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 = target;
 717         msg->msg_len = len;
 718         msg->msg_offset = offset;
 719
 720         if (len != 0)
 721                 lnet_setpayloadbuffer(msg);
 722
 723         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 724         msg->msg_hdr.type           = cpu_to_le32(type);
 725         /* dest_nid will be overwritten by lnet_select_pathway() */
 726         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 727         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 728         /* src_nid will be set later */
 729         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 730         msg->msg_hdr.payload_length = cpu_to_le32(len);
 731 }
 732
 733 void
 734 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 735 {
 736         void *priv = msg->msg_private;
 737         int rc;
 738
 739         LASSERT(!in_interrupt());
 740         LASSERT(nid_is_lo0(&ni->ni_nid) ||
 741                 (msg->msg_txcredit && msg->msg_peertxcredit));
 742
 743         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 744         if (rc < 0) {
 745                 msg->msg_no_resend = true;
 746                 lnet_finalize(msg, rc);
 747         }
 748 }
 749
 750 static int
 751 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 752 {
 753         int     rc;
 754
 755         LASSERT(!msg->msg_sending);
 756         LASSERT(msg->msg_receiving);
 757         LASSERT(!msg->msg_rx_ready_delay);
 758         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 759
 760         msg->msg_rx_ready_delay = 1;
 761         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 762                                                   &msg->msg_private);
 763         if (rc != 0) {
 764                 CERROR("recv from %s / send to %s aborted: "
 765                        "eager_recv failed %d\n",
 766                        libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
 767                        libcfs_id2str(msg->msg_target), rc);
 768                 LASSERT(rc < 0); /* required by my callers */
 769         }
 770
 771         return rc;
 772 }
 773
 774 static bool
 775 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 776 {
 777         time64_t deadline;
 778
 779         deadline = lpni->lpni_last_alive +
 780                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 781
 782         /*
 783          * assume peer_ni is alive as long as we're within the configured
 784          * peer timeout
 785          */
 786         if (deadline > now)
 787                 return false;
 788
 789         return true;
 790 }
 791
 792 /* NB: returns 1 when alive, 0 when dead, negative when error;
 793  *     may drop the lnet_net_lock */
 794 static int
 795 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 796                        struct lnet_msg *msg)
 797 {
 798         time64_t now = ktime_get_seconds();
 799
 800         if (!lnet_peer_aliveness_enabled(lpni))
 801                 return -ENODEV;
 802
 803         /*
 804          * If we're resending a message, let's attempt to send it even if
 805          * the peer is down to fulfill our resend quota on the message
 806          */
 807         if (msg->msg_retry_count > 0)
 808                 return 1;
 809
 810         /* try and send recovery messages irregardless */
 811         if (msg->msg_recovery)
 812                 return 1;
 813
 814         /* always send any responses */
 815         if (lnet_msg_is_response(msg))
 816                 return 1;
 817
 818         if (!lnet_is_peer_deadline_passed(lpni, now))
 819                 return true;
 820
 821         return lnet_is_peer_ni_alive(lpni);
 822 }
 823
 824 /**
 825  * \param msg The message to be sent.
 826  * \param do_send True if lnet_ni_send() should be called in this function.
 827  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 828  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 829  *
 830  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 831  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 832  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 833  * \retval -ECANCELED If the MD of the message has been unlinked.
