
  60         lnet_nid_t sd_rtr_nid;
  61         int sd_cpt;
  62         int sd_md_cpt;
  63         __u32 sd_send_case;
  64 };
  65
  66 static inline bool
  67 lnet_msg_is_response(struct lnet_msg *msg)
  68 {
  69         return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
  70 }
  71
  72 static inline bool
  73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
  74 {
  75         if (md_options & LNET_MD_NO_TRACK_RESPONSE)
  76                 /* Explicitly disabled in MD options */
  77                 return false;
  78
  79         if (md_options & LNET_MD_TRACK_RESPONSE)
  80                 /* Explicity enabled in MD options */
  81                 return true;
  82
  83         if (lnet_response_tracking == 3)
  84                 /* Enabled for all message types */
  85                 return true;
  86
  87         if (msg_type == LNET_MSG_PUT)
  88                 return lnet_response_tracking == 2;
  89
  90         if (msg_type == LNET_MSG_GET)
  91                 return lnet_response_tracking == 1;
  92
  93         return false;
  94 }
  95
  96 static inline struct lnet_comm_count *
  97 get_stats_counts(struct lnet_element_stats *stats,
  98                  enum lnet_stats_type stats_type)
  99 {
 100         switch (stats_type) {
 101         case LNET_STATS_TYPE_SEND:
 102                 return &stats->el_send_stats;
 103         case LNET_STATS_TYPE_RECV:
 104                 return &stats->el_recv_stats;
 105         case LNET_STATS_TYPE_DROP:
 106                 return &stats->el_drop_stats;
 107         default:
 108                 CERROR("Unknown stats type\n");
 109         }
 110
 111         return NULL;
 112 }
 113
 114 void lnet_incr_stats(struct lnet_element_stats *stats,
 115                      enum lnet_msg_type msg_type,
 116                      enum lnet_stats_type stats_type)
 117 {
 118         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 119         if (!counts)
 120                 return;
 121
 122         switch (msg_type) {
 123         case LNET_MSG_ACK:
 124                 atomic_inc(&counts->co_ack_count);
 125                 break;
 126         case LNET_MSG_PUT:
 127                 atomic_inc(&counts->co_put_count);
 128                 break;
 129         case LNET_MSG_GET:
 130                 atomic_inc(&counts->co_get_count);
 131                 break;
 132         case LNET_MSG_REPLY:
 133                 atomic_inc(&counts->co_reply_count);
 134                 break;
 135         case LNET_MSG_HELLO:
 136                 atomic_inc(&counts->co_hello_count);
 137                 break;
 138         default:
 139                 CERROR("There is a BUG in the code. Unknown message type\n");
 140                 break;
 141         }
 142 }
 143
 144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 145                      enum lnet_stats_type stats_type)
 146 {
 147         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 148         if (!counts)
 149                 return 0;
 150
 151         return (atomic_read(&counts->co_ack_count) +
 152                 atomic_read(&counts->co_put_count) +
 153                 atomic_read(&counts->co_get_count) +
 154                 atomic_read(&counts->co_reply_count) +
 155                 atomic_read(&counts->co_hello_count));
 156 }
 157
 158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 159                                 struct lnet_comm_count *counts)
 160 {
 161         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 162         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 163         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 164         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 165         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 166 }
 167
 168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 169                               struct lnet_element_stats *stats)
 170 {
 171         struct lnet_comm_count *counts;
 172
 173         LASSERT(msg_stats);
 174         LASSERT(stats);
 175
 176         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 177         if (!counts)
 178                 return;
 179         assign_stats(&msg_stats->im_send_stats, counts);
 180
 181         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 182         if (!counts)
 183                 return;
 184         assign_stats(&msg_stats->im_recv_stats, counts);
 185
 186         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 187         if (!counts)
 188                 return;
 189         assign_stats(&msg_stats->im_drop_stats, counts);
 190 }
 191
 192 int
 193 lnet_fail_nid(lnet_nid_t 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 (!list_empty(&cull)) {
 233                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 234
 235                 list_del(&tp->tp_list);
 236                 LIBCFS_FREE(tp, sizeof(*tp));
 237         }
 238         return 0;
 239 }
 240
 241 static int
 242 fail_peer(lnet_nid_t nid4, int outgoing)
 243 {
 244         struct lnet_test_peer *tp;
 245         struct list_head *el;
 246         struct list_head *next;
 247         struct lnet_nid nid;
 248         LIST_HEAD(cull);
 249         int fail = 0;
 250
 251         lnet_nid4_to_nid(nid4, &nid);
 252         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 253         lnet_net_lock(0);
 254
 255         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 256                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 257
 258                 if (tp->tp_threshold == 0) {
 259                         /* zombie entry */
 260                         if (outgoing) {
 261                                 /* only cull zombies on outgoing tests,
 262                                  * since we may be at interrupt priority on
 263                                  * incoming messages. */
 264                                 list_move(&tp->tp_list, &cull);
 265                         }
 266                         continue;
 267                 }
 268
 269                 if (LNET_NID_IS_ANY(&tp->tp_nid) ||     /* fail every peer */
 270                     nid_same(&nid, &tp->tp_nid)) {      /* fail this peer */
 271                         fail = 1;
 272
 273                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 274                                 tp->tp_threshold--;
 275                                 if (outgoing &&
 276                                     tp->tp_threshold == 0) {
 277                                         /* see above */
 278                                         list_move(&tp->tp_list, &cull);
 279                                 }
 280                         }
 281                         break;
 282                 }
 283         }
 284
 285         lnet_net_unlock(0);
 286
 287         while (!list_empty(&cull)) {
 288                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 289                 list_del(&tp->tp_list);
 290
 291                 LIBCFS_FREE(tp, sizeof(*tp));
 292         }
 293
 294         return fail;
 295 }
 296
 297 unsigned int
 298 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 299 {
 300         unsigned int nob = 0;
 301
 302         LASSERT(niov == 0 || iov != NULL);
 303         while (niov-- > 0)
 304                 nob += (iov++)->iov_len;
 305
 306         return (nob);
 307 }
 308 EXPORT_SYMBOL(lnet_iov_nob);
 309
 310 void
 311 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 312                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 313                   unsigned int nob)
 314 {
 315         /* NB diov, siov are READ-ONLY */
 316         unsigned int this_nob;
 317
 318         if (nob == 0)
 319                 return;
 320
 321         /* skip complete frags before 'doffset' */
 322         LASSERT(ndiov > 0);
 323         while (doffset >= diov->iov_len) {
 324                 doffset -= diov->iov_len;
 325                 diov++;
 326                 ndiov--;
 327                 LASSERT(ndiov > 0);
 328         }
 329
 330         /* skip complete frags before 'soffset' */
 331         LASSERT(nsiov > 0);
 332         while (soffset >= siov->iov_len) {
 333                 soffset -= siov->iov_len;
 334                 siov++;
 335                 nsiov--;
 336                 LASSERT(nsiov > 0);
 337         }
 338
 339         do {
 340                 LASSERT(ndiov > 0);
 341                 LASSERT(nsiov > 0);
 342                 this_nob = min3((unsigned int)diov->iov_len - doffset,
 343                                 (unsigned int)siov->iov_len - soffset,
 344                                 nob);
 345
 346                 memcpy((char *)diov->iov_base + doffset,
 347                        (char *)siov->iov_base + soffset, this_nob);
 348                 nob -= this_nob;
 349
 350                 if (diov->iov_len > doffset + this_nob) {
 351                         doffset += this_nob;
 352                 } else {
 353                         diov++;
 354                         ndiov--;
 355                         doffset = 0;
 356                 }
 357
 358                 if (siov->iov_len > soffset + this_nob) {
 359                         soffset += this_nob;
 360                 } else {
 361                         siov++;
 362                         nsiov--;
 363                         soffset = 0;
 364                 }
 365         } while (nob > 0);
 366 }
 367 EXPORT_SYMBOL(lnet_copy_iov2iov);
 368
 369 unsigned int
 370 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 371 {
 372         unsigned int  nob = 0;
 373
 374         LASSERT(niov == 0 || kiov != NULL);
 375         while (niov-- > 0)
 376                 nob += (kiov++)->bv_len;
 377
 378         return (nob);
 379 }
 380 EXPORT_SYMBOL(lnet_kiov_nob);
 381
 382 void
 383 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
 384                     unsigned int doffset,
 385                     unsigned int nsiov, struct bio_vec *siov,
 386                     unsigned int soffset,
 387                     unsigned int nob)
 388 {
 389         /* NB diov, siov are READ-ONLY */
 390         unsigned int    this_nob;
 391         char           *daddr = NULL;
 392         char           *saddr = NULL;
 393
 394         if (nob == 0)
 395                 return;
 396
 397         LASSERT (!in_interrupt ());
 398
 399         LASSERT (ndiov > 0);
 400         while (doffset >= diov->bv_len) {
 401                 doffset -= diov->bv_len;
 402                 diov++;
 403                 ndiov--;
 404                 LASSERT(ndiov > 0);
 405         }
 406
 407         LASSERT(nsiov > 0);
 408         while (soffset >= siov->bv_len) {
 409                 soffset -= siov->bv_len;
 410                 siov++;
 411                 nsiov--;
 412                 LASSERT(nsiov > 0);
 413         }
 414
 415         do {
 416                 LASSERT(ndiov > 0);
 417                 LASSERT(nsiov > 0);
 418                 this_nob = min3(diov->bv_len - doffset,
 419                                 siov->bv_len - soffset,
 420                                 nob);
 421
 422                 if (daddr == NULL)
 423                         daddr = ((char *)kmap(diov->bv_page)) +
 424                                 diov->bv_offset + doffset;
 425                 if (saddr == NULL)
 426                         saddr = ((char *)kmap(siov->bv_page)) +
 427                                 siov->bv_offset + soffset;
 428
 429                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 430                  * However in practice at least one of the kiovs will be mapped
 431                  * kernel pages and the map/unmap will be NOOPs */
 432
 433                 memcpy (daddr, saddr, this_nob);
 434                 nob -= this_nob;
 435
 436                 if (diov->bv_len > doffset + this_nob) {
 437                         daddr += this_nob;
 438                         doffset += this_nob;
 439                 } else {
 440                         kunmap(diov->bv_page);
 441                         daddr = NULL;
 442                         diov++;
 443                         ndiov--;
 444                         doffset = 0;
 445                 }
 446
 447                 if (siov->bv_len > soffset + this_nob) {
 448                         saddr += this_nob;
 449                         soffset += this_nob;
 450                 } else {
 451                         kunmap(siov->bv_page);
 452                         saddr = NULL;
 453                         siov++;
 454                         nsiov--;
 455                         soffset = 0;
 456                 }
 457         } while (nob > 0);
 458
 459         if (daddr != NULL)
 460                 kunmap(diov->bv_page);
 461         if (saddr != NULL)
 462                 kunmap(siov->bv_page);
 463 }
 464 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 465
 466 void
 467 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 468                     unsigned int nkiov, struct bio_vec *kiov,
 469                     unsigned int kiovoffset,
 470                     unsigned int nob)
 471 {
 472         /* NB iov, kiov are READ-ONLY */
 473         unsigned int    this_nob;
 474         char           *addr = NULL;
 475
 476         if (nob == 0)
 477                 return;
 478
 479         LASSERT (!in_interrupt ());
 480
 481         LASSERT (niov > 0);
 482         while (iovoffset >= iov->iov_len) {
 483                 iovoffset -= iov->iov_len;
 484                 iov++;
 485                 niov--;
 486                 LASSERT(niov > 0);
 487         }
 488
 489         LASSERT(nkiov > 0);
 490         while (kiovoffset >= kiov->bv_len) {
 491                 kiovoffset -= kiov->bv_len;
 492                 kiov++;
 493                 nkiov--;
 494                 LASSERT(nkiov > 0);
 495         }
 496
 497         do {
 498                 LASSERT(niov > 0);
 499                 LASSERT(nkiov > 0);
 500                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 501                                 (unsigned int)kiov->bv_len - kiovoffset,
 502                                 nob);
 503
 504                 if (addr == NULL)
 505                         addr = ((char *)kmap(kiov->bv_page)) +
 506                                 kiov->bv_offset + kiovoffset;
 507
 508                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 509                 nob -= this_nob;
 510
 511                 if (iov->iov_len > iovoffset + this_nob) {
 512                         iovoffset += this_nob;
 513                 } else {
 514                         iov++;
 515                         niov--;
 516                         iovoffset = 0;
 517                 }
 518
 519                 if (kiov->bv_len > kiovoffset + this_nob) {
 520                         addr += this_nob;
 521                         kiovoffset += this_nob;
 522                 } else {
 523                         kunmap(kiov->bv_page);
 524                         addr = NULL;
 525                         kiov++;
 526                         nkiov--;
 527                         kiovoffset = 0;
 528                 }
 529
 530         } while (nob > 0);
 531
 532         if (addr != NULL)
 533                 kunmap(kiov->bv_page);
 534 }
 535 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 536
 537 void
 538 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
 539                    unsigned int kiovoffset,
 540                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 541                    unsigned int nob)
 542 {
 543         /* NB kiov, iov are READ-ONLY */
 544         unsigned int    this_nob;
 545         char           *addr = NULL;
 546
 547         if (nob == 0)
 548                 return;
 549
 550         LASSERT (!in_interrupt ());
 551
 552         LASSERT (nkiov > 0);
 553         while (kiovoffset >= kiov->bv_len) {
 554                 kiovoffset -= kiov->bv_len;
 555                 kiov++;
 556                 nkiov--;
 557                 LASSERT(nkiov > 0);
 558         }
 559
 560         LASSERT(niov > 0);
 561         while (iovoffset >= iov->iov_len) {
 562                 iovoffset -= iov->iov_len;
 563                 iov++;
 564                 niov--;
 565                 LASSERT(niov > 0);
 566         }
 567
 568         do {
 569                 LASSERT(nkiov > 0);
 570                 LASSERT(niov > 0);
 571                 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
 572                                 (unsigned int)iov->iov_len - iovoffset,
 573                                 nob);
 574
 575                 if (addr == NULL)
 576                         addr = ((char *)kmap(kiov->bv_page)) +
 577                                 kiov->bv_offset + kiovoffset;
 578
 579                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 580                 nob -= this_nob;
 581
 582                 if (kiov->bv_len > kiovoffset + this_nob) {
 583                         addr += this_nob;
 584                         kiovoffset += this_nob;
 585                 } else {
 586                         kunmap(kiov->bv_page);
 587                         addr = NULL;
 588                         kiov++;
 589                         nkiov--;
 590                         kiovoffset = 0;
 591                 }
 592
 593                 if (iov->iov_len > iovoffset + this_nob) {
 594                         iovoffset += this_nob;
 595                 } else {
 596                         iov++;
 597                         niov--;
 598                         iovoffset = 0;
 599                 }
 600         } while (nob > 0);
 601
 602         if (addr != NULL)
 603                 kunmap(kiov->bv_page);
 604 }
 605 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 606
 607 int
 608 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
 609                   int src_niov, struct bio_vec *src,
 610                   unsigned int offset, unsigned int len)
 611 {
 612         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 613          * for exactly 'len' bytes, and return the number of entries.
 614          * NB not destructive to 'src' */
 615         unsigned int    frag_len;
 616         unsigned int    niov;
 617
 618         if (len == 0)                           /* no data => */
 619                 return (0);                     /* no frags */
 620
 621         LASSERT(src_niov > 0);
 622         while (offset >= src->bv_len) {      /* skip initial frags */
 623                 offset -= src->bv_len;
 624                 src_niov--;
 625                 src++;
 626                 LASSERT(src_niov > 0);
 627         }
 628
 629         niov = 1;
 630         for (;;) {
 631                 LASSERT(src_niov > 0);
 632                 LASSERT((int)niov <= dst_niov);
 633
 634                 frag_len = src->bv_len - offset;
 635                 dst->bv_page = src->bv_page;
 636                 dst->bv_offset = src->bv_offset + offset;
 637
 638                 if (len <= frag_len) {
 639                         dst->bv_len = len;
 640                         LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 641                         return niov;
 642                 }
 643
 644                 dst->bv_len = frag_len;
 645                 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 646
 647                 len -= frag_len;
 648                 dst++;
 649                 src++;
 650                 niov++;
 651                 src_niov--;
 652                 offset = 0;
 653         }
 654 }
 655 EXPORT_SYMBOL(lnet_extract_kiov);
 656
 657 void
 658 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 659              int delayed, unsigned int offset, unsigned int mlen,
 660              unsigned int rlen)
 661 {
 662         unsigned int niov = 0;
 663         struct kvec *iov = NULL;
 664         struct bio_vec  *kiov = NULL;
 665         int rc;
 666
 667         LASSERT (!in_interrupt ());
 668         LASSERT (mlen == 0 || msg != NULL);
 669
 670         if (msg != NULL) {
 671                 LASSERT(msg->msg_receiving);
 672                 LASSERT(!msg->msg_sending);
 673                 LASSERT(rlen == msg->msg_len);
 674                 LASSERT(mlen <= msg->msg_len);
 675                 LASSERT(msg->msg_offset == offset);
 676                 LASSERT(msg->msg_wanted == mlen);
 677
 678                 msg->msg_receiving = 0;
 679
 680                 if (mlen != 0) {
 681                         niov = msg->msg_niov;
 682                         kiov = msg->msg_kiov;
 683
 684                         LASSERT (niov > 0);
 685                         LASSERT ((iov == NULL) != (kiov == NULL));
 686                 }
 687         }
 688
 689         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 690                                              niov, kiov, offset, mlen,
 691                                              rlen);
 692         if (rc < 0)
 693                 lnet_finalize(msg, rc);
 694 }
 695
 696 static void
 697 lnet_setpayloadbuffer(struct lnet_msg *msg)
 698 {
 699         struct lnet_libmd *md = msg->msg_md;
 700
 701         LASSERT(msg->msg_len > 0);
 702         LASSERT(!msg->msg_routing);
 703         LASSERT(md != NULL);
 704         LASSERT(msg->msg_niov == 0);
 705         LASSERT(msg->msg_kiov == NULL);
 706
 707         msg->msg_niov = md->md_niov;
 708         msg->msg_kiov = md->md_kiov;
 709 }
 710
 711 void
 712 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 713                unsigned int offset, unsigned int len)
 714 {
 715         msg->msg_type = type;
 716         msg->msg_target.pid = target.pid;
 717         lnet_nid4_to_nid(target.nid, &msg->msg_target.nid);
 718         msg->msg_len = len;
 719         msg->msg_offset = offset;
 720
 721         if (len != 0)
 722                 lnet_setpayloadbuffer(msg);
 723
 724         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 725         msg->msg_hdr.type           = cpu_to_le32(type);
 726         /* dest_nid will be overwritten by lnet_select_pathway() */
 727         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 728         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 729         /* src_nid will be set later */
 730         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 731         msg->msg_hdr.payload_length = cpu_to_le32(len);
 732 }
 733
 734 void
 735 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 736 {
 737         void *priv = msg->msg_private;
 738         int rc;
 739
 740         LASSERT(!in_interrupt());
 741         LASSERT(nid_is_lo0(&ni->ni_nid) ||
 742                 (msg->msg_txcredit && msg->msg_peertxcredit));
 743
 744         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 745         if (rc < 0) {
 746                 msg->msg_no_resend = true;
 747                 lnet_finalize(msg, rc);
 748         }
 749 }
 750
 751 static int
 752 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 753 {
 754         int     rc;
 755
 756         LASSERT(!msg->msg_sending);
 757         LASSERT(msg->msg_receiving);
 758         LASSERT(!msg->msg_rx_ready_delay);
 759         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 760
 761         msg->msg_rx_ready_delay = 1;
 762         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 763                                                   &msg->msg_private);
 764         if (rc != 0) {
 765                 CERROR("recv from %s / send to %s aborted: "
 766                        "eager_recv failed %d\n",
 767                        libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
 768                        libcfs_idstr(&msg->msg_target), rc);
 769                 LASSERT(rc < 0); /* required by my callers */
 770         }
 771
 772         return rc;
 773 }
 774
 775 static bool
 776 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 777 {
 778         time64_t deadline;
 779
 780         deadline = lpni->lpni_last_alive +
 781                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 782
 783         /*
 784          * assume peer_ni is alive as long as we're within the configured
 785          * peer timeout
 786          */
 787         if (deadline > now)
 788                 return false;
 789
 790         return true;
 791 }
 792
 793 /* NB: returns 1 when alive, 0 when dead, negative when error;
 794  *     may drop the lnet_net_lock */
 795 static int
 796 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 797                        struct lnet_msg *msg)
 798 {
 799         time64_t now = ktime_get_seconds();
 800
 801         if (!lnet_peer_aliveness_enabled(lpni))
 802                 return -ENODEV;
 803
 804         /*
 805          * If we're resending a message, let's attempt to send it even if
 806          * the peer is down to fulfill our resend quota on the message
 807          */
 808         if (msg->msg_retry_count > 0)
 809                 return 1;
 810
 811         /* try and send recovery messages irregardless */
 812         if (msg->msg_recovery)
 813                 return 1;
 814
 815         /* always send any responses */
 816         if (lnet_msg_is_response(msg))
 817                 return 1;
 818
 819         if (!lnet_is_peer_deadline_passed(lpni, now))
 820                 return true;
 821
 822         return lnet_is_peer_ni_alive(lpni);
 823 }
 824
 825 /**
 826  * \param msg The message to be sent.