 834  */
 835 static int
 836 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 837 {
 838         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 839         struct lnet_ni          *ni = msg->msg_txni;
 840         int                     cpt = msg->msg_tx_cpt;
 841         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 842
 843         /* non-lnet_send() callers have checked before */
 844         LASSERT(!do_send || msg->msg_tx_delayed);
 845         LASSERT(!msg->msg_receiving);
 846         LASSERT(msg->msg_tx_committed);
 847
 848         /* can't get here if we're sending to the loopback interface */
 849         if (the_lnet.ln_loni)
 850                 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
 851
 852         /* NB 'lp' is always the next hop */
 853         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 854             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 855                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 856                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 857                         msg->msg_len;
 858                 lnet_net_unlock(cpt);
 859                 if (msg->msg_txpeer)
 860                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 861                                         msg->msg_type,
 862                                         LNET_STATS_TYPE_DROP);
 863                 if (msg->msg_txni)
 864                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 865                                         msg->msg_type,
 866                                         LNET_STATS_TYPE_DROP);
 867
 868                 CNETERR("Dropping message for %s: peer not alive\n",
 869                         libcfs_id2str(msg->msg_target));
 870                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 871                 if (do_send)
 872                         lnet_finalize(msg, -EHOSTUNREACH);
 873
 874                 lnet_net_lock(cpt);
 875                 return -EHOSTUNREACH;
 876         }
 877
 878         if (msg->msg_md != NULL &&
 879             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 880                 lnet_net_unlock(cpt);
 881
 882                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 883                         "called on the MD/ME.\n",
 884                         libcfs_id2str(msg->msg_target));
 885                 if (do_send) {
 886                         msg->msg_no_resend = true;
 887                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 888                                msg, libcfs_id2str(msg->msg_target));
 889                         lnet_finalize(msg, -ECANCELED);
 890                 }
 891
 892                 lnet_net_lock(cpt);
 893                 return -ECANCELED;
 894         }
 895
 896         if (!msg->msg_peertxcredit) {
 897                 spin_lock(&lp->lpni_lock);
 898                 LASSERT((lp->lpni_txcredits < 0) ==
 899                         !list_empty(&lp->lpni_txq));
 900
 901                 msg->msg_peertxcredit = 1;
 902                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 903                 lp->lpni_txcredits--;
 904
 905                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 906                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 907
 908                 if (lp->lpni_txcredits < 0) {
 909                         msg->msg_tx_delayed = 1;
 910                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 911                         spin_unlock(&lp->lpni_lock);
 912                         return LNET_CREDIT_WAIT;
 913                 }
 914                 spin_unlock(&lp->lpni_lock);
 915         }
 916
 917         if (!msg->msg_txcredit) {
 918                 LASSERT((tq->tq_credits < 0) ==
 919                         !list_empty(&tq->tq_delayed));
 920
 921                 msg->msg_txcredit = 1;
 922                 tq->tq_credits--;
 923                 atomic_dec(&ni->ni_tx_credits);
 924
 925                 if (tq->tq_credits < tq->tq_credits_min)
 926                         tq->tq_credits_min = tq->tq_credits;
 927
 928                 if (tq->tq_credits < 0) {
 929                         msg->msg_tx_delayed = 1;
 930                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933         }
 934
 935         if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
 936             lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
 937                 msg->msg_tx_delayed = 1;
 938                 return LNET_CREDIT_WAIT;
 939         }
 940
 941         /* unset the tx_delay flag as we're going to send it now */
 942         msg->msg_tx_delayed = 0;
 943
 944         if (do_send) {
 945                 lnet_net_unlock(cpt);
 946                 lnet_ni_send(ni, msg);
 947                 lnet_net_lock(cpt);
 948         }
 949         return LNET_CREDIT_OK;
 950 }
 951
 952
 953 static struct lnet_rtrbufpool *
 954 lnet_msg2bufpool(struct lnet_msg *msg)
 955 {
 956         struct lnet_rtrbufpool  *rbp;
 957         int                     cpt;
 958
 959         LASSERT(msg->msg_rx_committed);
 960
 961         cpt = msg->msg_rx_cpt;
 962         rbp = &the_lnet.ln_rtrpools[cpt][0];
 963
 964         LASSERT(msg->msg_len <= LNET_MTU);
 965         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 966                 rbp++;
 967                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 968         }
 969
 970         return rbp;
 971 }
 972
 973 static int
 974 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 975 {
 976         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 977          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 978          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 979          * received or OK to receive */
 980         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 981         struct lnet_peer *lp;
 982         struct lnet_rtrbufpool *rbp;
 983         struct lnet_rtrbuf *rb;
 984
 985         LASSERT(msg->msg_kiov == NULL);
 986         LASSERT(msg->msg_niov == 0);
 987         LASSERT(msg->msg_routing);
 988         LASSERT(msg->msg_receiving);
 989         LASSERT(!msg->msg_sending);
 990         LASSERT(lpni->lpni_peer_net);
 991         LASSERT(lpni->lpni_peer_net->lpn_peer);
 992
 993         lp = lpni->lpni_peer_net->lpn_peer;
 994
 995         /* non-lnet_parse callers only receive delayed messages */
 996         LASSERT(!do_recv || msg->msg_rx_delayed);
 997
 998         if (!msg->msg_peerrtrcredit) {
 999                 /* lpni_lock protects the credit manipulation */
1000                 spin_lock(&lpni->lpni_lock);
1001
1002                 msg->msg_peerrtrcredit = 1;
1003                 lpni->lpni_rtrcredits--;
1004                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1005                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1006
1007                 if (lpni->lpni_rtrcredits < 0) {
1008                         spin_unlock(&lpni->lpni_lock);
1009                         /* must have checked eager_recv before here */
1010                         LASSERT(msg->msg_rx_ready_delay);
1011                         msg->msg_rx_delayed = 1;
1012                         /* lp_lock protects the lp_rtrq */
1013                         spin_lock(&lp->lp_lock);
1014                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1015                         spin_unlock(&lp->lp_lock);
1016                         return LNET_CREDIT_WAIT;
1017                 }
1018                 spin_unlock(&lpni->lpni_lock);
1019         }
1020
1021         rbp = lnet_msg2bufpool(msg);
1022
1023         if (!msg->msg_rtrcredit) {
1024                 msg->msg_rtrcredit = 1;
1025                 rbp->rbp_credits--;
1026                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1027                         rbp->rbp_mincredits = rbp->rbp_credits;
1028
1029                 if (rbp->rbp_credits < 0) {
1030                         /* must have checked eager_recv before here */
1031                         LASSERT(msg->msg_rx_ready_delay);
1032                         msg->msg_rx_delayed = 1;
1033                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1034                         return LNET_CREDIT_WAIT;
1035                 }
1036         }
1037
1038         LASSERT(!list_empty(&rbp->rbp_bufs));
1039         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1040         list_del(&rb->rb_list);
1041
1042         msg->msg_niov = rbp->rbp_npages;
1043         msg->msg_kiov = &rb->rb_kiov[0];
1044
1045         /* unset the msg-rx_delayed flag since we're receiving the message */
1046         msg->msg_rx_delayed = 0;
1047
1048         if (do_recv) {
1049                 int cpt = msg->msg_rx_cpt;
1050
1051                 lnet_net_unlock(cpt);
1052                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1053                              0, msg->msg_len, msg->msg_len);
1054                 lnet_net_lock(cpt);
1055         }
1056         return LNET_CREDIT_OK;
1057 }
1058
1059 void
1060 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1061 {
1062         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1063         struct lnet_ni          *txni = msg->msg_txni;
1064         struct lnet_msg         *msg2;
1065
1066         if (msg->msg_txcredit) {
1067                 struct lnet_ni       *ni = msg->msg_txni;
1068                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1069
1070                 /* give back NI txcredits */
1071                 msg->msg_txcredit = 0;
1072
1073                 LASSERT((tq->tq_credits < 0) ==
1074                         !list_empty(&tq->tq_delayed));
1075
1076                 tq->tq_credits++;
1077                 atomic_inc(&ni->ni_tx_credits);
1078                 if (tq->tq_credits <= 0) {
1079                         msg2 = list_entry(tq->tq_delayed.next,
1080                                           struct lnet_msg, msg_list);
1081                         list_del(&msg2->msg_list);
1082
1083                         LASSERT(msg2->msg_txni == ni);
1084                         LASSERT(msg2->msg_tx_delayed);
1085                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1086
1087                         (void) lnet_post_send_locked(msg2, 1);
1088                 }
1089         }
1090
1091         if (msg->msg_peertxcredit) {
1092                 /* give back peer txcredits */
1093                 msg->msg_peertxcredit = 0;
1094
1095                 spin_lock(&txpeer->lpni_lock);
1096                 LASSERT((txpeer->lpni_txcredits < 0) ==
1097                         !list_empty(&txpeer->lpni_txq));
1098
1099                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1100                 LASSERT(txpeer->lpni_txqnob >= 0);
1101
1102                 txpeer->lpni_txcredits++;
1103                 if (txpeer->lpni_txcredits <= 0) {
1104                         int msg2_cpt;
1105
1106                         msg2 = list_entry(txpeer->lpni_txq.next,
1107                                               struct lnet_msg, msg_list);
1108                         list_del(&msg2->msg_list);
1109                         spin_unlock(&txpeer->lpni_lock);
1110
1111                         LASSERT(msg2->msg_txpeer == txpeer);
1112                         LASSERT(msg2->msg_tx_delayed);
1113
1114                         msg2_cpt = msg2->msg_tx_cpt;
1115
1116                         /*
1117                          * The msg_cpt can be different from the msg2_cpt
1118                          * so we need to make sure we lock the correct cpt
1119                          * for msg2.