 827  * \param do_send True if lnet_ni_send() should be called in this function.
 828  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 829  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 830  *
 831  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 832  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 833  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 834  * \retval -ECANCELED If the MD of the message has been unlinked.
 835  */
 836 static int
 837 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 838 {
 839         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 840         struct lnet_ni          *ni = msg->msg_txni;
 841         int                     cpt = msg->msg_tx_cpt;
 842         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 843
 844         /* non-lnet_send() callers have checked before */
 845         LASSERT(!do_send || msg->msg_tx_delayed);
 846         LASSERT(!msg->msg_receiving);
 847         LASSERT(msg->msg_tx_committed);
 848
 849         /* can't get here if we're sending to the loopback interface */
 850         if (the_lnet.ln_loni)
 851                 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
 852
 853         /* NB 'lp' is always the next hop */
 854         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 855             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 856                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 857                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 858                         msg->msg_len;
 859                 lnet_net_unlock(cpt);
 860                 if (msg->msg_txpeer)
 861                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 862                                         msg->msg_type,
 863                                         LNET_STATS_TYPE_DROP);
 864                 if (msg->msg_txni)
 865                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 866                                         msg->msg_type,
 867                                         LNET_STATS_TYPE_DROP);
 868
 869                 CNETERR("Dropping message for %s: peer not alive\n",
 870                         libcfs_idstr(&msg->msg_target));
 871                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 872                 if (do_send)
 873                         lnet_finalize(msg, -EHOSTUNREACH);
 874
 875                 lnet_net_lock(cpt);
 876                 return -EHOSTUNREACH;
 877         }
 878
 879         if (msg->msg_md != NULL &&
 880             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 881                 lnet_net_unlock(cpt);
 882
 883                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 884                         "called on the MD/ME.\n",
 885                         libcfs_idstr(&msg->msg_target));
 886                 if (do_send) {
 887                         msg->msg_no_resend = true;
 888                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 889                                msg, libcfs_idstr(&msg->msg_target));
 890                         lnet_finalize(msg, -ECANCELED);
 891                 }
 892
 893                 lnet_net_lock(cpt);
 894                 return -ECANCELED;
 895         }
 896
 897         if (!msg->msg_peertxcredit) {
 898                 spin_lock(&lp->lpni_lock);
 899                 LASSERT((lp->lpni_txcredits < 0) ==
 900                         !list_empty(&lp->lpni_txq));
 901
 902                 msg->msg_peertxcredit = 1;
 903                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 904                 lp->lpni_txcredits--;
 905
 906                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 907                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 908
 909                 if (lp->lpni_txcredits < 0) {
 910                         msg->msg_tx_delayed = 1;
 911                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 912                         spin_unlock(&lp->lpni_lock);
 913                         return LNET_CREDIT_WAIT;
 914                 }
 915                 spin_unlock(&lp->lpni_lock);
 916         }
 917
 918         if (!msg->msg_txcredit) {
 919                 LASSERT((tq->tq_credits < 0) ==
 920                         !list_empty(&tq->tq_delayed));
 921
 922                 msg->msg_txcredit = 1;
 923                 tq->tq_credits--;
 924                 atomic_dec(&ni->ni_tx_credits);
 925
 926                 if (tq->tq_credits < tq->tq_credits_min)
 927                         tq->tq_credits_min = tq->tq_credits;
 928
 929                 if (tq->tq_credits < 0) {
 930                         msg->msg_tx_delayed = 1;
 931                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 932                         return LNET_CREDIT_WAIT;
 933                 }
 934         }
 935
 936         if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
 937             lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
 938                 msg->msg_tx_delayed = 1;
 939                 return LNET_CREDIT_WAIT;
 940         }
 941
 942         /* unset the tx_delay flag as we're going to send it now */
 943         msg->msg_tx_delayed = 0;
 944
 945         if (do_send) {
 946                 lnet_net_unlock(cpt);
 947                 lnet_ni_send(ni, msg);
 948                 lnet_net_lock(cpt);
 949         }
 950         return LNET_CREDIT_OK;
 951 }
 952
 953
 954 static struct lnet_rtrbufpool *
 955 lnet_msg2bufpool(struct lnet_msg *msg)
 956 {
 957         struct lnet_rtrbufpool  *rbp;
 958         int                     cpt;
 959
 960         LASSERT(msg->msg_rx_committed);
 961
 962         cpt = msg->msg_rx_cpt;
 963         rbp = &the_lnet.ln_rtrpools[cpt][0];
 964
 965         LASSERT(msg->msg_len <= LNET_MTU);
 966         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 967                 rbp++;
 968                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 969         }
 970
 971         return rbp;
 972 }
 973
 974 static int
 975 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 976 {
 977         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 978          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 979          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 980          * received or OK to receive */
 981         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 982         struct lnet_peer *lp;
 983         struct lnet_rtrbufpool *rbp;
 984         struct lnet_rtrbuf *rb;
 985
 986         LASSERT(msg->msg_kiov == NULL);
 987         LASSERT(msg->msg_niov == 0);
 988         LASSERT(msg->msg_routing);
 989         LASSERT(msg->msg_receiving);
 990         LASSERT(!msg->msg_sending);
 991         LASSERT(lpni->lpni_peer_net);
 992         LASSERT(lpni->lpni_peer_net->lpn_peer);
 993
 994         lp = lpni->lpni_peer_net->lpn_peer;
 995
 996         /* non-lnet_parse callers only receive delayed messages */
 997         LASSERT(!do_recv || msg->msg_rx_delayed);
 998
 999         if (!msg->msg_peerrtrcredit) {
1000                 /* lpni_lock protects the credit manipulation */
1001                 spin_lock(&lpni->lpni_lock);
1002
1003                 msg->msg_peerrtrcredit = 1;
1004                 lpni->lpni_rtrcredits--;
1005                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1006                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1007
1008                 if (lpni->lpni_rtrcredits < 0) {
1009                         spin_unlock(&lpni->lpni_lock);
1010                         /* must have checked eager_recv before here */
1011                         LASSERT(msg->msg_rx_ready_delay);
1012                         msg->msg_rx_delayed = 1;
1013                         /* lp_lock protects the lp_rtrq */
1014                         spin_lock(&lp->lp_lock);
1015                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1016                         spin_unlock(&lp->lp_lock);
1017                         return LNET_CREDIT_WAIT;
1018                 }
1019                 spin_unlock(&lpni->lpni_lock);
1020         }
1021
1022         rbp = lnet_msg2bufpool(msg);
1023
1024         if (!msg->msg_rtrcredit) {
1025                 msg->msg_rtrcredit = 1;
1026                 rbp->rbp_credits--;
1027                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1028                         rbp->rbp_mincredits = rbp->rbp_credits;
1029
1030                 if (rbp->rbp_credits < 0) {
1031                         /* must have checked eager_recv before here */
1032                         LASSERT(msg->msg_rx_ready_delay);
1033                         msg->msg_rx_delayed = 1;
1034                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1035                         return LNET_CREDIT_WAIT;
1036                 }
1037         }
1038
1039         LASSERT(!list_empty(&rbp->rbp_bufs));
1040         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1041         list_del(&rb->rb_list);
1042
1043         msg->msg_niov = rbp->rbp_npages;
1044         msg->msg_kiov = &rb->rb_kiov[0];
1045
1046         /* unset the msg-rx_delayed flag since we're receiving the message */
1047         msg->msg_rx_delayed = 0;
1048
1049         if (do_recv) {
1050                 int cpt = msg->msg_rx_cpt;
1051
1052                 lnet_net_unlock(cpt);
1053                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1054                              0, msg->msg_len, msg->msg_len);
1055                 lnet_net_lock(cpt);
1056         }
1057         return LNET_CREDIT_OK;
1058 }
1059
1060 void
1061 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1062 {
1063         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1064         struct lnet_ni          *txni = msg->msg_txni;
1065         struct lnet_msg         *msg2;
1066
1067         if (msg->msg_txcredit) {
1068                 struct lnet_ni       *ni = msg->msg_txni;
1069                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1070
1071                 /* give back NI txcredits */
1072                 msg->msg_txcredit = 0;
1073
1074                 LASSERT((tq->tq_credits < 0) ==
1075                         !list_empty(&tq->tq_delayed));
1076
1077                 tq->tq_credits++;
1078                 atomic_inc(&ni->ni_tx_credits);
1079                 if (tq->tq_credits <= 0) {
1080                         msg2 = list_entry(tq->tq_delayed.next,
1081                                           struct lnet_msg, msg_list);
1082                         list_del(&msg2->msg_list);
1083
1084                         LASSERT(msg2->msg_txni == ni);
1085                         LASSERT(msg2->msg_tx_delayed);
1086                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1087
1088                         (void) lnet_post_send_locked(msg2, 1);
1089                 }
1090         }
1091
1092         if (msg->msg_peertxcredit) {
1093                 /* give back peer txcredits */
1094                 msg->msg_peertxcredit = 0;
1095
1096                 spin_lock(&txpeer->lpni_lock);
1097                 LASSERT((txpeer->lpni_txcredits < 0) ==
1098                         !list_empty(&txpeer->lpni_txq));
1099
1100                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1101                 LASSERT(txpeer->lpni_txqnob >= 0);
1102
1103                 txpeer->lpni_txcredits++;
1104                 if (txpeer->lpni_txcredits <= 0) {
1105                         int msg2_cpt;
1106
1107                         msg2 = list_entry(txpeer->lpni_txq.next,
1108                                               struct lnet_msg, msg_list);
1109                         list_del(&msg2->msg_list);
1110                         spin_unlock(&txpeer->lpni_lock);
1111
1112                         LASSERT(msg2->msg_txpeer == txpeer);
1113                         LASSERT(msg2->msg_tx_delayed);
1114
1115                         msg2_cpt = msg2->msg_tx_cpt;
1116
1117                         /*
1118                          * The msg_cpt can be different from the msg2_cpt
1119                          * so we need to make sure we lock the correct cpt
1120                          * for msg2.
1121                          * Once we call lnet_post_send_locked() it is no
1122                          * longer safe to access msg2, since it could've
1123                          * been freed by lnet_finalize(), but we still
1124                          * need to relock the correct cpt, so we cache the
1125                          * msg2_cpt for the purpose of the check that
1126                          * follows the call to lnet_pose_send_locked().
1127                          */
1128                         if (msg2_cpt != msg->msg_tx_cpt) {
1129                                 lnet_net_unlock(msg->msg_tx_cpt);
1130                                 lnet_net_lock(msg2_cpt);
1131                         }
1132                         (void) lnet_post_send_locked(msg2, 1);
1133                         if (msg2_cpt != msg->msg_tx_cpt) {
1134                                 lnet_net_unlock(msg2_cpt);
1135                                 lnet_net_lock(msg->msg_tx_cpt);
1136                         }
1137                 } else {
1138                         spin_unlock(&txpeer->lpni_lock);
1139                 }
1140         }
1141
1142         if (txni != NULL) {
1143                 msg->msg_txni = NULL;
1144                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1145         }
1146
1147         if (txpeer != NULL) {
1148                 msg->msg_txpeer = NULL;
1149                 lnet_peer_ni_decref_locked(txpeer);
1150         }
1151 }
1152
1153 void
1154 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1155 {
1156         struct lnet_msg *msg;
1157
1158         if (list_empty(&rbp->rbp_msgs))
1159                 return;
1160         msg = list_entry(rbp->rbp_msgs.next,
1161                          struct lnet_msg, msg_list);
1162         list_del(&msg->msg_list);
1163
1164         (void)lnet_post_routed_recv_locked(msg, 1);
1165 }
1166
1167 void
1168 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1169 {
1170         struct lnet_msg *msg;
1171         struct lnet_msg *tmp;
1172
1173         lnet_net_unlock(cpt);
1174
1175         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1176                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1177                              0, 0, 0, msg->msg_hdr.payload_length);
1178                 list_del_init(&msg->msg_list);
1179                 msg->msg_no_resend = true;
1180                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1181                 lnet_finalize(msg, -ECANCELED);
1182         }
1183
1184         lnet_net_lock(cpt);
1185 }
1186
1187 void
1188 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1189 {
1190         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1191         struct lnet_peer *lp;
1192         struct lnet_ni *rxni = msg->msg_rxni;
1193         struct lnet_msg *msg2;
1194
1195         if (msg->msg_rtrcredit) {
1196                 /* give back global router credits */
1197                 struct lnet_rtrbuf *rb;
1198                 struct lnet_rtrbufpool *rbp;
1199
1200                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1201                  * there until it gets one allocated, or aborts the wait
1202                  * itself */
1203                 LASSERT(msg->msg_kiov != NULL);
1204
1205                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1206                 rbp = rb->rb_pool;
1207
1208                 msg->msg_kiov = NULL;
1209                 msg->msg_rtrcredit = 0;
1210
1211                 LASSERT(rbp == lnet_msg2bufpool(msg));
1212
1213                 LASSERT((rbp->rbp_credits > 0) ==
1214                         !list_empty(&rbp->rbp_bufs));
1215
1216                 /* If routing is now turned off, we just drop this buffer and
1217                  * don't bother trying to return credits.  */
1218                 if (!the_lnet.ln_routing) {
1219                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1220                         goto routing_off;
1221                 }
1222
1223                 /* It is possible that a user has lowered the desired number of
1224                  * buffers in this pool.  Make sure we never put back
1225                  * more buffers than the stated number. */
1226                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1227                         /* Discard this buffer so we don't have too
1228                          * many. */
1229                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1230                         rbp->rbp_nbuffers--;
1231                 } else {
1232                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1233                         rbp->rbp_credits++;
1234                         if (rbp->rbp_credits <= 0)
1235                                 lnet_schedule_blocked_locked(rbp);
1236                 }
1237         }
1238
1239 routing_off:
1240         if (msg->msg_peerrtrcredit) {
1241                 LASSERT(rxpeerni);
1242                 LASSERT(rxpeerni->lpni_peer_net);
1243                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1244
1245                 /* give back peer router credits */
1246                 msg->msg_peerrtrcredit = 0;
1247
1248                 spin_lock(&rxpeerni->lpni_lock);
1249                 rxpeerni->lpni_rtrcredits++;
1250                 spin_unlock(&rxpeerni->lpni_lock);
1251
1252                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1253                 spin_lock(&lp->lp_lock);
1254
1255                 /* drop all messages which are queued to be routed on that
1256                  * peer. */
1257                 if (!the_lnet.ln_routing) {
1258                         LIST_HEAD(drop);
1259                         list_splice_init(&lp->lp_rtrq, &drop);
1260                         spin_unlock(&lp->lp_lock);
1261                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1262                 } else if (!list_empty(&lp->lp_rtrq)) {
1263                         int msg2_cpt;
1264
1265                         msg2 = list_entry(lp->lp_rtrq.next,
1266                                           struct lnet_msg, msg_list);
1267                         list_del(&msg2->msg_list);
1268                         msg2_cpt = msg2->msg_rx_cpt;
1269                         spin_unlock(&lp->lp_lock);
1270                         /*
1271                          * messages on the lp_rtrq can be from any NID in
1272                          * the peer, which means they might have different
1273                          * cpts. We need to make sure we lock the right
1274                          * one.
1275                          */
1276                         if (msg2_cpt != msg->msg_rx_cpt) {
1277                                 lnet_net_unlock(msg->msg_rx_cpt);
1278                                 lnet_net_lock(msg2_cpt);
1279                         }
1280                         (void) lnet_post_routed_recv_locked(msg2, 1);
1281                         if (msg2_cpt != msg->msg_rx_cpt) {
1282                                 lnet_net_unlock(msg2_cpt);
1283                                 lnet_net_lock(msg->msg_rx_cpt);
1284                         }
1285                 } else {
1286                         spin_unlock(&lp->lp_lock);
1287                 }
1288         }
1289         if (rxni != NULL) {
1290                 msg->msg_rxni = NULL;
1291                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1292         }
1293         if (rxpeerni != NULL) {
1294                 msg->msg_rxpeer = NULL;
1295                 lnet_peer_ni_decref_locked(rxpeerni);
1296         }
1297 }
1298
1299 static struct lnet_peer_ni *
1300 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1301                     struct lnet_peer *peer,
1302                     struct lnet_peer_ni *best_lpni,
1303                     struct lnet_peer_net *peer_net)
1304 {
1305         /*
1306          * Look at the peer NIs for the destination peer that connect
1307          * to the chosen net. If a peer_ni is preferred when using the
1308          * best_ni to communicate, we use that one. If there is no
1309          * preferred peer_ni, or there are multiple preferred peer_ni,
1310          * the available transmit credits are used. If the transmit
1311          * credits are equal, we round-robin over the peer_ni.
1312          */
1313         struct lnet_peer_ni *lpni = NULL;
1314         int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1315                 INT_MIN;
1316         int best_lpni_healthv = (best_lpni) ?
1317                 atomic_read(&best_lpni->lpni_healthv) : 0;
1318         bool best_lpni_is_preferred = false;
1319         bool lpni_is_preferred;
1320         int lpni_healthv;
1321         __u32 lpni_sel_prio;
1322         __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1323
1324         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1325                 /*
1326                  * if the best_ni we've chosen aleady has this lpni
1327                  * preferred, then let's use it
1328                  */
1329                 if (best_ni) {
1330                         lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1331                                 lpni, &best_ni->ni_nid);
1332                         CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1333                                libcfs_nidstr(&best_ni->ni_nid),
1334                                lpni_is_preferred);
1335                 } else {
1336                         lpni_is_preferred = false;
1337                 }
1338
1339                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1340                 lpni_sel_prio = lpni->lpni_sel_priority;
1341
1342                 if (best_lpni)
1343                         CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1344                                 libcfs_nidstr(&lpni->lpni_nid),
1345                                 libcfs_nidstr(&best_lpni->lpni_nid),
1346                                 lpni_healthv, best_lpni_healthv,
1347                                 lpni_sel_prio, best_sel_prio,
1348                                 lpni->lpni_txcredits, best_lpni_credits,
1349                                 lpni->lpni_seq, best_lpni->lpni_seq);
1350                 else
1351                         goto select_lpni;
1352
1353                 /* pick the healthiest peer ni */
1354                 if (lpni_healthv < best_lpni_healthv)
1355                         continue;
1356                 else if (lpni_healthv > best_lpni_healthv) {
1357                         if (best_lpni_is_preferred)
1358                                 best_lpni_is_preferred = false;
1359                         goto select_lpni;
1360                 }
1361
1362                 if (lpni_sel_prio > best_sel_prio)
1363                         continue;
1364                 else if (lpni_sel_prio < best_sel_prio) {
1365                         if (best_lpni_is_preferred)
1366                                 best_lpni_is_preferred = false;
1367                         goto select_lpni;
1368                 }
1369
1370                 /* if this is a preferred peer use it */
1371                 if (!best_lpni_is_preferred && lpni_is_preferred) {
1372                         best_lpni_is_preferred = true;
1373                         goto select_lpni;
1374                 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1375                         /* this is not the preferred peer so let's ignore
1376                          * it.
1377                          */
1378                         continue;
1379                 }
1380
1381                 if (lpni->lpni_txcredits < best_lpni_credits)
1382                         /* We already have a peer that has more credits
1383                          * available than this one. No need to consider
1384                          * this peer further.
1385                          */
1386                         continue;
1387                 else if (lpni->lpni_txcredits > best_lpni_credits)
1388                         goto select_lpni;
1389
1390                 /* The best peer found so far and the current peer
1391                  * have the same number of available credits let's
1392                  * make sure to select between them using Round Robin
1393                  */
1394                 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1395                         continue;
1396 select_lpni:
1397                 best_lpni_is_preferred = lpni_is_preferred;
1398                 best_lpni_healthv = lpni_healthv;
1399                 best_sel_prio = lpni_sel_prio;
1400                 best_lpni = lpni;
1401                 best_lpni_credits = lpni->lpni_txcredits;
1402         }
1403
1404         /* if we still can't find a peer ni then we can't reach it */
1405         if (!best_lpni) {
1406                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1407                         LNET_NIDNET(dst_nid);
1408                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1409                                 libcfs_net2str(net_id));
1410                 return NULL;
1411         }
1412
1413         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1414                libcfs_nidstr(&best_lpni->lpni_nid));
1415
1416         return best_lpni;
1417 }
1418
1419 /*
1420  * Prerequisite: the best_ni should already be set in the sd
1421  * Find the best lpni.
1422  * If the net id is provided then restrict lpni selection on
1423  * that particular net.
1424  * Otherwise find any reachable lpni. When dealing with an MR
1425  * gateway and it has multiple lpnis which we can use
1426  * we want to select the best one from the list of reachable
1427  * ones.