1120                          * Once we call lnet_post_send_locked() it is no
1121                          * longer safe to access msg2, since it could've
1122                          * been freed by lnet_finalize(), but we still
1123                          * need to relock the correct cpt, so we cache the
1124                          * msg2_cpt for the purpose of the check that
1125                          * follows the call to lnet_pose_send_locked().
1126                          */
1127                         if (msg2_cpt != msg->msg_tx_cpt) {
1128                                 lnet_net_unlock(msg->msg_tx_cpt);
1129                                 lnet_net_lock(msg2_cpt);
1130                         }
1131                         (void) lnet_post_send_locked(msg2, 1);
1132                         if (msg2_cpt != msg->msg_tx_cpt) {
1133                                 lnet_net_unlock(msg2_cpt);
1134                                 lnet_net_lock(msg->msg_tx_cpt);
1135                         }
1136                 } else {
1137                         spin_unlock(&txpeer->lpni_lock);
1138                 }
1139         }
1140
1141         if (txni != NULL) {
1142                 msg->msg_txni = NULL;
1143                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1144         }
1145
1146         if (txpeer != NULL) {
1147                 msg->msg_txpeer = NULL;
1148                 lnet_peer_ni_decref_locked(txpeer);
1149         }
1150 }
1151
1152 void
1153 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1154 {
1155         struct lnet_msg *msg;
1156
1157         if (list_empty(&rbp->rbp_msgs))
1158                 return;
1159         msg = list_entry(rbp->rbp_msgs.next,
1160                          struct lnet_msg, msg_list);
1161         list_del(&msg->msg_list);
1162
1163         (void)lnet_post_routed_recv_locked(msg, 1);
1164 }
1165
1166 void
1167 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1168 {
1169         struct lnet_msg *msg;
1170         struct lnet_msg *tmp;
1171
1172         lnet_net_unlock(cpt);
1173
1174         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1175                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1176                              0, 0, 0, msg->msg_hdr.payload_length);
1177                 list_del_init(&msg->msg_list);
1178                 msg->msg_no_resend = true;
1179                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1180                 lnet_finalize(msg, -ECANCELED);
1181         }
1182
1183         lnet_net_lock(cpt);
1184 }
1185
1186 void
1187 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1188 {
1189         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1190         struct lnet_peer *lp;
1191         struct lnet_ni *rxni = msg->msg_rxni;
1192         struct lnet_msg *msg2;
1193
1194         if (msg->msg_rtrcredit) {
1195                 /* give back global router credits */
1196                 struct lnet_rtrbuf *rb;
1197                 struct lnet_rtrbufpool *rbp;
1198
1199                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1200                  * there until it gets one allocated, or aborts the wait
1201                  * itself */
1202                 LASSERT(msg->msg_kiov != NULL);
1203
1204                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1205                 rbp = rb->rb_pool;
1206
1207                 msg->msg_kiov = NULL;
1208                 msg->msg_rtrcredit = 0;
1209
1210                 LASSERT(rbp == lnet_msg2bufpool(msg));
1211
1212                 LASSERT((rbp->rbp_credits > 0) ==
1213                         !list_empty(&rbp->rbp_bufs));
1214
1215                 /* If routing is now turned off, we just drop this buffer and
1216                  * don't bother trying to return credits.  */
1217                 if (!the_lnet.ln_routing) {
1218                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1219                         goto routing_off;
1220                 }
1221
1222                 /* It is possible that a user has lowered the desired number of
1223                  * buffers in this pool.  Make sure we never put back
1224                  * more buffers than the stated number. */
1225                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1226                         /* Discard this buffer so we don't have too
1227                          * many. */
1228                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1229                         rbp->rbp_nbuffers--;
1230                 } else {
1231                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1232                         rbp->rbp_credits++;
1233                         if (rbp->rbp_credits <= 0)
1234                                 lnet_schedule_blocked_locked(rbp);
1235                 }
1236         }
1237
1238 routing_off:
1239         if (msg->msg_peerrtrcredit) {
1240                 LASSERT(rxpeerni);
1241                 LASSERT(rxpeerni->lpni_peer_net);
1242                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1243
1244                 /* give back peer router credits */
1245                 msg->msg_peerrtrcredit = 0;
1246
1247                 spin_lock(&rxpeerni->lpni_lock);
1248                 rxpeerni->lpni_rtrcredits++;
1249                 spin_unlock(&rxpeerni->lpni_lock);
1250
1251                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1252                 spin_lock(&lp->lp_lock);
1253
1254                 /* drop all messages which are queued to be routed on that
1255                  * peer. */
1256                 if (!the_lnet.ln_routing) {
1257                         LIST_HEAD(drop);
1258                         list_splice_init(&lp->lp_rtrq, &drop);
1259                         spin_unlock(&lp->lp_lock);
1260                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1261                 } else if (!list_empty(&lp->lp_rtrq)) {
1262                         int msg2_cpt;
1263
1264                         msg2 = list_entry(lp->lp_rtrq.next,
1265                                           struct lnet_msg, msg_list);
1266                         list_del(&msg2->msg_list);
1267                         msg2_cpt = msg2->msg_rx_cpt;
1268                         spin_unlock(&lp->lp_lock);
1269                         /*
1270                          * messages on the lp_rtrq can be from any NID in
1271                          * the peer, which means they might have different
1272                          * cpts. We need to make sure we lock the right
1273                          * one.
1274                          */
1275                         if (msg2_cpt != msg->msg_rx_cpt) {
1276                                 lnet_net_unlock(msg->msg_rx_cpt);
1277                                 lnet_net_lock(msg2_cpt);
1278                         }
1279                         (void) lnet_post_routed_recv_locked(msg2, 1);
1280                         if (msg2_cpt != msg->msg_rx_cpt) {
1281                                 lnet_net_unlock(msg2_cpt);
1282                                 lnet_net_lock(msg->msg_rx_cpt);
1283                         }
1284                 } else {
1285                         spin_unlock(&lp->lp_lock);
1286                 }
1287         }
1288         if (rxni != NULL) {
1289                 msg->msg_rxni = NULL;
1290                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1291         }
1292         if (rxpeerni != NULL) {
1293                 msg->msg_rxpeer = NULL;
1294                 lnet_peer_ni_decref_locked(rxpeerni);
1295         }
1296 }
1297
1298 static struct lnet_peer_ni *
1299 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1300                     struct lnet_peer *peer,
1301                     struct lnet_peer_ni *best_lpni,
1302                     struct lnet_peer_net *peer_net)
1303 {
1304         /*
1305          * Look at the peer NIs for the destination peer that connect
1306          * to the chosen net. If a peer_ni is preferred when using the
1307          * best_ni to communicate, we use that one. If there is no
1308          * preferred peer_ni, or there are multiple preferred peer_ni,
1309          * the available transmit credits are used. If the transmit
1310          * credits are equal, we round-robin over the peer_ni.
1311          */
1312         struct lnet_peer_ni *lpni = NULL;
1313         int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1314                 INT_MIN;
1315         int best_lpni_healthv = (best_lpni) ?