1428  */
1429 static inline struct lnet_peer_ni *
1430 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1431                     struct lnet_peer *peer, __u32 net_id)
1432 {
1433         struct lnet_peer_net *peer_net;
1434
1435         /* find the best_lpni on any local network */
1436         if (net_id == LNET_NET_ANY) {
1437                 struct lnet_peer_ni *best_lpni = NULL;
1438                 struct lnet_peer_net *lpn;
1439                 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1440                         /* no net specified find any reachable peer ni */
1441                         if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1442                                 continue;
1443                         best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1444                                                         best_lpni, lpn);
1445                 }
1446
1447                 return best_lpni;
1448         }
1449         /* restrict on the specified net */
1450         peer_net = lnet_peer_get_net_locked(peer, net_id);
1451         if (peer_net)
1452                 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1453
1454         return NULL;
1455 }
1456
1457 static int
1458 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1459 {
1460         if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1461                 return 1;
1462
1463         if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1464                 return -1;
1465
1466         if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1467                 return 1;
1468
1469         if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1470                 return -1;
1471
1472         return 0;
1473 }
1474
1475 /* Compare route priorities and hop counts */
1476 static int
1477 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1478 {
1479         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1480         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1481
1482         if (r1->lr_priority < r2->lr_priority)
1483                 return 1;
1484
1485         if (r1->lr_priority > r2->lr_priority)
1486                 return -1;
1487
1488         if (r1_hops < r2_hops)
1489                 return 1;
1490
1491         if (r1_hops > r2_hops)
1492                 return -1;
1493
1494         return 0;
1495 }
1496
1497 static struct lnet_route *
1498 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1499                        struct lnet_peer_ni *remote_lpni,
1500                        struct lnet_route **prev_route,
1501                        struct lnet_peer_ni **gwni)
1502 {
1503         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1504         struct lnet_route *best_route;
1505         struct lnet_route *last_route;
1506         struct lnet_route *route;
1507         int rc;
1508         bool best_rte_is_preferred = false;
1509         struct lnet_nid *gw_pnid;
1510
1511         CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1512                libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1513
1514         best_route = last_route = NULL;
1515         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1516                 if (!lnet_is_route_alive(route))
1517                         continue;
1518                 gw_pnid = &route->lr_gateway->lp_primary_nid;
1519
1520                 /* no protection on below fields, but it's harmless */
1521                 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1522                         last_route = route;
1523
1524                 /* if the best route found is in the preferred list then
1525                  * tag it as preferred and use it later on. But if we
1526                  * didn't find any routes which are on the preferred list
1527                  * then just use the best route possible.
1528                  */
1529                 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1530
1531                 if (!best_route || (rc && !best_rte_is_preferred)) {
1532                         /* Restrict the selection of the router NI on the
1533                          * src_net provided. If the src_net is LNET_NID_ANY,
1534                          * then select the best interface available.
1535                          */
1536                         lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1537                                                    route->lr_gateway,
1538                                                    src_net);
1539                         if (!lpni) {
1540                                 CDEBUG(D_NET,
1541                                        "Gateway %s does not have a peer NI on net %s\n",
1542                                        libcfs_nidstr(gw_pnid),
1543                                        libcfs_net2str(src_net));
1544                                 continue;
1545                         }
1546                 }
1547
1548                 if (rc && !best_rte_is_preferred) {
1549                         /* This is the first preferred route we found,
1550                          * so it beats any route found previously
1551                          */
1552                         best_route = route;
1553                         if (!last_route)
1554                                 last_route = route;
1555                         best_gw_ni = lpni;
1556                         best_rte_is_preferred = true;
1557                         CDEBUG(D_NET, "preferred gw = %s\n",
1558                                libcfs_nidstr(gw_pnid));
1559                         continue;
1560                 } else if ((!rc) && best_rte_is_preferred)
1561                         /* The best route we found so far is in the preferred
1562                          * list, so it beats any non-preferred route
1563                          */
1564                         continue;
1565
1566                 if (!best_route) {
1567                         best_route = last_route = route;
1568                         best_gw_ni = lpni;
1569                         continue;
1570                 }
1571
1572                 rc = lnet_compare_routes(route, best_route);
1573                 if (rc == -1)
1574                         continue;
1575
1576                 /* Restrict the selection of the router NI on the
1577                  * src_net provided. If the src_net is LNET_NID_ANY,
1578                  * then select the best interface available.
1579                  */
1580                 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1581                                            route->lr_gateway,
1582                                            src_net);
1583                 if (!lpni) {
1584                         CDEBUG(D_NET,
1585                                "Gateway %s does not have a peer NI on net %s\n",
1586                                libcfs_nidstr(gw_pnid),
1587                                libcfs_net2str(src_net));
1588                         continue;
1589                 }
1590
1591                 if (rc == 1) {
1592                         best_route = route;
1593                         best_gw_ni = lpni;
1594                         continue;
1595                 }
1596
1597                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1598                 if (rc == -1)
1599                         continue;
1600
1601                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1602                         best_route = route;
1603                         best_gw_ni = lpni;
1604                         continue;
1605                 }
1606         }
1607
1608         *prev_route = last_route;
1609         *gwni = best_gw_ni;
1610
1611         return best_route;
1612 }
1613
1614 static inline unsigned int
1615 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1616 {
1617         if (dev_idx == UINT_MAX)
1618                 return UINT_MAX;
1619
1620         if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1621             !ni->ni_net->net_lnd->lnd_get_dev_prio)
1622                 return UINT_MAX;
1623
1624         return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1625 }
1626
1627 static struct lnet_ni *
1628 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1629                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1630                  struct lnet_msg *msg, int md_cpt)
1631 {
1632         struct lnet_libmd *md = msg->msg_md;
1633         unsigned int offset = msg->msg_offset;
1634         unsigned int shortest_distance;
1635         struct lnet_ni *ni = NULL;
1636         int best_credits;
1637         int best_healthv;
1638         __u32 best_sel_prio;
1639         unsigned int best_dev_prio;
1640         unsigned int dev_idx = UINT_MAX;
1641         struct page *page = lnet_get_first_page(md, offset);
1642         msg->msg_rdma_force = lnet_is_rdma_only_page(page);
1643
1644         if (msg->msg_rdma_force)
1645                 dev_idx = lnet_get_dev_idx(page);
1646
1647         /*
1648          * If there is no peer_ni that we can send to on this network,
1649          * then there is no point in looking for a new best_ni here.
1650         */
1651         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1652                 return best_ni;
1653
1654         if (best_ni == NULL) {
1655                 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1656                 shortest_distance = UINT_MAX;
1657                 best_dev_prio = UINT_MAX;
1658                 best_credits = INT_MIN;
1659                 best_healthv = 0;
1660         } else {
1661                 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1662                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1663                                                      best_ni->ni_dev_cpt);
1664                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1665                 best_healthv = atomic_read(&best_ni->ni_healthv);
1666                 best_sel_prio = best_ni->ni_sel_priority;
1667         }
1668
1669         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1670                 unsigned int distance;
1671                 int ni_credits;
1672                 int ni_healthv;
1673                 int ni_fatal;
1674                 __u32 ni_sel_prio;
1675                 unsigned int ni_dev_prio;
1676
1677                 ni_credits = atomic_read(&ni->ni_tx_credits);
1678                 ni_healthv = atomic_read(&ni->ni_healthv);
1679                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1680                 ni_sel_prio = ni->ni_sel_priority;
1681
1682                 /*
1683                  * calculate the distance from the CPT on which
1684                  * the message memory is allocated to the CPT of
1685                  * the NI's physical device
1686                  */
1687                 distance = cfs_cpt_distance(lnet_cpt_table(),
1688                                             md_cpt,
1689                                             ni->ni_dev_cpt);
1690
1691                 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1692
1693                 /*
1694                  * All distances smaller than the NUMA range
1695                  * are treated equally.
1696                  */
1697                 if (distance < lnet_numa_range)
1698                         distance = lnet_numa_range;
1699
1700                 /*
1701                  * Select on health, selection policy, direct dma prio,
1702                  * shorter distance, available credits, then round-robin.
1703                  */
1704                 if (ni_fatal)
1705                         continue;
1706
1707                 if (best_ni)
1708                         CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u, g:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u, g:%u]\n",
1709                                libcfs_nidstr(&ni->ni_nid), ni_credits, distance,
1710                                ni->ni_seq, ni_sel_prio, ni_dev_prio,
1711                                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1712                                : "not selected", best_credits, shortest_distance,
1713                                (best_ni) ? best_ni->ni_seq : 0,
1714                                best_sel_prio, best_dev_prio);
1715                 else
1716                         goto select_ni;
1717
1718                 if (ni_healthv < best_healthv)
1719                         continue;
1720                 else if (ni_healthv > best_healthv)
1721                         goto select_ni;
1722
1723                 if (ni_sel_prio > best_sel_prio)
1724                         continue;
1725                 else if (ni_sel_prio < best_sel_prio)
1726                         goto select_ni;
1727
1728                 if (ni_dev_prio > best_dev_prio)
1729                         continue;
1730                 else if (ni_dev_prio < best_dev_prio)
1731                         goto select_ni;
1732
1733                 if (distance > shortest_distance)
1734                         continue;
1735                 else if (distance < shortest_distance)
1736                         goto select_ni;
1737
1738                 if (ni_credits < best_credits)
1739                         continue;
1740                 else if (ni_credits > best_credits)
1741                         goto select_ni;
1742
1743                 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1744                         continue;
1745
1746 select_ni:
1747                 best_sel_prio = ni_sel_prio;
1748                 best_dev_prio = ni_dev_prio;
1749                 shortest_distance = distance;
1750                 best_healthv = ni_healthv;
1751                 best_ni = ni;
1752                 best_credits = ni_credits;
1753         }
1754
1755         CDEBUG(D_NET, "selected best_ni %s\n",
1756                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1757
1758         return best_ni;
1759 }
1760
1761 /*
1762  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1763  * because such traffic is required to perform discovery. We therefore
1764  * exclude all GET and PUT on that portal. We also exclude all ACK and
1765  * REPLY traffic, but that is because the portal is not tracked in the
1766  * message structure for these message types. We could restrict this
1767  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1768  */
1769 static bool
1770 lnet_msg_discovery(struct lnet_msg *msg)
1771 {
1772         if (msg->msg_type == LNET_MSG_PUT) {
1773                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1774                         return true;
1775         } else if (msg->msg_type == LNET_MSG_GET) {
1776                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1777                         return true;
1778         }
1779         return false;
1780 }
1781
1782 #define SRC_SPEC        0x0001
1783 #define SRC_ANY         0x0002
1784 #define LOCAL_DST       0x0004
1785 #define REMOTE_DST      0x0008
1786 #define MR_DST          0x0010
1787 #define NMR_DST         0x0020
1788 #define SND_RESP        0x0040
1789
1790 /* The following to defines are used for return codes */
1791 #define REPEAT_SEND     0x1000
1792 #define PASS_THROUGH    0x2000
1793
1794 /* The different cases lnet_select pathway needs to handle */
1795 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1796 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1797 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1798 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1799 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1800 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1801 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1802 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1803
1804 static int
1805 lnet_handle_lo_send(struct lnet_send_data *sd)
1806 {
1807         struct lnet_msg *msg = sd->sd_msg;
1808         int cpt = sd->sd_cpt;
1809
1810         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1811                 return -ESHUTDOWN;
1812
1813         /* No send credit hassles with LOLND */
1814         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1815         msg->msg_hdr.dest_nid =
1816                 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1817         if (!msg->msg_routing)
1818                 msg->msg_hdr.src_nid =
1819                         cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1820         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1821         lnet_msg_commit(msg, cpt);
1822         msg->msg_txni = the_lnet.ln_loni;
1823
1824         return LNET_CREDIT_OK;
1825 }
1826
1827 static int
1828 lnet_handle_send(struct lnet_send_data *sd)
1829 {
1830         struct lnet_ni *best_ni = sd->sd_best_ni;
1831         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1832         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1833         struct lnet_msg *msg = sd->sd_msg;
1834         int cpt2;
1835         __u32 send_case = sd->sd_send_case;
1836         int rc;
1837         __u32 routing = send_case & REMOTE_DST;
1838          struct lnet_rsp_tracker *rspt;
1839
1840         /* Increment sequence number of the selected peer, peer net,
1841          * local ni and local net so that we pick the next ones
1842          * in Round Robin.
1843          */
1844         best_lpni->lpni_peer_net->lpn_seq++;
1845         best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1846         best_ni->ni_net->net_seq++;
1847         best_ni->ni_seq = best_ni->ni_net->net_seq;
1848
1849         CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1850                libcfs_nidstr(&best_ni->ni_nid),
1851                best_ni->ni_seq, best_ni->ni_net->net_seq,
1852                atomic_read(&best_ni->ni_tx_credits),
1853                best_ni->ni_sel_priority,
1854                libcfs_nidstr(&best_lpni->lpni_nid),
1855                best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1856                best_lpni->lpni_txcredits,
1857                best_lpni->lpni_sel_priority);
1858
1859         /*
1860          * grab a reference on the peer_ni so it sticks around even if
1861          * we need to drop and relock the lnet_net_lock below.
1862          */
1863         lnet_peer_ni_addref_locked(best_lpni);
1864
1865         /*
1866          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1867          * message. This ensures that we get a CPT that is correct for
1868          * the NI when the NI has been restricted to a subset of all CPTs.
1869          * If the selected CPT differs from the one currently locked, we
1870          * must unlock and relock the lnet_net_lock(), and then check whether
1871          * the configuration has changed. We don't have a hold on the best_ni
1872          * yet, and it may have vanished.
1873          */
1874         cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1875         if (sd->sd_cpt != cpt2) {
1876                 __u32 seq = lnet_get_dlc_seq_locked();
1877                 lnet_net_unlock(sd->sd_cpt);
1878                 sd->sd_cpt = cpt2;
1879                 lnet_net_lock(sd->sd_cpt);
1880                 if (seq != lnet_get_dlc_seq_locked()) {
1881                         lnet_peer_ni_decref_locked(best_lpni);
1882                         return REPEAT_SEND;
1883                 }
1884         }
1885
1886         /*
1887          * store the best_lpni in the message right away to avoid having
1888          * to do the same operation under different conditions
1889          */
1890         msg->msg_txpeer = best_lpni;
1891         msg->msg_txni = best_ni;
1892
1893         /*
1894          * grab a reference for the best_ni since now it's in use in this
1895          * send. The reference will be dropped in lnet_finalize()
1896          */
1897         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1898
1899         /*
1900          * Always set the target.nid to the best peer picked. Either the
1901          * NID will be one of the peer NIDs selected, or the same NID as
1902          * what was originally set in the target or it will be the NID of
1903          * a router if this message should be routed
1904          */
1905         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1906
1907         /*
1908          * lnet_msg_commit assigns the correct cpt to the message, which
1909          * is used to decrement the correct refcount on the ni when it's
1910          * time to return the credits
1911          */
1912         lnet_msg_commit(msg, sd->sd_cpt);
1913
1914         /*
1915          * If we are routing the message then we keep the src_nid that was
1916          * set by the originator. If we are not routing then we are the
1917          * originator and set it here.
1918          */
1919         if (!msg->msg_routing)
1920                 msg->msg_hdr.src_nid =
1921                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txni->ni_nid));
1922
1923         if (routing) {
1924                 msg->msg_target_is_router = 1;
1925                 msg->msg_target.pid = LNET_PID_LUSTRE;
1926                 /*
1927                  * since we're routing we want to ensure that the
1928                  * msg_hdr.dest_nid is set to the final destination. When
1929                  * the router receives this message it knows how to route
1930                  * it.
1931                  *
1932                  * final_dst_lpni is set at the beginning of the
1933                  * lnet_select_pathway() function and is never changed.
1934                  * It's safe to use it here.
1935                  */
1936                 /* FIXME handle large-addr nid */
1937                 msg->msg_hdr.dest_nid =
1938                         cpu_to_le64(lnet_nid_to_nid4(&final_dst_lpni->lpni_nid));
1939         } else {
1940                 /*
1941                  * if we're not routing set the dest_nid to the best peer
1942                  * ni NID that we picked earlier in the algorithm.
1943                  */
1944                 msg->msg_hdr.dest_nid =
1945                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid));
1946         }
1947
1948         /*
1949          * if we have response tracker block update it with the next hop
1950          * nid
1951          */
1952         if (msg->msg_md) {
1953                 rspt = msg->msg_md->md_rspt_ptr;
1954                 if (rspt) {
1955                         rspt->rspt_next_hop_nid =
1956                                 msg->msg_txpeer->lpni_nid;
1957                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1958                                libcfs_nidstr(&rspt->rspt_next_hop_nid));
1959                 }
1960         }
1961
1962         rc = lnet_post_send_locked(msg, 0);
1963
1964         if (!rc)
1965                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1966                        libcfs_nid2str(msg->msg_hdr.src_nid),
1967                        libcfs_nidstr(&msg->msg_txni->ni_nid),
1968                        libcfs_nid2str(sd->sd_src_nid),
1969                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1970                        libcfs_nid2str(sd->sd_dst_nid),
1971                        libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1972                        libcfs_nid2str(sd->sd_rtr_nid),
1973                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1974
1975         return rc;
1976 }
1977
1978 static inline void
1979 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1980                          struct lnet_msg *msg)
1981 {
1982         if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1983             !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1984                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1985                        libcfs_nidstr(&lni->ni_nid),
1986                        libcfs_nidstr(&lpni->lpni_nid));
1987                 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
1988         }
1989 }
1990
1991 /*
1992  * Source Specified
1993  * Local Destination
1994  * non-mr peer
1995  *
1996  * use the source and destination NIDs as the pathway
1997  */
1998 static int
1999 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2000 {
2001         /* the destination lpni is set before we get here. */
2002
2003         /* find local NI */
2004         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2005         if (!sd->sd_best_ni) {
2006                 CERROR("Can't send to %s: src %s is not a "
2007                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2008                                 libcfs_nid2str(sd->sd_src_nid));
2009                 return -EINVAL;
2010         }
2011
2012         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2013
2014         return lnet_handle_send(sd);
2015 }
2016
2017 /*
2018  * Source Specified
2019  * Local Destination
2020  * MR Peer
2021  *
2022  * Don't run the selection algorithm on the peer NIs. By specifying the
2023  * local NID, we're also saying that we should always use the destination NID
2024  * provided. This handles the case where we should be using the same
2025  * destination NID for the all the messages which belong to the same RPC
2026  * request.
2027  */
2028 static int
2029 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2030 {
2031         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2032         if (!sd->sd_best_ni) {
2033                 CERROR("Can't send to %s: src %s is not a "
2034                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2035                                 libcfs_nid2str(sd->sd_src_nid));
2036                 return -EINVAL;
2037         }
2038
2039         if (sd->sd_best_lpni &&
2040             nid_same(&sd->sd_best_lpni->lpni_nid,
2041                       &the_lnet.ln_loni->ni_nid))
2042                 return lnet_handle_lo_send(sd);
2043         else if (sd->sd_best_lpni)
2044                 return lnet_handle_send(sd);
2045
2046         CERROR("can't send to %s. no NI on %s\n",
2047                libcfs_nid2str(sd->sd_dst_nid),
2048                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2049
2050         return -EHOSTUNREACH;
2051 }
2052
2053 struct lnet_ni *
2054 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2055                               struct lnet_peer *peer,
2056                               struct lnet_peer_net *peer_net,
2057                               struct lnet_msg *msg,
2058                               int cpt)
2059 {
2060         struct lnet_net *local_net;
2061         struct lnet_ni *best_ni;
2062
2063         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2064         if (!local_net)
2065                 return NULL;
2066
2067         /*
2068          * Iterate through the NIs in this local Net and select
2069          * the NI to send from. The selection is determined by
2070          * these 3 criterion in the following priority:
2071          *      1. NUMA
2072          *      2. NI available credits
2073          *      3. Round Robin
2074          */
2075         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2076                                    peer, peer_net, msg, cpt);
2077
2078         return best_ni;
2079 }
2080
2081 static int
2082 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2083                              int cpt)
2084 {
2085         struct lnet_peer *peer;
2086         struct lnet_peer_ni *new_lpni;
2087         int rc;
2088
2089         lnet_peer_ni_addref_locked(lpni);
2090
2091         peer = lpni->lpni_peer_net->lpn_peer;
2092
2093         if (lnet_peer_gw_discovery(peer)) {
2094                 lnet_peer_ni_decref_locked(lpni);
2095                 return 0;
2096         }
2097
2098         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2099                 lnet_peer_ni_decref_locked(lpni);
2100                 return 0;
2101         }
2102
2103         rc = lnet_discover_peer_locked(lpni, cpt, false);
2104         if (rc) {
2105                 lnet_peer_ni_decref_locked(lpni);
2106                 return rc;
2107         }
2108
2109         new_lpni = lnet_find_peer_ni_locked(lnet_nid_to_nid4(&lpni->lpni_nid));
2110         if (!new_lpni) {
2111                 lnet_peer_ni_decref_locked(lpni);
2112                 return -ENOENT;
2113         }
2114
2115         peer = new_lpni->lpni_peer_net->lpn_peer;
2116         spin_lock(&peer->lp_lock);
2117         if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2118                 /* The peer NI did not change and the peer is up to date.
2119                  * Nothing more to do.