1316                 atomic_read(&best_lpni->lpni_healthv) : 0;
1317         bool best_lpni_is_preferred = false;
1318         bool lpni_is_preferred;
1319         int lpni_healthv;
1320         __u32 lpni_sel_prio;
1321         __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1322
1323         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1324                 /*
1325                  * if the best_ni we've chosen aleady has this lpni
1326                  * preferred, then let's use it
1327                  */
1328                 if (best_ni) {
1329                         /* FIXME need to handle large-addr nid */
1330                         lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1331                                 lpni, lnet_nid_to_nid4(&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         lnet_nid_t 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 = lnet_nid_to_nid4(&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_nid2str(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_nid2str(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_nid2str(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 = lnet_nid_to_nid4(&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         /* FIXME handle large-addr nids */
1906         msg->msg_target.nid = lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid);
1907
1908         /*
1909          * lnet_msg_commit assigns the correct cpt to the message, which
1910          * is used to decrement the correct refcount on the ni when it's
1911          * time to return the credits
1912          */
1913         lnet_msg_commit(msg, sd->sd_cpt);
1914
1915         /*
1916          * If we are routing the message then we keep the src_nid that was
1917          * set by the originator. If we are not routing then we are the
1918          * originator and set it here.
1919          */
1920         if (!msg->msg_routing)
1921                 msg->msg_hdr.src_nid =
1922                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txni->ni_nid));
1923
1924         if (routing) {
1925                 msg->msg_target_is_router = 1;
1926                 msg->msg_target.pid = LNET_PID_LUSTRE;
1927                 /*
1928                  * since we're routing we want to ensure that the
1929                  * msg_hdr.dest_nid is set to the final destination. When
1930                  * the router receives this message it knows how to route
1931                  * it.
1932                  *
1933                  * final_dst_lpni is set at the beginning of the
1934                  * lnet_select_pathway() function and is never changed.
1935                  * It's safe to use it here.
1936                  */
1937                 /* FIXME handle large-addr nid */
1938                 msg->msg_hdr.dest_nid =
1939                         cpu_to_le64(lnet_nid_to_nid4(&final_dst_lpni->lpni_nid));
1940         } else {
1941                 /*
1942                  * if we're not routing set the dest_nid to the best peer
1943                  * ni NID that we picked earlier in the algorithm.
1944                  */
1945                 msg->msg_hdr.dest_nid =
1946                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid));
1947         }
1948
1949         /*
1950          * if we have response tracker block update it with the next hop
1951          * nid
1952          */
1953         if (msg->msg_md) {
1954                 rspt = msg->msg_md->md_rspt_ptr;
1955                 if (rspt) {
1956                         rspt->rspt_next_hop_nid =
1957                                 msg->msg_txpeer->lpni_nid;
1958                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1959                                libcfs_nidstr(&rspt->rspt_next_hop_nid));
1960                 }
1961         }
1962
1963         rc = lnet_post_send_locked(msg, 0);
1964
1965         if (!rc)
1966                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1967                        libcfs_nid2str(msg->msg_hdr.src_nid),
1968                        libcfs_nidstr(&msg->msg_txni->ni_nid),
1969                        libcfs_nid2str(sd->sd_src_nid),
1970                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1971                        libcfs_nid2str(sd->sd_dst_nid),
1972                        libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1973                        libcfs_nid2str(sd->sd_rtr_nid),
1974                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1975
1976         return rc;
1977 }
1978
1979 static inline void
1980 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1981                          struct lnet_msg *msg)
1982 {
1983         if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1984             !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1985                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1986                        libcfs_nidstr(&lni->ni_nid),
1987                        libcfs_nidstr(&lpni->lpni_nid));
1988                 lnet_peer_ni_set_non_mr_pref_nid(
1989                         lpni, lnet_nid_to_nid4(&lni->ni_nid));
1990         }
1991 }
1992
1993 /*
1994  * Source Specified
1995  * Local Destination
1996  * non-mr peer
1997  *
1998  * use the source and destination NIDs as the pathway
1999  */
2000 static int
2001 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2002 {
2003         /* the destination lpni is set before we get here. */
2004
2005         /* find local NI */
2006         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2007         if (!sd->sd_best_ni) {
2008                 CERROR("Can't send to %s: src %s is not a "
2009                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2010                                 libcfs_nid2str(sd->sd_src_nid));
2011                 return -EINVAL;
2012         }
2013
2014         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2015
2016         return lnet_handle_send(sd);
2017 }
2018
2019 /*
2020  * Source Specified
2021  * Local Destination
2022  * MR Peer
2023  *
2024  * Don't run the selection algorithm on the peer NIs. By specifying the
2025  * local NID, we're also saying that we should always use the destination NID
2026  * provided. This handles the case where we should be using the same
2027  * destination NID for the all the messages which belong to the same RPC
2028  * request.