2120                  */
2121                 spin_unlock(&peer->lp_lock);
2122                 lnet_peer_ni_decref_locked(lpni);
2123                 lnet_peer_ni_decref_locked(new_lpni);
2124                 return 0;
2125         }
2126         spin_unlock(&peer->lp_lock);
2127
2128         /* Either the peer NI changed during discovery, or the peer isn't up
2129          * to date. In both cases we want to queue the message on the
2130          * (possibly new) peer's pending queue and queue the peer for discovery
2131          */
2132         msg->msg_sending = 0;
2133         msg->msg_txpeer = NULL;
2134         lnet_net_unlock(cpt);
2135         lnet_peer_queue_message(peer, msg);
2136         lnet_net_lock(cpt);
2137
2138         lnet_peer_ni_decref_locked(lpni);
2139         lnet_peer_ni_decref_locked(new_lpni);
2140
2141         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2142                msg, libcfs_nidstr(&peer->lp_primary_nid));
2143
2144         return LNET_DC_WAIT;
2145 }
2146
2147 static int
2148 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2149                              lnet_nid_t dst_nid,
2150                              struct lnet_peer_ni **gw_lpni,
2151                              struct lnet_peer **gw_peer)
2152 {
2153         int rc;
2154         struct lnet_peer *gw;
2155         struct lnet_peer *lp;
2156         struct lnet_peer_net *lpn;
2157         struct lnet_peer_net *best_lpn = NULL;
2158         struct lnet_remotenet *rnet, *best_rnet = NULL;
2159         struct lnet_route *best_route = NULL;
2160         struct lnet_route *last_route = NULL;
2161         struct lnet_peer_ni *lpni = NULL;
2162         struct lnet_peer_ni *gwni = NULL;
2163         bool route_found = false;
2164         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2165                 (sd->sd_best_ni != NULL)
2166                 ? lnet_nid_to_nid4(&sd->sd_best_ni->ni_nid)
2167                 : LNET_NID_ANY;
2168         int best_lpn_healthv = 0;
2169         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2170
2171         CDEBUG(D_NET, "using src nid %s for route restriction\n",
2172                libcfs_nid2str(src_nid));
2173
2174         /* If a router nid was specified then we are replying to a GET or
2175          * sending an ACK. In this case we use the gateway associated with the
2176          * specified router nid.
2177          */
2178         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2179                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2180                 if (gwni) {
2181                         gw = gwni->lpni_peer_net->lpn_peer;
2182                         lnet_peer_ni_decref_locked(gwni);
2183                         if (gw->lp_rtr_refcount)
2184                                 route_found = true;
2185                 } else {
2186                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2187                                libcfs_nid2str(sd->sd_rtr_nid));
2188                 }
2189         }
2190
2191         if (!route_found) {
2192                 if (sd->sd_msg->msg_routing || src_nid != LNET_NID_ANY) {
2193                         /* If I'm routing this message then I need to find the
2194                          * next hop based on the destination NID
2195                          *
2196                          * We also find next hop based on the destination NID
2197                          * if the source NI was specified
2198                          */
2199                         best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2200                         if (!best_rnet) {
2201                                 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2202                                        src_nid != LNET_NID_ANY ?
2203                                                 libcfs_nid2str(src_nid) :
2204                                                 "any local NI",
2205                                        libcfs_nid2str(sd->sd_dst_nid));
2206                                 return -EHOSTUNREACH;
2207                         }
2208                 } else {
2209                         /* we've already looked up the initial lpni using
2210                          * dst_nid
2211                          */
2212                         lpni = sd->sd_best_lpni;
2213                         /* the peer tree must be in existence */
2214                         LASSERT(lpni && lpni->lpni_peer_net &&
2215                                 lpni->lpni_peer_net->lpn_peer);
2216                         lp = lpni->lpni_peer_net->lpn_peer;
2217
2218                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2219                                 /* is this remote network reachable?  */
2220                                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2221                                 if (!rnet)
2222                                         continue;
2223
2224                                 if (!best_lpn) {
2225                                         best_lpn = lpn;
2226                                         best_rnet = rnet;
2227                                 }
2228
2229                                 /* select the preferred peer net */
2230                                 if (best_lpn_healthv > lpn->lpn_healthv)
2231                                         continue;
2232                                 else if (best_lpn_healthv < lpn->lpn_healthv)
2233                                         goto use_lpn;
2234
2235                                 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2236                                         continue;
2237                                 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2238                                         goto use_lpn;
2239
2240                                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2241                                         continue;
2242 use_lpn:
2243                                 best_lpn_healthv = lpn->lpn_healthv;
2244                                 best_lpn_sel_prio = lpn->lpn_sel_priority;
2245                                 best_lpn = lpn;
2246                                 best_rnet = rnet;
2247                         }
2248
2249                         if (!best_lpn) {
2250                                 CERROR("peer %s has no available nets\n",
2251                                        libcfs_nid2str(sd->sd_dst_nid));
2252                                 return -EHOSTUNREACH;
2253                         }
2254
2255                         sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2256                                                                sd->sd_dst_nid,
2257                                                                lp,
2258                                                                best_lpn->lpn_net_id);
2259                         if (!sd->sd_best_lpni) {
2260                                 CERROR("peer %s is unreachable\n",
2261                                        libcfs_nid2str(sd->sd_dst_nid));
2262                                 return -EHOSTUNREACH;
2263                         }
2264
2265                         /* We're attempting to round robin over the remote peer
2266                          * NI's so update the final destination we selected
2267                          */
2268                         sd->sd_final_dst_lpni = sd->sd_best_lpni;
2269
2270                         /* Increment the sequence number of the remote lpni so
2271                          * we can round robin over the different interfaces of
2272                          * the remote lpni
2273                          */
2274                         sd->sd_best_lpni->lpni_seq++;
2275                 }
2276
2277                 /*
2278                  * find the best route. Restrict the selection on the net of the
2279                  * local NI if we've already picked the local NI to send from.
2280                  * Otherwise, let's pick any route we can find and then find
2281                  * a local NI we can reach the route's gateway on. Any route we
2282                  * select will be reachable by virtue of the restriction we have
2283                  * when adding a route.
2284                  */
2285                 best_route = lnet_find_route_locked(best_rnet,
2286                                                     LNET_NIDNET(src_nid),
2287                                                     sd->sd_best_lpni,
2288                                                     &last_route, &gwni);
2289
2290                 if (!best_route) {
2291                         CERROR("no route to %s from %s\n",
2292                                libcfs_nid2str(dst_nid),
2293                                libcfs_nid2str(src_nid));
2294                         return -EHOSTUNREACH;
2295                 }
2296
2297                 if (!gwni) {
2298                         CERROR("Internal Error. Route expected to %s from %s\n",
2299                                libcfs_nid2str(dst_nid),
2300                                libcfs_nid2str(src_nid));
2301                         return -EFAULT;
2302                 }
2303
2304                 gw = best_route->lr_gateway;
2305                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2306         }
2307
2308         /*
2309          * If the router checker is not active then discover the gateway here.
2310          * This ensures we are able to take advantage of multi-rail routing, but
2311          * if the router checker is active then we do not unecessarily delay
2312          * messages while the gateway is being checked by the dedicated monitor
2313          * thread.
2314          *
2315          * NB: We're only checking the alive_router_check_interval here, rather
2316          * than calling lnet_router_checker_active(), because the other
2317          * conditions that are checked by that function are either
2318          * irrelevant (the_lnet.ln_routing) or must be true (list of routers
2319          * is not empty)
2320          */
2321         if (alive_router_check_interval <= 0) {
2322                 rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2323                 if (rc)
2324                         return rc;
2325         }
2326
2327         if (!sd->sd_best_ni) {
2328                 lpn = gwni->lpni_peer_net;
2329                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2330                                                                sd->sd_msg,
2331                                                                sd->sd_md_cpt);
2332                 if (!sd->sd_best_ni) {
2333                         CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2334                                libcfs_net2str(lpn->lpn_net_id),
2335                                libcfs_nid2str(sd->sd_src_nid));
2336                         return -EFAULT;
2337                 }
2338         }
2339
2340         *gw_lpni = gwni;
2341         *gw_peer = gw;
2342
2343         /*
2344          * increment the sequence numbers since now we're sure we're
2345          * going to use this path
2346          */
2347         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2348                 LASSERT(best_route && last_route);
2349                 best_route->lr_seq = last_route->lr_seq + 1;
2350                 if (best_lpn)
2351                         best_lpn->lpn_seq++;
2352         }
2353
2354         return 0;
2355 }
2356
2357 /*
2358  * Handle two cases:
2359  *
2360  * Case 1:
2361  *  Source specified
2362  *  Remote destination
2363  *  Non-MR destination
2364  *
2365  * Case 2:
2366  *  Source specified
2367  *  Remote destination
2368  *  MR destination
2369  *
2370  * The handling of these two cases is similar. Even though the destination
2371  * can be MR or non-MR, we'll deal directly with the router.
2372  */
2373 static int
2374 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2375 {
2376         int rc;
2377         struct lnet_peer_ni *gw_lpni = NULL;
2378         struct lnet_peer *gw_peer = NULL;
2379
2380         /* find local NI */
2381         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2382         if (!sd->sd_best_ni) {
2383                 CERROR("Can't send to %s: src %s is not a "
2384                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2385                                 libcfs_nid2str(sd->sd_src_nid));
2386                 return -EINVAL;
2387         }
2388
2389         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2390                                      &gw_peer);
2391         if (rc)
2392                 return rc;
2393
2394         if (sd->sd_send_case & NMR_DST)
2395                 /*
2396                  * since the final destination is non-MR let's set its preferred
2397                  * NID before we send
2398                  */
2399                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2400                                          sd->sd_msg);
2401
2402         /*
2403          * We're going to send to the gw found so let's set its
2404          * info
2405          */
2406         sd->sd_peer = gw_peer;
2407         sd->sd_best_lpni = gw_lpni;
2408
2409         return lnet_handle_send(sd);
2410 }
2411
2412 struct lnet_ni *
2413 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2414                                struct lnet_msg *msg, bool discovery)
2415 {
2416         struct lnet_peer_net *lpn = NULL;
2417         struct lnet_peer_net *best_lpn = NULL;
2418         struct lnet_net *net = NULL;
2419         struct lnet_net *best_net = NULL;
2420         struct lnet_ni *best_ni = NULL;
2421         int best_lpn_healthv = 0;
2422         int best_net_healthv = 0;
2423         int net_healthv;
2424         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2425         __u32 lpn_sel_prio;
2426         __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2427         __u32 net_sel_prio;
2428         bool exit = false;
2429
2430         /*
2431          * The peer can have multiple interfaces, some of them can be on
2432          * the local network and others on a routed network. We should
2433          * prefer the local network. However if the local network is not
2434          * available then we need to try the routed network
2435          */
2436
2437         /* go through all the peer nets and find the best_ni */
2438         list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2439                 /*
2440                  * The peer's list of nets can contain non-local nets. We
2441                  * want to only examine the local ones.
2442                  */
2443                 net = lnet_get_net_locked(lpn->lpn_net_id);
2444                 if (!net)
2445                         continue;
2446
2447                 lpn_sel_prio = lpn->lpn_sel_priority;
2448                 net_healthv = lnet_get_net_healthv_locked(net);
2449                 net_sel_prio = net->net_sel_priority;
2450
2451                 /*
2452                  * if this is a discovery message and lp_disc_net_id is
2453                  * specified then use that net to send the discovery on.
2454                  */
2455                 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2456                     discovery) {
2457                         exit = true;
2458                         goto select_lpn;
2459                 }
2460
2461                 if (!best_lpn)
2462                         goto select_lpn;
2463
2464                 /* always select the lpn with the best health */
2465                 if (best_lpn_healthv > lpn->lpn_healthv)
2466                         continue;
2467                 else if (best_lpn_healthv < lpn->lpn_healthv)
2468                         goto select_lpn;
2469
2470                 /* select the preferred peer and local nets */
2471                 if (best_lpn_sel_prio < lpn_sel_prio)
2472                         continue;
2473                 else if (best_lpn_sel_prio > lpn_sel_prio)
2474                         goto select_lpn;
2475
2476                 if (best_net_healthv > net_healthv)
2477                         continue;
2478                 else if (best_net_healthv < net_healthv)
2479                         goto select_lpn;
2480
2481                 if (best_net_sel_prio < net_sel_prio)
2482                         continue;
2483                 else if (best_net_sel_prio > net_sel_prio)
2484                         goto select_lpn;
2485
2486                 if (best_lpn->lpn_seq < lpn->lpn_seq)
2487                         continue;
2488                 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2489                         goto select_lpn;
2490
2491                 /* round robin over the local networks */
2492                 if (best_net->net_seq <= net->net_seq)
2493                         continue;
2494
2495 select_lpn:
2496                 best_net_healthv = net_healthv;
2497                 best_net_sel_prio = net_sel_prio;
2498                 best_lpn_healthv = lpn->lpn_healthv;
2499                 best_lpn_sel_prio = lpn_sel_prio;
2500                 best_lpn = lpn;
2501                 best_net = net;
2502
2503                 if (exit)
2504                         break;
2505         }
2506
2507         if (best_lpn) {
2508                 /* Select the best NI on the same net as best_lpn chosen
2509                  * above
2510                  */
2511                 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2512                                                         msg, md_cpt);
2513         }
2514
2515         return best_ni;
2516 }
2517
2518 static struct lnet_ni *
2519 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2520 {
2521         struct lnet_ni *best_ni = NULL;
2522         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2523         struct lnet_peer_ni *lpni_entry;
2524
2525         /*
2526          * We must use a consistent source address when sending to a
2527          * non-MR peer. However, a non-MR peer can have multiple NIDs
2528          * on multiple networks, and we may even need to talk to this
2529          * peer on multiple networks -- certain types of
2530          * load-balancing configuration do this.
2531          *
2532          * So we need to pick the NI the peer prefers for this
2533          * particular network.
2534          */
2535         LASSERT(peer_net);
2536         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2537                             lpni_peer_nis) {
2538                 if (lpni_entry->lpni_pref_nnids == 0)
2539                         continue;
2540                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2541                 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2542                                                 cpt);
2543                 break;
2544         }
2545
2546         return best_ni;
2547 }
2548
2549 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2550 static int
2551 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2552 {
2553         struct lnet_ni *best_ni = NULL;
2554         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2555
2556         /*
2557          * We must use a consistent source address when sending to a
2558          * non-MR peer. However, a non-MR peer can have multiple NIDs
2559          * on multiple networks, and we may even need to talk to this
2560          * peer on multiple networks -- certain types of
2561          * load-balancing configuration do this.
2562          *
2563          * So we need to pick the NI the peer prefers for this
2564          * particular network.
2565          */
2566
2567         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2568                                                        sd->sd_cpt);
2569
2570         /* if best_ni is still not set just pick one */
2571         if (!best_ni) {
2572                 best_ni =
2573                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2574                                                 sd->sd_best_lpni->lpni_peer_net,
2575                                                 sd->sd_msg,
2576                                                 sd->sd_md_cpt);
2577                 /* If there is no best_ni we don't have a route */
2578                 if (!best_ni) {
2579                         CERROR("no path to %s from net %s\n",
2580                                 libcfs_nidstr(&best_lpni->lpni_nid),
2581                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2582                         return -EHOSTUNREACH;
2583                 }
2584         }
2585
2586         sd->sd_best_ni = best_ni;
2587
2588         /* Set preferred NI if necessary. */
2589         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2590
2591         return 0;
2592 }
2593
2594
2595 /*
2596  * Source not specified
2597  * Local destination
2598  * Non-MR Peer
2599  *
2600  * always use the same source NID for NMR peers
2601  * If we've talked to that peer before then we already have a preferred
2602  * source NI associated with it. Otherwise, we select a preferred local NI
2603  * and store it in the peer
2604  */
2605 static int
2606 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2607 {
2608         int rc = 0;
2609
2610         /* sd->sd_best_lpni is already set to the final destination */
2611
2612         /*
2613          * At this point we should've created the peer ni and peer. If we
2614          * can't find it, then something went wrong. Instead of assert
2615          * output a relevant message and fail the send
2616          */
2617         if (!sd->sd_best_lpni) {
2618                 CERROR("Internal fault. Unable to send msg %s to %s. "
2619                        "NID not known\n",
2620                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2621                        libcfs_nid2str(sd->sd_dst_nid));
2622                 return -EFAULT;
2623         }
2624
2625         if (sd->sd_msg->msg_routing) {
2626                 /* If I'm forwarding this message then I can choose any NI
2627                  * on the destination peer net
2628                  */
2629                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2630                                                                sd->sd_peer,
2631                                                                sd->sd_best_lpni->lpni_peer_net,
2632                                                                sd->sd_msg,
2633                                                                sd->sd_md_cpt);
2634                 if (!sd->sd_best_ni) {
2635                         CERROR("Unable to forward message to %s. No local NI available\n",
2636                                libcfs_nid2str(sd->sd_dst_nid));
2637                         rc = -EHOSTUNREACH;
2638                 }
2639         } else
2640                 rc = lnet_select_preferred_best_ni(sd);
2641
2642         if (!rc)
2643                 rc = lnet_handle_send(sd);
2644
2645         return rc;
2646 }
2647
2648 static int
2649 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2650 {
2651         /*
2652          * NOTE we've already handled the remote peer case. So we only
2653          * need to worry about the local case here.
2654          *
2655          * if we're sending a response, ACK or reply, we need to send it
2656          * to the destination NID given to us. At this point we already
2657          * have the peer_ni we're suppose to send to, so just find the
2658          * best_ni on the peer net and use that. Since we're sending to an
2659          * MR peer then we can just run the selection algorithm on our
2660          * local NIs and pick the best one.
2661          */
2662         if (sd->sd_send_case & SND_RESP) {
2663                 sd->sd_best_ni =
2664                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2665                                                 sd->sd_best_lpni->lpni_peer_net,
2666                                                 sd->sd_msg,
2667                                                 sd->sd_md_cpt);
2668
2669                 if (!sd->sd_best_ni) {
2670                         /*
2671                          * We're not going to deal with not able to send
2672                          * a response to the provided final destination
2673                          */
2674                         CERROR("Can't send response to %s. "
2675                                "No local NI available\n",
2676                                 libcfs_nid2str(sd->sd_dst_nid));
2677                         return -EHOSTUNREACH;
2678                 }
2679
2680                 return lnet_handle_send(sd);
2681         }
2682
2683         /*
2684          * If we get here that means we're sending a fresh request, PUT or
2685          * GET, so we need to run our standard selection algorithm.
2686          * First find the best local interface that's on any of the peer's
2687          * networks.
2688          */
2689         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2690                                         sd->sd_md_cpt,
2691                                         sd->sd_msg,
2692                                         lnet_msg_discovery(sd->sd_msg));
2693         if (sd->sd_best_ni) {
2694                 sd->sd_best_lpni =
2695                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2696                                       sd->sd_peer,
2697                                       sd->sd_best_ni->ni_net->net_id);
2698
2699                 /*
2700                  * if we're successful in selecting a peer_ni on the local
2701                  * network, then send to it. Otherwise fall through and
2702                  * try and see if we can reach it over another routed
2703                  * network
2704                  */
2705                 if (sd->sd_best_lpni &&
2706                     nid_same(&sd->sd_best_lpni->lpni_nid,
2707                              &the_lnet.ln_loni->ni_nid)) {
2708                         /*
2709                          * in case we initially started with a routed
2710                          * destination, let's reset to local
2711                          */
2712                         sd->sd_send_case &= ~REMOTE_DST;
2713                         sd->sd_send_case |= LOCAL_DST;
2714                         return lnet_handle_lo_send(sd);
2715                 } else if (sd->sd_best_lpni) {
2716                         /*
2717                          * in case we initially started with a routed
2718                          * destination, let's reset to local
2719                          */
2720                         sd->sd_send_case &= ~REMOTE_DST;
2721                         sd->sd_send_case |= LOCAL_DST;
2722                         return lnet_handle_send(sd);
2723                 }
2724
2725                 CERROR("Internal Error. Expected to have a best_lpni: "
2726                        "%s -> %s\n",
2727                        libcfs_nid2str(sd->sd_src_nid),
2728                        libcfs_nid2str(sd->sd_dst_nid));
2729
2730                 return -EFAULT;
2731         }
2732
2733         /*
2734          * Peer doesn't have a local network. Let's see if there is
2735          * a remote network we can reach it on.
2736          */
2737         return PASS_THROUGH;
2738 }
2739
2740 /*
2741  * Case 1:
2742  *      Source NID not specified
2743  *      Local destination
2744  *      MR peer
2745  *
2746  * Case 2:
2747  *      Source NID not speified
2748  *      Remote destination
2749  *      MR peer
2750  *
2751  * In both of these cases if we're sending a response, ACK or REPLY, then
2752  * we need to send to the destination NID provided.
2753  *
2754  * In the remote case let's deal with MR routers.
2755  *
2756  */
2757
2758 static int
2759 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2760 {
2761         int rc = 0;
2762         struct lnet_peer *gw_peer = NULL;
2763         struct lnet_peer_ni *gw_lpni = NULL;
2764
2765         /*
2766          * handle sending a response to a remote peer here so we don't
2767          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2768          */
2769         if (sd->sd_send_case & REMOTE_DST &&
2770             sd->sd_send_case & SND_RESP) {
2771                 struct lnet_peer_ni *gw;
2772                 struct lnet_peer *gw_peer;
2773
2774                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2775                                                   &gw_peer);
2776                 if (rc < 0) {
2777                         CERROR("Can't send response to %s. "
2778                                "No route available\n",
2779                                 libcfs_nid2str(sd->sd_dst_nid));
2780                         return -EHOSTUNREACH;
2781                 } else if (rc > 0) {
2782                         return rc;
2783                 }
2784
2785                 sd->sd_best_lpni = gw;
2786                 sd->sd_peer = gw_peer;
2787
2788                 return lnet_handle_send(sd);
2789         }
2790
2791         /*
2792          * Even though the NID for the peer might not be on a local network,
2793          * since the peer is MR there could be other interfaces on the
2794          * local network. In that case we'd still like to prefer the local
2795          * network over the routed network. If we're unable to do that
2796          * then we select the best router among the different routed networks,
2797          * and if the router is MR then we can deal with it as such.