2029  */
2030 static int
2031 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2032 {
2033         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2034         if (!sd->sd_best_ni) {
2035                 CERROR("Can't send to %s: src %s is not a "
2036                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2037                                 libcfs_nid2str(sd->sd_src_nid));
2038                 return -EINVAL;
2039         }
2040
2041         if (sd->sd_best_lpni &&
2042             nid_same(&sd->sd_best_lpni->lpni_nid,
2043                       &the_lnet.ln_loni->ni_nid))
2044                 return lnet_handle_lo_send(sd);
2045         else if (sd->sd_best_lpni)
2046                 return lnet_handle_send(sd);
2047
2048         CERROR("can't send to %s. no NI on %s\n",
2049                libcfs_nid2str(sd->sd_dst_nid),
2050                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2051
2052         return -EHOSTUNREACH;
2053 }
2054
2055 struct lnet_ni *
2056 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2057                               struct lnet_peer *peer,
2058                               struct lnet_peer_net *peer_net,
2059                               struct lnet_msg *msg,
2060                               int cpt)
2061 {
2062         struct lnet_net *local_net;
2063         struct lnet_ni *best_ni;
2064
2065         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2066         if (!local_net)
2067                 return NULL;
2068
2069         /*
2070          * Iterate through the NIs in this local Net and select
2071          * the NI to send from. The selection is determined by
2072          * these 3 criterion in the following priority:
2073          *      1. NUMA
2074          *      2. NI available credits
2075          *      3. Round Robin
2076          */
2077         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2078                                    peer, peer_net, msg, cpt);
2079
2080         return best_ni;
2081 }
2082
2083 static int
2084 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2085                              int cpt)
2086 {
2087         struct lnet_peer *peer;
2088         struct lnet_peer_ni *new_lpni;
2089         int rc;
2090
2091         lnet_peer_ni_addref_locked(lpni);
2092
2093         peer = lpni->lpni_peer_net->lpn_peer;
2094
2095         if (lnet_peer_gw_discovery(peer)) {
2096                 lnet_peer_ni_decref_locked(lpni);
2097                 return 0;
2098         }
2099
2100         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2101                 lnet_peer_ni_decref_locked(lpni);
2102                 return 0;
2103         }
2104
2105         rc = lnet_discover_peer_locked(lpni, cpt, false);
2106         if (rc) {
2107                 lnet_peer_ni_decref_locked(lpni);
2108                 return rc;
2109         }
2110
2111         new_lpni = lnet_find_peer_ni_locked(lnet_nid_to_nid4(&lpni->lpni_nid));
2112         if (!new_lpni) {
2113                 lnet_peer_ni_decref_locked(lpni);
2114                 return -ENOENT;
2115         }
2116
2117         peer = new_lpni->lpni_peer_net->lpn_peer;
2118         spin_lock(&peer->lp_lock);
2119         if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2120                 /* The peer NI did not change and the peer is up to date.
2121                  * Nothing more to do.