2798          */
2799         rc = lnet_handle_any_mr_dsta(sd);
2800         if (rc != PASS_THROUGH)
2801                 return rc;
2802
2803         /*
2804          * Now that we must route to the destination, we must consider the
2805          * MR case, where the destination has multiple interfaces, some of
2806          * which we can route to and others we do not. For this reason we
2807          * need to select the destination which we can route to and if
2808          * there are multiple, we need to round robin.
2809          */
2810         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2811                                           &gw_peer);
2812         if (rc)
2813                 return rc;
2814
2815         sd->sd_send_case &= ~LOCAL_DST;
2816         sd->sd_send_case |= REMOTE_DST;
2817
2818         sd->sd_peer = gw_peer;
2819         sd->sd_best_lpni = gw_lpni;
2820
2821         return lnet_handle_send(sd);
2822 }
2823
2824 /*
2825  * Source not specified
2826  * Remote destination
2827  * Non-MR peer
2828  *
2829  * Must send to the specified peer NID using the same source NID that
2830  * we've used before. If it's the first time to talk to that peer then
2831  * find the source NI and assign it as preferred to that peer
2832  */
2833 static int
2834 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2835 {
2836         int rc;
2837         struct lnet_peer_ni *gw_lpni = NULL;
2838         struct lnet_peer *gw_peer = NULL;
2839
2840         /*
2841          * Let's see if we have a preferred NI to talk to this NMR peer
2842          */
2843         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2844                                                               sd->sd_cpt);
2845
2846         /*
2847          * find the router and that'll find the best NI if we didn't find
2848          * it already.
2849          */
2850         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2851                                           &gw_peer);
2852         if (rc)
2853                 return rc;
2854
2855         /*
2856          * set the best_ni we've chosen as the preferred one for
2857          * this peer
2858          */
2859         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2860
2861         /* we'll be sending to the gw */
2862         sd->sd_best_lpni = gw_lpni;
2863         sd->sd_peer = gw_peer;
2864
2865         return lnet_handle_send(sd);
2866 }
2867
2868 static int
2869 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2870 {
2871         /*
2872          * turn off the SND_RESP bit.
2873          * It will be checked in the case handling
2874          */
2875         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2876
2877         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2878                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2879                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2880                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2881                 libcfs_nid2str(sd->sd_dst_nid),
2882                 (send_case & LOCAL_DST) ? "local" : "routed");
2883
2884         switch (send_case) {
2885         /*
2886          * For all cases where the source is specified, we should always
2887          * use the destination NID, whether it's an MR destination or not,
2888          * since we're continuing a series of related messages for the
2889          * same RPC
2890          */
2891         case SRC_SPEC_LOCAL_NMR_DST:
2892                 return lnet_handle_spec_local_nmr_dst(sd);
2893         case SRC_SPEC_LOCAL_MR_DST:
2894                 return lnet_handle_spec_local_mr_dst(sd);
2895         case SRC_SPEC_ROUTER_NMR_DST:
2896         case SRC_SPEC_ROUTER_MR_DST:
2897                 return lnet_handle_spec_router_dst(sd);
2898         case SRC_ANY_LOCAL_NMR_DST:
2899                 return lnet_handle_any_local_nmr_dst(sd);
2900         case SRC_ANY_LOCAL_MR_DST:
2901         case SRC_ANY_ROUTER_MR_DST:
2902                 return lnet_handle_any_mr_dst(sd);
2903         case SRC_ANY_ROUTER_NMR_DST:
2904                 return lnet_handle_any_router_nmr_dst(sd);
2905         default:
2906                 CERROR("Unknown send case\n");
2907                 return -1;
2908         }
2909 }
2910
2911 static int
2912 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2913                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2914 {
2915         struct lnet_peer_ni *lpni;
2916         struct lnet_peer *peer;
2917         struct lnet_send_data send_data;
2918         int cpt, rc;
2919         int md_cpt;
2920         __u32 send_case = 0;
2921         bool final_hop;
2922         bool mr_forwarding_allowed;
2923
2924         memset(&send_data, 0, sizeof(send_data));
2925
2926         /*
2927          * get an initial CPT to use for locking. The idea here is not to
2928          * serialize the calls to select_pathway, so that as many
2929          * operations can run concurrently as possible. To do that we use
2930          * the CPT where this call is being executed. Later on when we
2931          * determine the CPT to use in lnet_message_commit, we switch the
2932          * lock and check if there was any configuration change.  If none,
2933          * then we proceed, if there is, then we restart the operation.
2934          */
2935         cpt = lnet_net_lock_current();
2936
2937         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2938         if (md_cpt == CFS_CPT_ANY)
2939                 md_cpt = cpt;
2940
2941 again:
2942
2943         /*
2944          * If we're being asked to send to the loopback interface, there
2945          * is no need to go through any selection. We can just shortcut
2946          * the entire process and send over lolnd
2947          */
2948         send_data.sd_msg = msg;
2949         send_data.sd_cpt = cpt;
2950         if (dst_nid == LNET_NID_LO_0) {
2951                 rc = lnet_handle_lo_send(&send_data);
2952                 lnet_net_unlock(cpt);
2953                 return rc;
2954         }
2955
2956         /*
2957          * find an existing peer_ni, or create one and mark it as having been
2958          * created due to network traffic. This call will create the
2959          * peer->peer_net->peer_ni tree.
2960          */
2961         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2962         if (IS_ERR(lpni)) {
2963                 lnet_net_unlock(cpt);
2964                 return PTR_ERR(lpni);
2965         }
2966
2967         /*
2968          * Cache the original src_nid and rtr_nid. If we need to resend the
2969          * message then we'll need to know whether the src_nid was originally
2970          * specified for this message. If it was originally specified,
2971          * then we need to keep using the same src_nid since it's
2972          * continuing the same sequence of messages. Similarly, rtr_nid will
2973          * affect our choice of next hop.
2974          */
2975         lnet_nid4_to_nid(src_nid, &msg->msg_src_nid_param);
2976         lnet_nid4_to_nid(rtr_nid, &msg->msg_rtr_nid_param);
2977
2978         /*
2979          * If necessary, perform discovery on the peer that owns this peer_ni.
2980          * Note, this can result in the ownership of this peer_ni changing
2981          * to another peer object.
2982          */
2983         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2984         if (rc) {
2985                 lnet_peer_ni_decref_locked(lpni);
2986                 lnet_net_unlock(cpt);
2987                 return rc;
2988         }
2989         lnet_peer_ni_decref_locked(lpni);
2990
2991         peer = lpni->lpni_peer_net->lpn_peer;
2992
2993         /*
2994          * Identify the different send cases
2995          */
2996         if (src_nid == LNET_NID_ANY) {
2997                 send_case |= SRC_ANY;
2998                 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2999                         send_case |= LOCAL_DST;
3000                 else
3001                         send_case |= REMOTE_DST;
3002         } else {
3003                 send_case |= SRC_SPEC;
3004                 if (LNET_NIDNET(src_nid) == LNET_NIDNET(dst_nid))
3005                         send_case |= LOCAL_DST;
3006                 else
3007                         send_case |= REMOTE_DST;
3008         }
3009
3010         final_hop = false;
3011         if (msg->msg_routing && (send_case & LOCAL_DST))
3012                 final_hop = true;
3013
3014         /* Determine whether to allow MR forwarding for this message.
3015          * NB: MR forwarding is allowed if the message originator and the
3016          * destination are both MR capable, and the destination lpni that was
3017          * originally chosen by the originator is unhealthy or down.
3018          * We check the MR capability of the destination further below
3019          */
3020         mr_forwarding_allowed = false;
3021         if (final_hop) {
3022                 struct lnet_peer *src_lp;
3023                 struct lnet_peer_ni *src_lpni;
3024
3025                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
3026                                                   LNET_NID_ANY, cpt);
3027                 /* We don't fail the send if we hit any errors here. We'll just
3028                  * try to send it via non-multi-rail criteria
3029                  */
3030                 if (!IS_ERR(src_lpni)) {
3031                         /* Drop ref taken by lnet_nid2peerni_locked() */
3032                         lnet_peer_ni_decref_locked(src_lpni);
3033                         src_lp = lpni->lpni_peer_net->lpn_peer;
3034                         if (lnet_peer_is_multi_rail(src_lp) &&
3035                             !lnet_is_peer_ni_alive(lpni))
3036                                 mr_forwarding_allowed = true;
3037
3038                 }
3039                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3040                        mr_forwarding_allowed ? "allowed" : "not allowed");
3041         }
3042
3043         /*
3044          * Deal with the peer as NMR in the following cases:
3045          * 1. the peer is NMR
3046          * 2. We're trying to recover a specific peer NI
3047          * 3. I'm a router sending to the final destination and MR forwarding is
3048          *    not allowed for this message (as determined above).
3049          *    In this case the source of the message would've
3050          *    already selected the final destination so my job
3051          *    is to honor the selection.
3052          */
3053         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3054             (final_hop && !mr_forwarding_allowed))
3055                 send_case |= NMR_DST;
3056         else
3057                 send_case |= MR_DST;
3058
3059         if (lnet_msg_is_response(msg))
3060                 send_case |= SND_RESP;
3061
3062         /* assign parameters to the send_data */
3063         send_data.sd_rtr_nid = rtr_nid;
3064         send_data.sd_src_nid = src_nid;
3065         send_data.sd_dst_nid = dst_nid;
3066         send_data.sd_best_lpni = lpni;
3067         /*
3068          * keep a pointer to the final destination in case we're going to
3069          * route, so we'll need to access it later
3070          */
3071         send_data.sd_final_dst_lpni = lpni;
3072         send_data.sd_peer = peer;
3073         send_data.sd_md_cpt = md_cpt;
3074         send_data.sd_send_case = send_case;
3075
3076         rc = lnet_handle_send_case_locked(&send_data);
3077
3078         /*
3079          * Update the local cpt since send_data.sd_cpt might've been
3080          * updated as a result of calling lnet_handle_send_case_locked().
3081          */
3082         cpt = send_data.sd_cpt;
3083
3084         if (rc == REPEAT_SEND)
3085                 goto again;
3086
3087         lnet_net_unlock(cpt);
3088
3089         return rc;
3090 }
3091
3092 int
3093 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3094 {
3095         lnet_nid_t dst_nid = lnet_nid_to_nid4(&msg->msg_target.nid);
3096         int rc;
3097
3098         /*
3099          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3100          * but we might want to use pre-determined router for ACK/REPLY
3101          * in the future
3102          */
3103         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3104         LASSERT(msg->msg_txpeer == NULL);
3105         LASSERT(msg->msg_txni == NULL);
3106         LASSERT(!msg->msg_sending);
3107         LASSERT(!msg->msg_target_is_router);
3108         LASSERT(!msg->msg_receiving);
3109
3110         msg->msg_sending = 1;
3111
3112         LASSERT(!msg->msg_tx_committed);
3113
3114         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3115         if (rc < 0) {
3116                 if (rc == -EHOSTUNREACH)
3117                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3118                 else
3119                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3120                 return rc;
3121         }
3122
3123         if (rc == LNET_CREDIT_OK)
3124                 lnet_ni_send(msg->msg_txni, msg);
3125
3126         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3127         return 0;
3128 }
3129
3130 enum lnet_mt_event_type {
3131         MT_TYPE_LOCAL_NI = 0,
3132         MT_TYPE_PEER_NI
3133 };
3134
3135 struct lnet_mt_event_info {
3136         enum lnet_mt_event_type mt_type;
3137         lnet_nid_t mt_nid;
3138 };
3139
3140 /* called with res_lock held */
3141 void
3142 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3143 {
3144         struct lnet_rsp_tracker *rspt;
3145
3146         /*
3147          * msg has a refcount on the MD so the MD is not going away.
3148          * The rspt queue for the cpt is protected by
3149          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3150          */
3151         if (!md->md_rspt_ptr)
3152                 return;
3153
3154         rspt = md->md_rspt_ptr;
3155
3156         /* debug code */
3157         LASSERT(rspt->rspt_cpt == cpt);
3158
3159         md->md_rspt_ptr = NULL;
3160
3161         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3162                 /*
3163                  * The monitor thread has invalidated this handle because the
3164                  * response timed out, but it failed to lookup the MD. That
3165                  * means this response tracker is on the zombie list. We can
3166                  * safely remove it under the resource lock (held by caller) and
3167                  * free the response tracker block.
3168                  */
3169                 list_del(&rspt->rspt_on_list);
3170                 lnet_rspt_free(rspt, cpt);
3171         } else {
3172                 /*
3173                  * invalidate the handle to indicate that a response has been
3174                  * received, which will then lead the monitor thread to clean up
3175                  * the rspt block.
3176                  */
3177                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3178         }
3179 }
3180
3181 void
3182 lnet_clean_zombie_rstqs(void)
3183 {
3184         struct lnet_rsp_tracker *rspt, *tmp;
3185         int i;
3186
3187         cfs_cpt_for_each(i, lnet_cpt_table()) {
3188                 list_for_each_entry_safe(rspt, tmp,
3189                                          the_lnet.ln_mt_zombie_rstqs[i],
3190                                          rspt_on_list) {
3191                         list_del(&rspt->rspt_on_list);
3192                         lnet_rspt_free(rspt, i);
3193                 }
3194         }
3195
3196         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3197 }
3198
3199 static void
3200 lnet_finalize_expired_responses(void)
3201 {
3202         struct lnet_libmd *md;
3203         struct lnet_rsp_tracker *rspt, *tmp;
3204         ktime_t now;
3205         int i;
3206
3207         if (the_lnet.ln_mt_rstq == NULL)
3208                 return;
3209
3210         cfs_cpt_for_each(i, lnet_cpt_table()) {
3211                 LIST_HEAD(local_queue);
3212
3213                 lnet_net_lock(i);
3214                 if (!the_lnet.ln_mt_rstq[i]) {
3215                         lnet_net_unlock(i);
3216                         continue;
3217                 }
3218                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3219                 lnet_net_unlock(i);
3220
3221                 now = ktime_get();
3222
3223                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3224                         /*
3225                          * The rspt mdh will be invalidated when a response
3226                          * is received or whenever we want to discard the
3227                          * block the monitor thread will walk the queue
3228                          * and clean up any rsts with an invalid mdh.
3229                          * The monitor thread will walk the queue until
3230                          * the first unexpired rspt block. This means that
3231                          * some rspt blocks which received their
3232                          * corresponding responses will linger in the
3233                          * queue until they are cleaned up eventually.
3234                          */
3235                         lnet_res_lock(i);
3236                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3237                                 lnet_res_unlock(i);
3238                                 list_del(&rspt->rspt_on_list);
3239                                 lnet_rspt_free(rspt, i);
3240                                 continue;
3241                         }
3242
3243                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3244                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3245                                 struct lnet_peer_ni *lpni;
3246                                 struct lnet_nid nid;
3247
3248                                 md = lnet_handle2md(&rspt->rspt_mdh);
3249                                 if (!md) {
3250                                         /* MD has been queued for unlink, but
3251                                          * rspt hasn't been detached (Note we've
3252                                          * checked above that the rspt_mdh is
3253                                          * valid). Since we cannot lookup the MD
3254                                          * we're unable to detach the rspt
3255                                          * ourselves. Thus, move the rspt to the
3256                                          * zombie list where we'll wait for
3257                                          * either:
3258                                          *   1. The remaining operations on the
3259                                          *   MD to complete. In this case the
3260                                          *   final operation will result in
3261                                          *   lnet_msg_detach_md()->
3262                                          *   lnet_detach_rsp_tracker() where
3263                                          *   we will clean up this response
3264                                          *   tracker.
3265                                          *   2. LNet to shutdown. In this case
3266                                          *   we'll wait until after all LND Nets
3267                                          *   have shutdown and then we can
3268                                          *   safely free any remaining response
3269                                          *   tracker blocks on the zombie list.
3270                                          * Note: We need to hold the resource
3271                                          * lock when adding to the zombie list
3272                                          * because we may have concurrent access
3273                                          * with lnet_detach_rsp_tracker().
3274                                          */
3275                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
3276                                         list_move(&rspt->rspt_on_list,
3277                                                   the_lnet.ln_mt_zombie_rstqs[i]);
3278                                         lnet_res_unlock(i);
3279                                         continue;
3280                                 }
3281                                 LASSERT(md->md_rspt_ptr == rspt);
3282                                 md->md_rspt_ptr = NULL;
3283                                 lnet_res_unlock(i);
3284
3285                                 LNetMDUnlink(rspt->rspt_mdh);
3286
3287                                 nid = rspt->rspt_next_hop_nid;
3288
3289                                 list_del(&rspt->rspt_on_list);
3290                                 lnet_rspt_free(rspt, i);
3291
3292                                 /* If we're shutting down we just want to clean
3293                                  * up the rspt blocks
3294                                  */
3295                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3296                                         continue;
3297
3298                                 lnet_net_lock(i);
3299                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3300                                 lnet_net_unlock(i);
3301
3302                                 CDEBUG(D_NET,
3303                                        "Response timeout: md = %p: nid = %s\n",
3304                                        md, libcfs_nidstr(&nid));
3305
3306                                 /*
3307                                  * If there is a timeout on the response
3308                                  * from the next hop decrement its health
3309                                  * value so that we don't use it
3310                                  */
3311                                 lnet_net_lock(0);
3312                                 lpni = lnet_peer_ni_find_locked(&nid);
3313                                 if (lpni) {
3314                                         lnet_handle_remote_failure_locked(lpni);
3315                                         lnet_peer_ni_decref_locked(lpni);
3316                                 }
3317                                 lnet_net_unlock(0);
3318                         } else {
3319                                 lnet_res_unlock(i);
3320                                 break;
3321                         }
3322                 }
3323
3324                 if (!list_empty(&local_queue)) {
3325                         lnet_net_lock(i);
3326                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3327                         lnet_net_unlock(i);
3328                 }
3329         }
3330 }
3331
3332 static void
3333 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3334 {
3335         struct lnet_msg *msg;
3336
3337         while (!list_empty(resendq)) {
3338                 struct lnet_peer_ni *lpni;
3339
3340                 msg = list_entry(resendq->next, struct lnet_msg,
3341                                  msg_list);
3342
3343                 list_del_init(&msg->msg_list);
3344
3345                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3346                 if (!lpni) {
3347                         lnet_net_unlock(cpt);
3348                         CERROR("Expected that a peer is already created for %s\n",
3349                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3350                         msg->msg_no_resend = true;
3351                         lnet_finalize(msg, -EFAULT);
3352                         lnet_net_lock(cpt);
3353                 } else {
3354                         int rc;
3355
3356                         lnet_peer_ni_decref_locked(lpni);
3357
3358                         lnet_net_unlock(cpt);
3359                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3360                                libcfs_nidstr(&msg->msg_src_nid_param),
3361                                libcfs_idstr(&msg->msg_target),
3362                                lnet_msgtyp2str(msg->msg_type),
3363                                msg->msg_recovery,
3364                                msg->msg_retry_count);
3365                         rc = lnet_send(lnet_nid_to_nid4(&msg->msg_src_nid_param),
3366                                        msg,
3367                                        lnet_nid_to_nid4(&msg->msg_rtr_nid_param));
3368                         if (rc) {
3369                                 CERROR("Error sending %s to %s: %d\n",
3370                                        lnet_msgtyp2str(msg->msg_type),
3371                                        libcfs_idstr(&msg->msg_target), rc);
3372                                 msg->msg_no_resend = true;
3373                                 lnet_finalize(msg, rc);
3374                         }
3375                         lnet_net_lock(cpt);
3376                         if (!rc)
3377                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3378                 }
3379         }
3380 }
3381
3382 static void
3383 lnet_resend_pending_msgs(void)
3384 {
3385         int i;
3386
3387         cfs_cpt_for_each(i, lnet_cpt_table()) {
3388                 lnet_net_lock(i);
3389                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3390                 lnet_net_unlock(i);
3391         }
3392 }
3393
3394 /* called with cpt and ni_lock held */
3395 static void
3396 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3397 {
3398         struct lnet_handle_md recovery_mdh;
3399
3400         LNetInvalidateMDHandle(&recovery_mdh);
3401
3402         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3403             force) {
3404                 recovery_mdh = ni->ni_ping_mdh;
3405                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3406         }
3407         lnet_ni_unlock(ni);
3408         lnet_net_unlock(cpt);
3409         if (!LNetMDHandleIsInvalid(recovery_mdh))
3410                 LNetMDUnlink(recovery_mdh);
3411         lnet_net_lock(cpt);
3412         lnet_ni_lock(ni);
3413 }
3414
3415 static void
3416 lnet_recover_local_nis(void)
3417 {
3418         struct lnet_mt_event_info *ev_info;
3419         LIST_HEAD(processed_list);
3420         LIST_HEAD(local_queue);
3421         struct lnet_handle_md mdh;
3422         struct lnet_ni *tmp;
3423         struct lnet_ni *ni;
3424         lnet_nid_t nid;
3425         int healthv;
3426         int rc;
3427         time64_t now;
3428
3429         /*
3430          * splice the recovery queue on a local queue. We will iterate
3431          * through the local queue and update it as needed. Once we're
3432          * done with the traversal, we'll splice the local queue back on
3433          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3434          * will be traversed in the next iteration.
3435          */
3436         lnet_net_lock(0);
3437         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3438                          &local_queue);
3439         lnet_net_unlock(0);
3440
3441         now = ktime_get_seconds();
3442
3443         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3444                 /*
3445                  * if an NI is being deleted or it is now healthy, there
3446                  * is no need to keep it around in the recovery queue.
3447                  * The monitor thread is the only thread responsible for
3448                  * removing the NI from the recovery queue.
3449                  * Multiple threads can be adding NIs to the recovery
3450                  * queue.
3451                  */
3452                 healthv = atomic_read(&ni->ni_healthv);
3453
3454                 lnet_net_lock(0);
3455                 lnet_ni_lock(ni);
3456                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3457                     healthv == LNET_MAX_HEALTH_VALUE) {
3458                         list_del_init(&ni->ni_recovery);
3459                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3460                         lnet_ni_unlock(ni);
3461                         lnet_ni_decref_locked(ni, 0);
3462                         lnet_net_unlock(0);
3463                         continue;
3464                 }
3465
3466                 /*
3467                  * if the local NI failed recovery we must unlink the md.
3468                  * But we want to keep the local_ni on the recovery queue
3469                  * so we can continue the attempts to recover it.
3470                  */
3471                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3472                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3473                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3474                 }
3475
3476
3477                 lnet_ni_unlock(ni);
3478
3479                 if (now < ni->ni_next_ping) {
3480                         lnet_net_unlock(0);
3481                         continue;
3482                 }
3483
3484                 lnet_net_unlock(0);
3485
3486                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3487                        libcfs_nidstr(&ni->ni_nid));
3488
3489                 lnet_ni_lock(ni);
3490                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3491                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3492                         lnet_ni_unlock(ni);
3493
3494                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3495                         if (!ev_info) {
3496                                 CERROR("out of memory. Can't recover %s\n",
3497                                        libcfs_nidstr(&ni->ni_nid));
3498                                 lnet_ni_lock(ni);
3499                                 ni->ni_recovery_state &=
3500                                   ~LNET_NI_RECOVERY_PENDING;
3501                                 lnet_ni_unlock(ni);
3502                                 continue;
3503                         }
3504
3505                         mdh = ni->ni_ping_mdh;
3506                         /*
3507                          * Invalidate the ni mdh in case it's deleted.
3508                          * We'll unlink the mdh in this case below.
3509                          */
3510                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3511                         /* FIXME need to handle large-addr nid */
3512                         nid = lnet_nid_to_nid4(&ni->ni_nid);
3513
3514                         /*
3515                          * remove the NI from the local queue and drop the
3516                          * reference count to it while we're recovering
3517                          * it. The reason for that, is that the NI could
3518                          * be deleted, and the way the code is structured
3519                          * is if we don't drop the NI, then the deletion
3520                          * code will enter a loop waiting for the
3521                          * reference count to be removed while holding the
3522                          * ln_mutex_lock(). When we look up the peer to
3523                          * send to in lnet_select_pathway() we will try to
3524                          * lock the ln_mutex_lock() as well, leading to
3525                          * a deadlock. By dropping the refcount and
3526                          * removing it from the list, we allow for the NI
3527                          * to be removed, then we use the cached NID to
3528                          * look it up again. If it's gone, then we just
3529                          * continue examining the rest of the queue.
3530                          */
3531                         lnet_net_lock(0);
3532                         list_del_init(&ni->ni_recovery);
3533                         lnet_ni_decref_locked(ni, 0);
3534                         lnet_net_unlock(0);
3535
3536                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3537                         ev_info->mt_nid = nid;
3538                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3539                                             ev_info, the_lnet.ln_mt_handler,
3540                                             true);
3541                         /* lookup the nid again */
3542                         lnet_net_lock(0);
3543                         ni = lnet_nid2ni_locked(nid, 0);
3544                         if (!ni) {
3545                                 /*
3546                                  * the NI has been deleted when we dropped
3547                                  * the ref count
3548                                  */
3549                                 lnet_net_unlock(0);
3550                                 LNetMDUnlink(mdh);
3551                                 continue;
3552                         }
3553                         ni->ni_ping_count++;
3554
3555                         ni->ni_ping_mdh = mdh;
3556                         lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3557                                                         now);
3558
3559                         if (rc) {
3560                                 lnet_ni_lock(ni);
3561                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3562                                 lnet_ni_unlock(ni);
3563                         }
3564                         lnet_net_unlock(0);
3565                 } else
3566                         lnet_ni_unlock(ni);
3567         }
3568
3569         /*
3570          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3571          * reexamined in the next iteration.
3572          */
3573         list_splice_init(&processed_list, &local_queue);
3574         lnet_net_lock(0);
3575         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3576         lnet_net_unlock(0);
3577 }
3578
3579 static int
3580 lnet_resendqs_create(void)
3581 {
3582         struct list_head **resendqs;
3583         resendqs = lnet_create_array_of_queues();
3584
3585         if (!resendqs)
3586                 return -ENOMEM;
3587
3588         lnet_net_lock(LNET_LOCK_EX);
3589         the_lnet.ln_mt_resendqs = resendqs;
3590         lnet_net_unlock(LNET_LOCK_EX);
3591
3592         return 0;
3593 }
3594
3595 static void
3596 lnet_clean_local_ni_recoveryq(void)
3597 {
3598         struct lnet_ni *ni;
3599
3600         /* This is only called when the monitor thread has stopped */
3601         lnet_net_lock(0);
3602
3603         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3604                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3605                                 struct lnet_ni, ni_recovery);
3606                 list_del_init(&ni->ni_recovery);
3607                 lnet_ni_lock(ni);
3608                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3609                 lnet_ni_unlock(ni);
3610                 lnet_ni_decref_locked(ni, 0);
3611         }
3612
3613         lnet_net_unlock(0);
3614 }
3615
3616 static void
3617 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3618                                      bool force)
3619 {
3620         struct lnet_handle_md recovery_mdh;
3621
3622         LNetInvalidateMDHandle(&recovery_mdh);
3623
3624         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3625                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3626                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3627         }
3628         spin_unlock(&lpni->lpni_lock);
3629         lnet_net_unlock(cpt);
3630         if (!LNetMDHandleIsInvalid(recovery_mdh))
3631                 LNetMDUnlink(recovery_mdh);
3632         lnet_net_lock(cpt);
3633         spin_lock(&lpni->lpni_lock);
3634 }
3635
3636 static void
3637 lnet_clean_peer_ni_recoveryq(void)
3638 {
3639         struct lnet_peer_ni *lpni, *tmp;
3640
3641         lnet_net_lock(LNET_LOCK_EX);
3642
3643         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3644                                  lpni_recovery) {
3645                 list_del_init(&lpni->lpni_recovery);
3646                 spin_lock(&lpni->lpni_lock);
3647                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3648                 spin_unlock(&lpni->lpni_lock);
3649                 lnet_peer_ni_decref_locked(lpni);
3650         }
3651
3652         lnet_net_unlock(LNET_LOCK_EX);
3653 }
3654
3655 static void
3656 lnet_clean_resendqs(void)
3657 {
3658         struct lnet_msg *msg, *tmp;
3659         LIST_HEAD(msgs);
3660         int i;
3661
3662         cfs_cpt_for_each(i, lnet_cpt_table()) {
3663                 lnet_net_lock(i);
3664                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3665                 lnet_net_unlock(i);
3666                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3667                         list_del_init(&msg->msg_list);
3668                         msg->msg_no_resend = true;
3669                         lnet_finalize(msg, -ESHUTDOWN);
3670                 }
3671         }
3672
3673         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3674 }
3675
3676 static void
3677 lnet_recover_peer_nis(void)
3678 {
3679         struct lnet_mt_event_info *ev_info;
3680         LIST_HEAD(processed_list);
3681         LIST_HEAD(local_queue);
3682         struct lnet_handle_md mdh;
3683         struct lnet_peer_ni *lpni;
3684         struct lnet_peer_ni *tmp;
3685         lnet_nid_t nid;
3686         int healthv;
3687         int rc;
3688         time64_t now;
3689
3690         /*
3691          * Always use cpt 0 for locking across all interactions with
3692          * ln_mt_peerNIRecovq
3693          */
3694         lnet_net_lock(0);
3695         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3696                          &local_queue);
3697         lnet_net_unlock(0);
3698
3699         now = ktime_get_seconds();
3700
3701         list_for_each_entry_safe(lpni, tmp, &local_queue,
3702                                  lpni_recovery) {
3703                 /*
3704                  * The same protection strategy is used here as is in the
3705                  * local recovery case.
3706                  */
3707                 lnet_net_lock(0);
3708                 healthv = atomic_read(&lpni->lpni_healthv);
3709                 spin_lock(&lpni->lpni_lock);
3710                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3711                     healthv == LNET_MAX_HEALTH_VALUE) {
3712                         list_del_init(&lpni->lpni_recovery);
3713                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3714                         spin_unlock(&lpni->lpni_lock);
3715                         lnet_peer_ni_decref_locked(lpni);
3716                         lnet_net_unlock(0);
3717                         continue;
3718                 }
3719
3720                 /*
3721                  * If the peer NI has failed recovery we must unlink the
3722                  * md. But we want to keep the peer ni on the recovery
3723                  * queue so we can try to continue recovering it
3724                  */
3725                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3726                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3727                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3728                 }
3729
3730                 spin_unlock(&lpni->lpni_lock);
3731
3732                 if (now < lpni->lpni_next_ping) {
3733                         lnet_net_unlock(0);
3734                         continue;
3735                 }
3736
3737                 lnet_net_unlock(0);
3738
3739                 /*
3740                  * NOTE: we're racing with peer deletion from user space.
3741                  * It's possible that a peer is deleted after we check its
3742                  * state. In this case the recovery can create a new peer
3743                  */
3744                 spin_lock(&lpni->lpni_lock);
3745                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3746                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3747                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3748                         spin_unlock(&lpni->lpni_lock);
3749
3750                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3751                         if (!ev_info) {
3752                                 CERROR("out of memory. Can't recover %s\n",
3753                                        libcfs_nidstr(&lpni->lpni_nid));
3754                                 spin_lock(&lpni->lpni_lock);
3755                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3756                                 spin_unlock(&lpni->lpni_lock);
3757                                 continue;
3758                         }
3759
3760                         /* look at the comments in lnet_recover_local_nis() */
3761                         mdh = lpni->lpni_recovery_ping_mdh;
3762                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3763                         /* FIXME handle large-addr nid */
3764                         nid = lnet_nid_to_nid4(&lpni->lpni_nid);
3765                         lnet_net_lock(0);
3766                         list_del_init(&lpni->lpni_recovery);
3767                         lnet_peer_ni_decref_locked(lpni);
3768                         lnet_net_unlock(0);
3769
3770                         ev_info->mt_type = MT_TYPE_PEER_NI;
3771                         ev_info->mt_nid = nid;
3772                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3773                                             ev_info, the_lnet.ln_mt_handler,
3774                                             true);
3775                         lnet_net_lock(0);
3776                         /*
3777                          * lnet_find_peer_ni_locked() grabs a refcount for
3778                          * us. No need to take it explicitly.
3779                          */
3780                         lpni = lnet_find_peer_ni_locked(nid);
3781                         if (!lpni) {
3782                                 lnet_net_unlock(0);
3783                                 LNetMDUnlink(mdh);
3784                                 continue;
3785                         }
3786
3787                         lpni->lpni_ping_count++;
3788
3789                         lpni->lpni_recovery_ping_mdh = mdh;
3790
3791                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
3792                                                              &processed_list,
3793                                                              now);
3794                         if (rc) {
3795                                 spin_lock(&lpni->lpni_lock);
3796                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3797                                 spin_unlock(&lpni->lpni_lock);
3798                         }
3799
3800                         /* Drop the ref taken by lnet_find_peer_ni_locked() */
3801                         lnet_peer_ni_decref_locked(lpni);
3802                         lnet_net_unlock(0);
3803                 } else
3804                         spin_unlock(&lpni->lpni_lock);
3805         }
3806
3807         list_splice_init(&processed_list, &local_queue);
3808         lnet_net_lock(0);
3809         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3810         lnet_net_unlock(0);
3811 }
3812
3813 static int
3814 lnet_monitor_thread(void *arg)
3815 {
3816         time64_t rsp_timeout = 0;
3817         time64_t now;
3818
3819         wait_for_completion(&the_lnet.ln_started);
3820         /*
3821          * The monitor thread takes care of the following:
3822          *  1. Checks the aliveness of routers
3823          *  2. Checks if there are messages on the resend queue to resend
3824          *     them.
3825          *  3. Check if there are any NIs on the local recovery queue and
3826          *     pings them
3827          *  4. Checks if there are any NIs on the remote recovery queue
3828          *     and pings them.
3829          */
3830         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3831                 now = ktime_get_real_seconds();
3832
3833                 if (lnet_router_checker_active())
3834                         lnet_check_routers();
3835
3836                 lnet_resend_pending_msgs();
3837
3838                 if (now >= rsp_timeout) {
3839                         lnet_finalize_expired_responses();
3840                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3841                 }
3842
3843                 lnet_recover_local_nis();
3844                 lnet_recover_peer_nis();
3845
3846                 /*
3847                  * TODO do we need to check if we should sleep without
3848                  * timeout?  Technically, an active system will always
3849                  * have messages in flight so this check will always
3850                  * evaluate to false. And on an idle system do we care
3851                  * if we wake up every 1 second? Although, we've seen
3852                  * cases where we get a complaint that an idle thread
3853                  * is waking up unnecessarily.
3854                  */
3855                 wait_for_completion_interruptible_timeout(
3856                         &the_lnet.ln_mt_wait_complete,
3857                         cfs_time_seconds(1));
3858                 /* Must re-init the completion before testing anything,
3859                  * including ln_mt_state.
3860                  */
3861                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3862         }
3863
3864         /* Shutting down */
3865         lnet_net_lock(LNET_LOCK_EX);
3866         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3867         lnet_net_unlock(LNET_LOCK_EX);
3868
3869         /* signal that the monitor thread is exiting */
3870         up(&the_lnet.ln_mt_signal);
3871
3872         return 0;
3873 }
3874
3875 /*
3876  * lnet_send_ping
3877  * Sends a ping.
3878  * Returns == 0 if success
3879  * Returns > 0 if LNetMDBind or prior fails
3880  * Returns < 0 if LNetGet fails
3881  */
3882 int
3883 lnet_send_ping(lnet_nid_t dest_nid,
3884                struct lnet_handle_md *mdh, int nnis,
3885                void *user_data, lnet_handler_t handler, bool recovery)
3886 {
3887         struct lnet_md md = { NULL };
3888         struct lnet_process_id id;
3889         struct lnet_ping_buffer *pbuf;
3890         int rc;
3891
3892         if (dest_nid == LNET_NID_ANY) {
3893                 rc = -EHOSTUNREACH;
3894                 goto fail_error;
3895         }
3896
3897         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3898         if (!pbuf) {
3899                 rc = ENOMEM;
3900                 goto fail_error;
3901         }
3902
3903         /* initialize md content */
3904         md.start     = &pbuf->pb_info;
3905         md.length    = LNET_PING_INFO_SIZE(nnis);
3906         md.threshold = 2; /* GET/REPLY */
3907         md.max_size  = 0;
3908         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3909         md.user_ptr  = user_data;
3910         md.handler   = handler;
3911
3912         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3913         if (rc) {
3914                 lnet_ping_buffer_decref(pbuf);
3915                 CERROR("Can't bind MD: %d\n", rc);
3916                 rc = -rc; /* change the rc to positive */
3917                 goto fail_error;
3918         }
3919         id.pid = LNET_PID_LUSTRE;
3920         id.nid = dest_nid;
3921
3922         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3923                      LNET_RESERVED_PORTAL,
3924                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3925
3926         if (rc)
3927                 goto fail_unlink_md;
3928
3929         return 0;
3930
3931 fail_unlink_md:
3932         LNetMDUnlink(*mdh);
3933         LNetInvalidateMDHandle(mdh);
3934 fail_error:
3935         return rc;
3936 }
3937
3938 static void
3939 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3940                            int status, bool send, bool unlink_event)
3941 {
3942         lnet_nid_t nid = ev_info->mt_nid;
3943
3944         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3945                 struct lnet_ni *ni;
3946
3947                 lnet_net_lock(0);
3948                 ni = lnet_nid2ni_locked(nid, 0);
3949                 if (!ni) {
3950                         lnet_net_unlock(0);
3951                         return;
3952                 }
3953                 lnet_ni_lock(ni);
3954                 if (!send || (send && status != 0))
3955                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3956                 if (status)
3957                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3958                 lnet_ni_unlock(ni);
3959                 lnet_net_unlock(0);
3960
3961                 if (status != 0) {
3962                         CERROR("local NI (%s) recovery failed with %d\n",
3963                                libcfs_nid2str(nid), status);
3964                         return;
3965                 }
3966                 /*
3967                  * need to increment healthv for the ni here, because in
3968                  * the lnet_finalize() path we don't have access to this
3969                  * NI. And in order to get access to it, we'll need to
3970                  * carry forward too much information.
3971                  * In the peer case, it'll naturally be incremented
3972                  */
3973                 if (!unlink_event)
3974                         lnet_inc_healthv(&ni->ni_healthv,
3975                                          lnet_health_sensitivity);
3976         } else {
3977                 struct lnet_peer_ni *lpni;
3978                 int cpt;
3979
3980                 cpt = lnet_net_lock_current();
3981                 lpni = lnet_find_peer_ni_locked(nid);
3982                 if (!lpni) {
3983                         lnet_net_unlock(cpt);
3984                         return;
3985                 }
3986                 spin_lock(&lpni->lpni_lock);
3987                 if (!send || (send && status != 0))
3988                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3989                 if (status)
3990                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3991                 spin_unlock(&lpni->lpni_lock);
3992                 lnet_peer_ni_decref_locked(lpni);
3993                 lnet_net_unlock(cpt);
3994
3995                 if (status != 0)
3996                         CERROR("peer NI (%s) recovery failed with %d\n",
3997                                libcfs_nid2str(nid), status);
3998         }
3999 }
4000
4001 void
4002 lnet_mt_event_handler(struct lnet_event *event)
4003 {
4004         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
4005         struct lnet_ping_buffer *pbuf;
4006
4007         /* TODO: remove assert */
4008         LASSERT(event->type == LNET_EVENT_REPLY ||
4009                 event->type == LNET_EVENT_SEND ||
4010                 event->type == LNET_EVENT_UNLINK);
4011
4012         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4013                event->status);
4014
4015         switch (event->type) {
4016         case LNET_EVENT_UNLINK:
4017                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4018                        libcfs_nid2str(ev_info->mt_nid));
4019                 /* fallthrough */
4020         case LNET_EVENT_REPLY:
4021                 lnet_handle_recovery_reply(ev_info, event->status, false,
4022                                            event->type == LNET_EVENT_UNLINK);
4023                 break;
4024         case LNET_EVENT_SEND:
4025                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4026                                libcfs_nid2str(ev_info->mt_nid),
4027                                (event->status) ? "unsuccessfully" :
4028                                "successfully", event->status);
4029                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4030                 break;
4031         default:
4032                 CERROR("Unexpected event: %d\n", event->type);
4033                 break;
4034         }
4035         if (event->unlinked) {
4036                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4037                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4038                 lnet_ping_buffer_decref(pbuf);
4039         }
4040 }
4041
4042 static int
4043 lnet_rsp_tracker_create(void)
4044 {
4045         struct list_head **rstqs;
4046         rstqs = lnet_create_array_of_queues();
4047
4048         if (!rstqs)
4049                 return -ENOMEM;
4050
4051         the_lnet.ln_mt_rstq = rstqs;
4052
4053         return 0;
4054 }
4055
4056 static void
4057 lnet_rsp_tracker_clean(void)
4058 {
4059         lnet_finalize_expired_responses();
4060
4061         cfs_percpt_free(the_lnet.ln_mt_rstq);
4062         the_lnet.ln_mt_rstq = NULL;
4063 }
4064
4065 int lnet_monitor_thr_start(void)
4066 {
4067         int rc = 0;
4068         struct task_struct *task;
4069
4070         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4071                 return -EALREADY;
4072
4073         rc = lnet_resendqs_create();
4074         if (rc)
4075                 return rc;
4076
4077         rc = lnet_rsp_tracker_create();
4078         if (rc)
4079                 goto clean_queues;
4080
4081         sema_init(&the_lnet.ln_mt_signal, 0);
4082
4083         lnet_net_lock(LNET_LOCK_EX);
4084         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4085         lnet_net_unlock(LNET_LOCK_EX);
4086         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4087         if (IS_ERR(task)) {
4088                 rc = PTR_ERR(task);
4089                 CERROR("Can't start monitor thread: %d\n", rc);
4090                 goto clean_thread;
4091         }
4092
4093         return 0;
4094
4095 clean_thread:
4096         lnet_net_lock(LNET_LOCK_EX);
4097         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4098         lnet_net_unlock(LNET_LOCK_EX);
4099         /* block until event callback signals exit */
4100         down(&the_lnet.ln_mt_signal);
4101         /* clean up */
4102         lnet_net_lock(LNET_LOCK_EX);
4103         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4104         lnet_net_unlock(LNET_LOCK_EX);
4105         lnet_rsp_tracker_clean();
4106         lnet_clean_local_ni_recoveryq();
4107         lnet_clean_peer_ni_recoveryq();
4108         lnet_clean_resendqs();
4109         the_lnet.ln_mt_handler = NULL;
4110         return rc;
4111 clean_queues:
4112         lnet_rsp_tracker_clean();
4113         lnet_clean_local_ni_recoveryq();
4114         lnet_clean_peer_ni_recoveryq();
4115         lnet_clean_resendqs();
4116         return rc;
4117 }
4118
4119 void lnet_monitor_thr_stop(void)
4120 {
4121         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4122                 return;
4123
4124         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4125         lnet_net_lock(LNET_LOCK_EX);
4126         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4127         lnet_net_unlock(LNET_LOCK_EX);
4128
4129         /* tell the monitor thread that we're shutting down */
4130         complete(&the_lnet.ln_mt_wait_complete);
4131
4132         /* block until monitor thread signals that it's done */
4133         down(&the_lnet.ln_mt_signal);
4134         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4135
4136         /* perform cleanup tasks */
4137         lnet_rsp_tracker_clean();
4138         lnet_clean_local_ni_recoveryq();
4139         lnet_clean_peer_ni_recoveryq();
4140         lnet_clean_resendqs();
4141 }
4142
4143 void
4144 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4145                   __u32 msg_type)
4146 {
4147         lnet_net_lock(cpt);
4148         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4149         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4150         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4151         lnet_net_unlock(cpt);
4152
4153         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4154 }
4155
4156 static void
4157 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4158 {
4159         struct lnet_hdr *hdr = &msg->msg_hdr;
4160
4161         if (msg->msg_wanted != 0)
4162                 lnet_setpayloadbuffer(msg);
4163
4164         lnet_build_msg_event(msg, LNET_EVENT_PUT);
4165
4166         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
4167          * it back into the ACK during lnet_finalize() */
4168         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4169                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4170
4171         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4172                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4173 }
4174
4175 static int
4176 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4177 {
4178         struct lnet_hdr         *hdr = &msg->msg_hdr;
4179         struct lnet_match_info  info;
4180         int                     rc;
4181         bool                    ready_delay;
4182
4183         /* Convert put fields to host byte order */
4184         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4185         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
4186         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
4187
4188         /* Primary peer NID. */
4189         info.mi_id.nid = msg->msg_initiator;
4190         info.mi_id.pid  = hdr->src_pid;
4191         info.mi_opc     = LNET_MD_OP_PUT;
4192         info.mi_portal  = hdr->msg.put.ptl_index;
4193         info.mi_rlength = hdr->payload_length;
4194         info.mi_roffset = hdr->msg.put.offset;
4195         info.mi_mbits   = hdr->msg.put.match_bits;
4196         info.mi_cpt     = lnet_nid2cpt(&msg->msg_initiator, ni);
4197
4198         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4199         ready_delay = msg->msg_rx_ready_delay;
4200
4201  again:
4202         rc = lnet_ptl_match_md(&info, msg);
4203         switch (rc) {
4204         default:
4205                 LBUG();
4206
4207         case LNET_MATCHMD_OK:
4208                 lnet_recv_put(ni, msg);
4209                 return 0;
4210
4211         case LNET_MATCHMD_NONE:
4212                 if (ready_delay)
4213                         /* no eager_recv or has already called it, should
4214                          * have been attached on delayed list */
4215                         return 0;
4216
4217                 rc = lnet_ni_eager_recv(ni, msg);
4218                 if (rc == 0) {
4219                         ready_delay = true;
4220                         goto again;
4221                 }
4222                 /* fall through */
4223
4224         case LNET_MATCHMD_DROP:
4225                 CNETERR("Dropping PUT from %s portal %d match %llu"
4226                         " offset %d length %d: %d\n",
4227                         libcfs_idstr(&info.mi_id), info.mi_portal,
4228                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4229
4230                 return -ENOENT; /* -ve: OK but no match */
4231         }
4232 }
4233
4234 static int
4235 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4236 {
4237         struct lnet_match_info info;
4238         struct lnet_hdr *hdr = &msg->msg_hdr;
4239         struct lnet_process_id source_id;
4240         struct lnet_handle_wire reply_wmd;
4241         int rc;
4242
4243         /* Convert get fields to host byte order */
4244         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
4245         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
4246         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
4247         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
4248
4249         source_id.nid = hdr->src_nid;
4250         source_id.pid = hdr->src_pid;
4251         /* Primary peer NID */
4252         info.mi_id.nid  = msg->msg_initiator;
4253         info.mi_id.pid  = hdr->src_pid;
4254         info.mi_opc     = LNET_MD_OP_GET;
4255         info.mi_portal  = hdr->msg.get.ptl_index;
4256         info.mi_rlength = hdr->msg.get.sink_length;
4257         info.mi_roffset = hdr->msg.get.src_offset;
4258         info.mi_mbits   = hdr->msg.get.match_bits;
4259         info.mi_cpt     = lnet_nid2cpt(&msg->msg_initiator, ni);
4260
4261         rc = lnet_ptl_match_md(&info, msg);
4262         if (rc == LNET_MATCHMD_DROP) {
4263                 CNETERR("Dropping GET from %s portal %d match %llu"
4264                         " offset %d length %d\n",
4265                         libcfs_idstr(&info.mi_id), info.mi_portal,
4266                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
4267                 return -ENOENT; /* -ve: OK but no match */
4268         }
4269
4270         LASSERT(rc == LNET_MATCHMD_OK);
4271
4272         lnet_build_msg_event(msg, LNET_EVENT_GET);
4273
4274         reply_wmd = hdr->msg.get.return_wmd;
4275
4276         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4277                        msg->msg_offset, msg->msg_wanted);
4278
4279         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4280
4281         if (rdma_get) {
4282                 /* The LND completes the REPLY from her recv procedure */
4283                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4284                              msg->msg_offset, msg->msg_len, msg->msg_len);
4285                 return 0;
4286         }
4287
4288         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4289         msg->msg_receiving = 0;
4290
4291         /* FIXME need to handle large-addr nid */
4292         rc = lnet_send(lnet_nid_to_nid4(&ni->ni_nid), msg,
4293                        lnet_nid_to_nid4(&msg->msg_from));
4294         if (rc < 0) {
4295                 /* didn't get as far as lnet_ni_send() */
4296                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4297                        libcfs_nidstr(&ni->ni_nid),
4298                        libcfs_idstr(&info.mi_id), rc);
4299
4300                 lnet_finalize(msg, rc);
4301         }
4302
4303         return 0;
4304 }
4305
4306 static int
4307 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4308 {
4309         void *private = msg->msg_private;
4310         struct lnet_hdr *hdr = &msg->msg_hdr;
4311         struct lnet_process_id src = {0};
4312         struct lnet_libmd *md;
4313         unsigned int rlength;
4314         unsigned int mlength;
4315         int cpt;
4316
4317         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4318         lnet_res_lock(cpt);
4319
4320         src.nid = hdr->src_nid;
4321         src.pid = hdr->src_pid;
4322
4323         /* NB handles only looked up by creator (no flips) */
4324         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4325         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4326                 CNETERR("%s: Dropping REPLY from %s for %s "
4327                         "MD %#llx.%#llx\n",
4328                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4329                         (md == NULL) ? "invalid" : "inactive",
4330                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4331                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4332                 if (md != NULL && md->md_me != NULL)
4333                         CERROR("REPLY MD also attached to portal %d\n",
4334                                md->md_me->me_portal);
4335
4336                 lnet_res_unlock(cpt);
4337                 return -ENOENT; /* -ve: OK but no match */
4338         }
4339
4340         LASSERT(md->md_offset == 0);
4341
4342         rlength = hdr->payload_length;
4343         mlength = min(rlength, md->md_length);
4344
4345         if (mlength < rlength &&
4346             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4347                 CNETERR("%s: Dropping REPLY from %s length %d "
4348                         "for MD %#llx would overflow (%d)\n",
4349                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4350                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4351                         mlength);
4352                 lnet_res_unlock(cpt);
4353                 return -ENOENT; /* -ve: OK but no match */
4354         }
4355
4356         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4357                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4358                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4359
4360         lnet_msg_attach_md(msg, md, 0, mlength);
4361
4362         if (mlength != 0)
4363                 lnet_setpayloadbuffer(msg);
4364
4365         lnet_res_unlock(cpt);
4366
4367         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4368
4369         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4370         return 0;
4371 }
4372
4373 static int
4374 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4375 {
4376         struct lnet_hdr *hdr = &msg->msg_hdr;
4377         struct lnet_process_id src = {0};
4378         struct lnet_libmd *md;
4379         int cpt;
4380
4381         src.nid = hdr->src_nid;
4382         src.pid = hdr->src_pid;
4383
4384         /* Convert ack fields to host byte order */
4385         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4386         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4387
4388         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4389         lnet_res_lock(cpt);
4390
4391         /* NB handles only looked up by creator (no flips) */
4392         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4393         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4394                 /* Don't moan; this is expected */
4395                 CDEBUG(D_NET,
4396                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4397                        libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4398                        (md == NULL) ? "invalid" : "inactive",
4399                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4400                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4401                 if (md != NULL && md->md_me != NULL)
4402                         CERROR("Source MD also attached to portal %d\n",
4403                                md->md_me->me_portal);
4404
4405                 lnet_res_unlock(cpt);
4406                 return -ENOENT;                  /* -ve! */
4407         }
4408
4409         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4410                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4411                hdr->msg.ack.dst_wmd.wh_object_cookie);
4412
4413         lnet_msg_attach_md(msg, md, 0, 0);
4414
4415         lnet_res_unlock(cpt);
4416
4417         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4418
4419         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4420         return 0;
4421 }
4422
4423 /**
4424  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4425  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4426  * \retval -ve                  error code
4427  */
4428 int
4429 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4430 {
4431         int     rc = 0;
4432
4433         if (!the_lnet.ln_routing)
4434                 return -ECANCELED;
4435
4436         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4437             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4438                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4439                         msg->msg_rx_ready_delay = 1;
4440                 } else {
4441                         lnet_net_unlock(msg->msg_rx_cpt);
4442                         rc = lnet_ni_eager_recv(ni, msg);
4443                         lnet_net_lock(msg->msg_rx_cpt);
4444                 }
4445         }
4446
4447         if (rc == 0)
4448                 rc = lnet_post_routed_recv_locked(msg, 0);
4449         return rc;
4450 }
4451
4452 int
4453 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4454 {
4455         int     rc;
4456
4457         switch (msg->msg_type) {
4458         case LNET_MSG_ACK:
4459                 rc = lnet_parse_ack(ni, msg);
4460                 break;
4461         case LNET_MSG_PUT:
4462                 rc = lnet_parse_put(ni, msg);
4463                 break;
4464         case LNET_MSG_GET:
4465                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4466                 break;
4467         case LNET_MSG_REPLY:
4468                 rc = lnet_parse_reply(ni, msg);
4469                 break;
4470         default: /* prevent an unused label if !kernel */
4471                 LASSERT(0);
4472                 return -EPROTO;
4473         }
4474
4475         LASSERT(rc == 0 || rc == -ENOENT);
4476         return rc;
4477 }
4478
4479 char *
4480 lnet_msgtyp2str (int type)
4481 {
4482         switch (type) {
4483         case LNET_MSG_ACK:
4484                 return ("ACK");
4485         case LNET_MSG_PUT:
4486                 return ("PUT");
4487         case LNET_MSG_GET:
4488                 return ("GET");
4489         case LNET_MSG_REPLY:
4490                 return ("REPLY");
4491         case LNET_MSG_HELLO:
4492                 return ("HELLO");
4493         default:
4494                 return ("<UNKNOWN>");
4495         }
4496 }
4497 EXPORT_SYMBOL(lnet_msgtyp2str);
4498
4499 int
4500 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid4,
4501            void *private, int rdma_req)
4502 {
4503         struct lnet_peer_ni *lpni;
4504         struct lnet_msg *msg;
4505         __u32 payload_length;
4506         lnet_pid_t dest_pid;
4507         lnet_nid_t dest_nid;
4508         lnet_nid_t src_nid;
4509         struct lnet_nid from_nid;
4510         bool push = false;
4511         int for_me;
4512         __u32 type;
4513         int rc = 0;
4514         int cpt;
4515
4516         LASSERT (!in_interrupt ());
4517
4518         lnet_nid4_to_nid(from_nid4, &from_nid);
4519
4520         type = le32_to_cpu(hdr->type);
4521         src_nid = le64_to_cpu(hdr->src_nid);
4522         dest_nid = le64_to_cpu(hdr->dest_nid);
4523         dest_pid = le32_to_cpu(hdr->dest_pid);
4524         payload_length = le32_to_cpu(hdr->payload_length);
4525
4526         /* FIXME handle large-addr nids */
4527         for_me = (lnet_nid_to_nid4(&ni->ni_nid) == dest_nid);
4528         cpt = lnet_cpt_of_nid(from_nid4, ni);
4529
4530         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4531                 libcfs_nid2str(dest_nid),
4532                 libcfs_nidstr(&ni->ni_nid),
4533                 libcfs_nid2str(src_nid),
4534                 lnet_msgtyp2str(type),
4535                 (for_me) ? "for me" : "routed");
4536
4537         switch (type) {
4538         case LNET_MSG_ACK:
4539         case LNET_MSG_GET:
4540                 if (payload_length > 0) {
4541                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4542                                libcfs_nid2str(from_nid4),
4543                                libcfs_nid2str(src_nid),
4544                                lnet_msgtyp2str(type), payload_length);
4545                         return -EPROTO;
4546                 }
4547                 break;
4548
4549         case LNET_MSG_PUT:
4550         case LNET_MSG_REPLY:
4551                 if (payload_length >
4552                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4553                         CERROR("%s, src %s: bad %s payload %d "
4554                                "(%d max expected)\n",
4555                                libcfs_nid2str(from_nid4),
4556                                libcfs_nid2str(src_nid),
4557                                lnet_msgtyp2str(type),
4558                                payload_length,
4559                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4560                         return -EPROTO;
4561                 }
4562                 break;
4563
4564         default:
4565                 CERROR("%s, src %s: Bad message type 0x%x\n",
4566                        libcfs_nid2str(from_nid4),
4567                        libcfs_nid2str(src_nid), type);
4568                 return -EPROTO;
4569         }
4570
4571         if (the_lnet.ln_routing &&
4572             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4573                 lnet_ni_lock(ni);
4574                 spin_lock(&ni->ni_net->net_lock);
4575                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4576                 spin_unlock(&ni->ni_net->net_lock);
4577                 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4578                 lnet_ni_unlock(ni);
4579         }
4580
4581         if (push)
4582                 lnet_push_update_to_peers(1);
4583
4584         /* Regard a bad destination NID as a protocol error.  Senders should
4585          * know what they're doing; if they don't they're misconfigured, buggy
4586          * or malicious so we chop them off at the knees :) */
4587
4588         if (!for_me) {
4589                 if (LNET_NIDNET(dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4590                         /* should have gone direct */
4591                         CERROR("%s, src %s: Bad dest nid %s "
4592                                "(should have been sent direct)\n",
4593                                 libcfs_nid2str(from_nid4),
4594                                 libcfs_nid2str(src_nid),
4595                                 libcfs_nid2str(dest_nid));
4596                         return -EPROTO;
4597                 }
4598
4599                 if (lnet_islocalnid4(dest_nid)) {
4600                         /* dest is another local NI; sender should have used
4601                          * this node's NID on its own network */
4602                         CERROR("%s, src %s: Bad dest nid %s "
4603                                "(it's my nid but on a different network)\n",
4604                                 libcfs_nid2str(from_nid4),
4605                                 libcfs_nid2str(src_nid),
4606                                 libcfs_nid2str(dest_nid));
4607                         return -EPROTO;
4608                 }
4609
4610                 if (rdma_req && type == LNET_MSG_GET) {
4611                         CERROR("%s, src %s: Bad optimized GET for %s "
4612                                "(final destination must be me)\n",
4613                                 libcfs_nid2str(from_nid4),
4614                                 libcfs_nid2str(src_nid),
4615                                 libcfs_nid2str(dest_nid));
4616                         return -EPROTO;
4617                 }
4618
4619                 if (!the_lnet.ln_routing) {
4620                         CERROR("%s, src %s: Dropping message for %s "
4621                                "(routing not enabled)\n",
4622                                 libcfs_nid2str(from_nid4),
4623                                 libcfs_nid2str(src_nid),
4624                                 libcfs_nid2str(dest_nid));
4625                         goto drop;
4626                 }
4627         }
4628
4629         /* Message looks OK; we're not going to return an error, so we MUST
4630          * call back lnd_recv() come what may... */
4631
4632         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4633             fail_peer(src_nid, 0)) {                    /* shall we now? */
4634                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4635                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4636                        lnet_msgtyp2str(type));
4637                 goto drop;
4638         }
4639
4640         /* FIXME need to support large-addr nid */
4641         if (!list_empty(&the_lnet.ln_drop_rules) &&
4642             lnet_drop_rule_match(hdr, lnet_nid_to_nid4(&ni->ni_nid), NULL)) {
4643                 CDEBUG(D_NET,
4644                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4645                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4646                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4647                 goto drop;
4648         }
4649
4650         msg = lnet_msg_alloc();
4651         if (msg == NULL) {
4652                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4653                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4654                        lnet_msgtyp2str(type));
4655                 goto drop;
4656         }
4657
4658         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4659          * pointers NULL etc */
4660
4661         msg->msg_type = type;
4662         msg->msg_private = private;
4663         msg->msg_receiving = 1;
4664         msg->msg_rdma_get = rdma_req;
4665         msg->msg_len = msg->msg_wanted = payload_length;
4666         msg->msg_offset = 0;
4667         msg->msg_hdr = *hdr;
4668         /* for building message event */
4669         msg->msg_from = from_nid;
4670         if (!for_me) {
4671                 msg->msg_target.pid     = dest_pid;
4672                 lnet_nid4_to_nid(dest_nid, &msg->msg_target.nid);
4673                 msg->msg_routing        = 1;
4674
4675         } else {
4676                 /* convert common msg->hdr fields to host byteorder */
4677                 msg->msg_hdr.type       = type;
4678                 msg->msg_hdr.src_nid    = src_nid;
4679                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4680                 msg->msg_hdr.dest_nid   = dest_nid;
4681                 msg->msg_hdr.dest_pid   = dest_pid;
4682                 msg->msg_hdr.payload_length = payload_length;
4683         }
4684
4685         lnet_net_lock(cpt);
4686         lpni = lnet_peerni_by_nid_locked(&from_nid, &ni->ni_nid, cpt);
4687         if (IS_ERR(lpni)) {
4688                 lnet_net_unlock(cpt);
4689                 rc = PTR_ERR(lpni);
4690                 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4691                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4692                        lnet_msgtyp2str(type), rc);
4693                 lnet_msg_free(msg);
4694                 if (rc == -ESHUTDOWN)
4695                         /* We are shutting down.  Don't do anything more */
4696                         return 0;
4697                 goto drop;
4698         }
4699
4700         /* If this message was forwarded to us from a router then we may need
4701          * to update router aliveness or check for an asymmetrical route
4702          * (or both)
4703          */
4704         if (((lnet_drop_asym_route && for_me) ||
4705              !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4706             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid4)) {
4707                 __u32 src_net_id = LNET_NIDNET(src_nid);
4708                 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4709                 struct lnet_route *route;
4710                 bool found = false;
4711
4712                 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4713                         if (route->lr_net == src_net_id) {
4714                                 found = true;
4715                                 /* If we're transitioning the gateway from
4716                                  * dead -> alive, and discovery is disabled
4717                                  * locally or on the gateway, then we need to
4718                                  * update the cached route aliveness for each
4719                                  * route to the src_nid's net.
4720                                  *
4721                                  * Otherwise, we're only checking for
4722                                  * symmetrical route, and we can break the
4723                                  * loop
4724                                  */
4725                                 if (!gw->lp_alive &&
4726                                     lnet_is_discovery_disabled(gw))
4727                                         lnet_set_route_aliveness(route, true);
4728                                 else
4729                                         break;
4730                         }
4731                 }
4732                 if (lnet_drop_asym_route && for_me && !found) {
4733                         /* Drop ref taken by lnet_nid2peerni_locked() */
4734                         lnet_peer_ni_decref_locked(lpni);
4735                         lnet_net_unlock(cpt);
4736                         /* we would not use from_nid to route a message to
4737                          * src_nid
4738                          * => asymmetric routing detected but forbidden
4739                          */
4740                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4741                                libcfs_nid2str(from_nid4),
4742                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4743                         lnet_msg_free(msg);
4744                         goto drop;
4745                 }
4746                 if (!gw->lp_alive) {
4747                         struct lnet_peer_net *lpn;
4748                         struct lnet_peer_ni *lpni2;
4749
4750                         gw->lp_alive = true;
4751                         /* Mark all remote NIs on src_nid's net UP */
4752                         lpn = lnet_peer_get_net_locked(gw, src_net_id);
4753                         if (lpn)
4754                                 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4755                                                     lpni_peer_nis)
4756                                         lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4757                 }
4758         }
4759
4760         lpni->lpni_last_alive = ktime_get_seconds();
4761
4762         msg->msg_rxpeer = lpni;
4763         msg->msg_rxni = ni;
4764         lnet_ni_addref_locked(ni, cpt);
4765         /* Multi-Rail: Primary NID of source. */
4766         lnet_peer_primary_nid_locked(src_nid, &msg->msg_initiator);
4767
4768         /*
4769          * mark the status of this lpni as UP since we received a message
4770          * from it. The ping response reports back the ns_status which is
4771          * marked on the remote as up or down and we cache it here.
4772          */
4773         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4774
4775         lnet_msg_commit(msg, cpt);
4776
4777         /* message delay simulation */
4778         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4779                      lnet_delay_rule_match_locked(hdr, msg))) {
4780                 lnet_net_unlock(cpt);
4781                 return 0;
4782         }
4783
4784         if (!for_me) {
4785                 rc = lnet_parse_forward_locked(ni, msg);
4786                 lnet_net_unlock(cpt);
4787
4788                 if (rc < 0)
4789                         goto free_drop;
4790
4791                 if (rc == LNET_CREDIT_OK) {
4792                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4793                                      0, payload_length, payload_length);
4794                 }
4795                 return 0;
4796         }
4797
4798         lnet_net_unlock(cpt);
4799
4800         rc = lnet_parse_local(ni, msg);
4801         if (rc != 0)
4802                 goto free_drop;
4803         return 0;
4804
4805  free_drop:
4806         LASSERT(msg->msg_md == NULL);
4807         lnet_finalize(msg, rc);
4808
4809  drop:
4810         lnet_drop_message(ni, cpt, private, payload_length, type);
4811         return 0;
4812 }
4813 EXPORT_SYMBOL(lnet_parse);
4814
4815 void
4816 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4817 {
4818         while (!list_empty(head)) {
4819                 struct lnet_process_id id = {0};
4820                 struct lnet_msg *msg;
4821
4822                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4823                 list_del(&msg->msg_list);
4824
4825                 id.nid = msg->msg_hdr.src_nid;
4826                 id.pid = msg->msg_hdr.src_pid;
4827
4828                 LASSERT(msg->msg_md == NULL);
4829                 LASSERT(msg->msg_rx_delayed);
4830                 LASSERT(msg->msg_rxpeer != NULL);
4831                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4832
4833                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4834                       " offset %d length %d: %s\n",
4835                       libcfs_id2str(id),
4836                       msg->msg_hdr.msg.put.ptl_index,
4837                       msg->msg_hdr.msg.put.match_bits,
4838                       msg->msg_hdr.msg.put.offset,
4839                       msg->msg_hdr.payload_length, reason);
4840
4841                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4842                  * called lnet_drop_message(), so I just hang onto msg as well
4843                  * until that's done */
4844
4845                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4846                                   msg->msg_private, msg->msg_len,
4847                                   msg->msg_type);
4848
4849                 msg->msg_no_resend = true;
4850                 /*
4851                  * NB: message will not generate event because w/o attached MD,
4852                  * but we still should give error code so lnet_msg_decommit()
4853                  * can skip counters operations and other checks.
4854                  */
4855                 lnet_finalize(msg, -ENOENT);
4856         }
4857 }
4858
4859 void
4860 lnet_recv_delayed_msg_list(struct list_head *head)
4861 {
4862         while (!list_empty(head)) {
4863                 struct lnet_msg *msg;
4864                 struct lnet_process_id id;
4865
4866                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4867                 list_del(&msg->msg_list);
4868
4869                 /* md won't disappear under me, since each msg
4870                  * holds a ref on it */
4871
4872                 id.nid = msg->msg_hdr.src_nid;
4873                 id.pid = msg->msg_hdr.src_pid;
4874
4875                 LASSERT(msg->msg_rx_delayed);
4876                 LASSERT(msg->msg_md != NULL);
4877                 LASSERT(msg->msg_rxpeer != NULL);
4878                 LASSERT(msg->msg_rxni != NULL);
4879                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4880
4881                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4882                        "match %llu offset %d length %d.\n",
4883                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4884                         msg->msg_hdr.msg.put.match_bits,
4885                         msg->msg_hdr.msg.put.offset,
4886                         msg->msg_hdr.payload_length);
4887
4888                 lnet_recv_put(msg->msg_rxni, msg);
4889         }
4890 }
4891
4892 static void
4893 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4894                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4895 {
4896         s64 timeout_ns;
4897         struct lnet_rsp_tracker *local_rspt;
4898
4899         /*
4900          * MD has a refcount taken by message so it's not going away.
4901          * The MD however can be looked up. We need to secure the access
4902          * to the md_rspt_ptr by taking the res_lock.
4903          * The rspt can be accessed without protection up to when it gets
4904          * added to the list.
4905          */
4906
4907         lnet_res_lock(cpt);
4908         local_rspt = md->md_rspt_ptr;
4909         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4910         if (local_rspt != NULL) {
4911                 /*
4912                  * we already have an rspt attached to the md, so we'll
4913                  * update the deadline on that one.
4914                  */
4915                 lnet_rspt_free(rspt, cpt);
4916         } else {
4917                 /* new md */
4918                 rspt->rspt_mdh = mdh;
4919                 rspt->rspt_cpt = cpt;
4920                 /* store the rspt so we can access it when we get the REPLY */
4921                 md->md_rspt_ptr = rspt;
4922                 local_rspt = rspt;
4923         }
4924         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4925
4926         /*
4927          * add to the list of tracked responses. It's added to tail of the
4928          * list in order to expire all the older entries first.
4929          */
4930         lnet_net_lock(cpt);
4931         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4932         lnet_net_unlock(cpt);
4933         lnet_res_unlock(cpt);
4934 }
4935
4936 /**
4937  * Initiate an asynchronous PUT operation.
4938  *
4939  * There are several events associated with a PUT: completion of the send on
4940  * the initiator node (LNET_EVENT_SEND), and when the send completes
4941  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4942  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4943  * used at the target node to indicate the completion of incoming data
4944  * delivery.
4945  *
4946  * The local events will be logged in the EQ associated with the MD pointed to
4947  * by \a mdh handle. Using a MD without an associated EQ results in these
4948  * events being discarded. In this case, the caller must have another
4949  * mechanism (e.g., a higher level protocol) for determining when it is safe
4950  * to modify the memory region associated with the MD.
4951  *
4952  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4953  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4954  *
4955  * \param self Indicates the NID of a local interface through which to send
4956  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4957  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4958  * must be "free floating" (See LNetMDBind()).
4959  * \param ack Controls whether an acknowledgment is requested.
4960  * Acknowledgments are only sent when they are requested by the initiating
4961  * process and the target MD enables them.
4962  * \param target A process identifier for the target process.
4963  * \param portal The index in the \a target's portal table.
4964  * \param match_bits The match bits to use for MD selection at the target
4965  * process.
4966  * \param offset The offset into the target MD (only used when the target
4967  * MD has the LNET_MD_MANAGE_REMOTE option set).
4968  * \param hdr_data 64 bits of user data that can be included in the message
4969  * header. This data is written to an event queue entry at the target if an
4970  * EQ is present on the matching MD.
4971  *
4972  * \retval  0      Success, and only in this case events will be generated
4973  * and logged to EQ (if it exists).
4974  * \retval -EIO    Simulated failure.
4975  * \retval -ENOMEM Memory allocation failure.
4976  * \retval -ENOENT Invalid MD object.
4977  *
4978  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4979  */
4980 int
4981 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4982         struct lnet_process_id target, unsigned int portal,
4983         __u64 match_bits, unsigned int offset,
4984         __u64 hdr_data)
4985 {
4986         struct lnet_msg *msg;
4987         struct lnet_libmd *md;
4988         int cpt;
4989         int rc;
4990         struct lnet_rsp_tracker *rspt = NULL;
4991
4992         LASSERT(the_lnet.ln_refcount > 0);
4993
4994         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4995             fail_peer(target.nid, 1)) {                 /* shall we now? */
4996                 CERROR("Dropping PUT to %s: simulated failure\n",
4997                        libcfs_id2str(target));
4998                 return -EIO;
4999         }
5000
5001         msg = lnet_msg_alloc();
5002         if (msg == NULL) {
5003                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
5004                        libcfs_id2str(target));
5005                 return -ENOMEM;
5006         }
5007         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
5008
5009         cpt = lnet_cpt_of_cookie(mdh.cookie);
5010
5011         if (ack == LNET_ACK_REQ) {
5012                 rspt = lnet_rspt_alloc(cpt);
5013                 if (!rspt) {
5014                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5015                                 libcfs_id2str(target));
5016                         return -ENOMEM;
5017                 }
5018                 INIT_LIST_HEAD(&rspt->rspt_on_list);
5019         }
5020
5021         lnet_res_lock(cpt);
5022
5023         md = lnet_handle2md(&mdh);
5024         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5025                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5026                        match_bits, portal, libcfs_id2str(target),
5027                        md == NULL ? -1 : md->md_threshold);
5028                 if (md != NULL && md->md_me != NULL)
5029                         CERROR("Source MD also attached to portal %d\n",
5030                                md->md_me->me_portal);
5031                 lnet_res_unlock(cpt);
5032
5033                 if (rspt)
5034                         lnet_rspt_free(rspt, cpt);
5035
5036                 lnet_msg_free(msg);
5037                 return -ENOENT;
5038         }
5039
5040         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
5041
5042         lnet_msg_attach_md(msg, md, 0, 0);
5043
5044         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5045
5046         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5047         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5048         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5049         msg->msg_hdr.msg.put.hdr_data = hdr_data;
5050
5051         /* NB handles only looked up by creator (no flips) */
5052         if (ack == LNET_ACK_REQ) {
5053                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5054                         the_lnet.ln_interface_cookie;
5055                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5056                         md->md_lh.lh_cookie;
5057         } else {
5058                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5059                         LNET_WIRE_HANDLE_COOKIE_NONE;
5060                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5061                         LNET_WIRE_HANDLE_COOKIE_NONE;
5062         }
5063
5064         lnet_res_unlock(cpt);
5065
5066         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5067
5068         if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5069                                                    md->md_options))
5070                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5071         else if (rspt)
5072                 lnet_rspt_free(rspt, cpt);
5073
5074         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5075                                  CFS_FAIL_ONCE))
5076                 rc = -EIO;
5077         else
5078                 rc = lnet_send(self, msg, LNET_NID_ANY);
5079
5080         if (rc != 0) {
5081                 CNETERR("Error sending PUT to %s: %d\n",
5082                         libcfs_id2str(target), rc);
5083                 msg->msg_no_resend = true;
5084                 lnet_finalize(msg, rc);
5085         }
5086
5087         /* completion will be signalled by an event */
5088         return 0;
5089 }
5090 EXPORT_SYMBOL(LNetPut);
5091
5092 /*
5093  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
5094  * returns a msg for the LND to pass to lnet_finalize() when the sink
5095  * data has been received.
5096  *
5097  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5098  * lnet_finalize() is called on it, so the LND must call this first
5099  */
5100 struct lnet_msg *
5101 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5102 {
5103         struct lnet_msg *msg = lnet_msg_alloc();
5104         struct lnet_libmd *getmd = getmsg->msg_md;
5105         struct lnet_processid *peer_id = &getmsg->msg_target;
5106         int cpt;
5107
5108         LASSERT(!getmsg->msg_target_is_router);
5109         LASSERT(!getmsg->msg_routing);
5110
5111         if (msg == NULL) {
5112                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5113                        libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id));
5114                 goto drop;
5115         }
5116
5117         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5118         lnet_res_lock(cpt);
5119
5120         LASSERT(getmd->md_refcount > 0);
5121
5122         if (getmd->md_threshold == 0) {
5123                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5124                         libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id),
5125                         getmd);
5126                 lnet_res_unlock(cpt);
5127                 goto drop;
5128         }
5129
5130         LASSERT(getmd->md_offset == 0);
5131
5132         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5133                libcfs_nidstr(&ni->ni_nid), libcfs_idstr(peer_id), getmd);
5134
5135         /* setup information for lnet_build_msg_event */
5136         msg->msg_initiator =
5137                 getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
5138         msg->msg_from = peer_id->nid;
5139         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5140         msg->msg_hdr.src_nid = lnet_nid_to_nid4(&peer_id->nid);
5141         msg->msg_hdr.payload_length = getmd->md_length;
5142         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5143
5144         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5145         lnet_res_unlock(cpt);
5146
5147         cpt = lnet_nid2cpt(&peer_id->nid, ni);
5148
5149         lnet_net_lock(cpt);
5150         lnet_msg_commit(msg, cpt);
5151         lnet_net_unlock(cpt);
5152
5153         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5154
5155         return msg;
5156
5157  drop:
5158         cpt = lnet_nid2cpt(&peer_id->nid, ni);
5159
5160         lnet_net_lock(cpt);
5161         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5162         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5163         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5164                 getmd->md_length;
5165         lnet_net_unlock(cpt);
5166
5167         if (msg != NULL)
5168                 lnet_msg_free(msg);
5169
5170         return NULL;
5171 }
5172 EXPORT_SYMBOL(lnet_create_reply_msg);
5173
5174 void
5175 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5176                        unsigned int len)
5177 {
5178         /* Set the REPLY length, now the RDMA that elides the REPLY message has
5179          * completed and I know it. */
5180         LASSERT(reply != NULL);
5181         LASSERT(reply->msg_type == LNET_MSG_GET);
5182         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5183
5184         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
5185          * the end of my buffer, I might as well be dead. */
5186         LASSERT(len <= reply->msg_ev.mlength);
5187
5188         reply->msg_ev.mlength = len;
5189 }
5190 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5191
5192 /**
5193  * Initiate an asynchronous GET operation.
5194  *
5195  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5196  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5197  * the target node in the REPLY has been written to local MD.
5198  *
5199  * On the target node, an LNET_EVENT_GET is logged when the GET request
5200  * arrives and is accepted into a MD.
5201  *
5202  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5203  * \param mdh A handle for the MD that describes the memory into which the
5204  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5205  *
5206  * \retval  0      Success, and only in this case events will be generated
5207  * and logged to EQ (if it exists) of the MD.
5208  * \retval -EIO    Simulated failure.
5209  * \retval -ENOMEM Memory allocation failure.
5210  * \retval -ENOENT Invalid MD object.
5211  */
5212 int
5213 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
5214         struct lnet_process_id target, unsigned int portal,
5215         __u64 match_bits, unsigned int offset, bool recovery)
5216 {
5217         struct lnet_msg *msg;
5218         struct lnet_libmd *md;
5219         struct lnet_rsp_tracker *rspt;
5220         int cpt;
5221         int rc;
5222
5223         LASSERT(the_lnet.ln_refcount > 0);
5224
5225         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
5226             fail_peer(target.nid, 1))                   /* shall we now? */
5227         {
5228                 CERROR("Dropping GET to %s: simulated failure\n",
5229                        libcfs_id2str(target));
5230                 return -EIO;
5231         }
5232
5233         msg = lnet_msg_alloc();
5234         if (!msg) {
5235                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5236                        libcfs_id2str(target));
5237                 return -ENOMEM;
5238         }
5239
5240         cpt = lnet_cpt_of_cookie(mdh.cookie);
5241
5242         rspt = lnet_rspt_alloc(cpt);
5243         if (!rspt) {
5244                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5245                        libcfs_id2str(target));
5246                 return -ENOMEM;
5247         }
5248         INIT_LIST_HEAD(&rspt->rspt_on_list);
5249
5250         msg->msg_recovery = recovery;
5251
5252         lnet_res_lock(cpt);
5253
5254         md = lnet_handle2md(&mdh);
5255         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5256                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5257                        match_bits, portal, libcfs_id2str(target),
5258                        md == NULL ? -1 : md->md_threshold);
5259                 if (md != NULL && md->md_me != NULL)
5260                         CERROR("REPLY MD also attached to portal %d\n",
5261                                md->md_me->me_portal);
5262
5263                 lnet_res_unlock(cpt);
5264
5265                 lnet_msg_free(msg);
5266                 lnet_rspt_free(rspt, cpt);
5267                 return -ENOENT;
5268         }
5269
5270         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5271
5272         lnet_msg_attach_md(msg, md, 0, 0);
5273
5274         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5275
5276         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5277         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5278         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5279         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5280
5281         /* NB handles only looked up by creator (no flips) */
5282         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5283                 the_lnet.ln_interface_cookie;
5284         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5285                 md->md_lh.lh_cookie;
5286
5287         lnet_res_unlock(cpt);
5288
5289         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5290
5291         if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5292                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5293         else
5294                 lnet_rspt_free(rspt, cpt);
5295
5296         rc = lnet_send(self, msg, LNET_NID_ANY);
5297         if (rc < 0) {
5298                 CNETERR("Error sending GET to %s: %d\n",
5299                         libcfs_id2str(target), rc);
5300                 msg->msg_no_resend = true;
5301                 lnet_finalize(msg, rc);
5302         }
5303
5304         /* completion will be signalled by an event */
5305         return 0;
5306 }
5307 EXPORT_SYMBOL(LNetGet);
5308
5309 /**
5310  * Calculate distance to node at \a dstnid.
5311  *
5312  * \param dstnid Target NID.
5313  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5314  * is saved here.
5315  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5316  * here.
5317  *
5318  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5319  * local_nid_dist_zero is set, which is the default.
5320  * \retval positives Distance to target NID, i.e. number of hops plus one.
5321  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5322  */
5323 int
5324 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5325 {
5326         struct list_head *e;
5327         struct lnet_ni *ni = NULL;
5328         struct lnet_remotenet *rnet;
5329         __u32 dstnet = LNET_NIDNET(dstnid);
5330         int hops;
5331         int cpt;
5332         __u32 order = 2;
5333         struct list_head *rn_list;
5334         bool matched_dstnet = false;
5335
5336         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5337          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5338          * keep order 0 free for 0@lo and order 1 free for a local NID
5339          * match */
5340
5341         LASSERT(the_lnet.ln_refcount > 0);
5342
5343         cpt = lnet_net_lock_current();
5344
5345         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5346                 /* FIXME support large-addr nid */
5347                 if (lnet_nid_to_nid4(&ni->ni_nid) == dstnid) {
5348                         if (srcnidp != NULL)
5349                                 *srcnidp = dstnid;
5350                         if (orderp != NULL) {
5351                                 if (dstnid == LNET_NID_LO_0)
5352                                         *orderp = 0;
5353                                 else
5354                                         *orderp = 1;
5355                         }
5356                         lnet_net_unlock(cpt);
5357
5358                         return local_nid_dist_zero ? 0 : 1;
5359                 }
5360
5361                 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5362                         matched_dstnet = true;
5363                         /* We matched the destination net, but we may have
5364                          * additional local NIs to inspect.
5365                          *
5366                          * We record the nid and order as appropriate, but
5367                          * they may be overwritten if we match local NI above.
5368                          */
5369                         if (srcnidp)
5370                                 /* FIXME support large-addr nids */
5371                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5372
5373                         if (orderp) {
5374                                 /* Check if ni was originally created in
5375                                  * current net namespace.
5376                                  * If not, assign order above 0xffff0000,
5377                                  * to make this ni not a priority.
5378                                  */
5379                                 if (current->nsproxy &&
5380                                     !net_eq(ni->ni_net_ns,
5381                                             current->nsproxy->net_ns))
5382                                         *orderp = order + 0xffff0000;
5383                                 else
5384                                         *orderp = order;
5385                         }
5386                 }
5387
5388                 order++;
5389         }
5390
5391         if (matched_dstnet) {
5392                 lnet_net_unlock(cpt);
5393                 return 1;
5394         }
5395
5396         rn_list = lnet_net2rnethash(dstnet);
5397         list_for_each(e, rn_list) {
5398                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5399
5400                 if (rnet->lrn_net == dstnet) {
5401                         struct lnet_route *route;
5402                         struct lnet_route *shortest = NULL;
5403                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5404                         __u32 route_hops;
5405
5406                         LASSERT(!list_empty(&rnet->lrn_routes));
5407
5408                         list_for_each_entry(route, &rnet->lrn_routes,
5409                                             lr_list) {
5410                                 route_hops = route->lr_hops;
5411                                 if (route_hops == LNET_UNDEFINED_HOPS)
5412                                         route_hops = 1;
5413                                 if (shortest == NULL ||
5414                                     route_hops < shortest_hops) {
5415                                         shortest = route;
5416                                         shortest_hops = route_hops;
5417                                 }
5418                         }
5419
5420                         LASSERT(shortest != NULL);
5421                         hops = shortest_hops;
5422                         if (srcnidp != NULL) {
5423                                 struct lnet_net *net;
5424                                 net = lnet_get_net_locked(shortest->lr_lnet);
5425                                 LASSERT(net);
5426                                 ni = lnet_get_next_ni_locked(net, NULL);
5427                                 /* FIXME support large-addr nids */
5428                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5429                         }
5430                         if (orderp != NULL)
5431                                 *orderp = order;
5432                         lnet_net_unlock(cpt);
5433                         return hops + 1;
5434                 }
5435                 order++;
5436         }
5437
5438         lnet_net_unlock(cpt);
5439         return -EHOSTUNREACH;
5440 }
5441 EXPORT_SYMBOL(LNetDist);