2122                  */
2123                 spin_unlock(&peer->lp_lock);
2124                 lnet_peer_ni_decref_locked(lpni);
2125                 lnet_peer_ni_decref_locked(new_lpni);
2126                 return 0;
2127         }
2128         spin_unlock(&peer->lp_lock);
2129
2130         /* Either the peer NI changed during discovery, or the peer isn't up
2131          * to date. In both cases we want to queue the message on the
2132          * (possibly new) peer's pending queue and queue the peer for discovery
2133          */
2134         msg->msg_sending = 0;
2135         msg->msg_txpeer = NULL;
2136         lnet_net_unlock(cpt);
2137         lnet_peer_queue_message(peer, msg);
2138         lnet_net_lock(cpt);
2139
2140         lnet_peer_ni_decref_locked(lpni);
2141         lnet_peer_ni_decref_locked(new_lpni);
2142
2143         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2144                msg, libcfs_nidstr(&peer->lp_primary_nid));
2145
2146         return LNET_DC_WAIT;
2147 }
2148
2149 static int
2150 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2151                              lnet_nid_t dst_nid,
2152                              struct lnet_peer_ni **gw_lpni,
2153                              struct lnet_peer **gw_peer)
2154 {
2155         int rc;
2156         struct lnet_peer *gw;
2157         struct lnet_peer *lp;
2158         struct lnet_peer_net *lpn;
2159         struct lnet_peer_net *best_lpn = NULL;
2160         struct lnet_remotenet *rnet, *best_rnet = NULL;
2161         struct lnet_route *best_route = NULL;
2162         struct lnet_route *last_route = NULL;
2163         struct lnet_peer_ni *lpni = NULL;
2164         struct lnet_peer_ni *gwni = NULL;
2165         bool route_found = false;
2166         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2167                 (sd->sd_best_ni != NULL)
2168                 ? lnet_nid_to_nid4(&sd->sd_best_ni->ni_nid)
2169                 : LNET_NID_ANY;
2170         int best_lpn_healthv = 0;
2171         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2172
2173         CDEBUG(D_NET, "using src nid %s for route restriction\n",
2174                libcfs_nid2str(src_nid));
2175
2176         /* If a router nid was specified then we are replying to a GET or
2177          * sending an ACK. In this case we use the gateway associated with the
2178          * specified router nid.
2179          */
2180         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2181                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2182                 if (gwni) {
2183                         gw = gwni->lpni_peer_net->lpn_peer;
2184                         lnet_peer_ni_decref_locked(gwni);
2185                         if (gw->lp_rtr_refcount)
2186                                 route_found = true;
2187                 } else {
2188                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2189                                libcfs_nid2str(sd->sd_rtr_nid));
2190                 }
2191         }
2192
2193         if (!route_found) {
2194                 if (sd->sd_msg->msg_routing) {
2195                         /* If I'm routing this message then I need to find the
2196                          * next hop based on the destination NID
2197                          */
2198                         best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2199                         if (!best_rnet) {
2200                                 CERROR("Unable to route message to %s - Route table may be misconfigured\n",
2201                                        libcfs_nid2str(sd->sd_dst_nid));
2202                                 return -EHOSTUNREACH;
2203                         }
2204                 } else {
2205                         /* we've already looked up the initial lpni using
2206                          * dst_nid
2207                          */
2208                         lpni = sd->sd_best_lpni;
2209                         /* the peer tree must be in existence */
2210                         LASSERT(lpni && lpni->lpni_peer_net &&
2211                                 lpni->lpni_peer_net->lpn_peer);
2212                         lp = lpni->lpni_peer_net->lpn_peer;
2213
2214                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2215                                 /* is this remote network reachable?  */
2216                                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2217                                 if (!rnet)
2218                                         continue;
2219
2220                                 if (!best_lpn) {
2221                                         best_lpn = lpn;
2222                                         best_rnet = rnet;
2223                                 }
2224
2225                                 /* select the preferred peer net */
2226                                 if (best_lpn_healthv > lpn->lpn_healthv)
2227                                         continue;
2228                                 else if (best_lpn_healthv < lpn->lpn_healthv)
2229                                         goto use_lpn;
2230
2231                                 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2232                                         continue;
2233                                 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2234                                         goto use_lpn;
2235
2236                                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2237                                         continue;
2238 use_lpn:
2239                                 best_lpn_healthv = lpn->lpn_healthv;
2240                                 best_lpn_sel_prio = lpn->lpn_sel_priority;
2241                                 best_lpn = lpn;
2242                                 best_rnet = rnet;
2243                         }
2244
2245                         if (!best_lpn) {
2246                                 CERROR("peer %s has no available nets\n",
2247                                        libcfs_nid2str(sd->sd_dst_nid));
2248                                 return -EHOSTUNREACH;
2249                         }
2250
2251                         sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2252                                                                sd->sd_dst_nid,
2253                                                                lp,
2254                                                                best_lpn->lpn_net_id);
2255                         if (!sd->sd_best_lpni) {
2256                                 CERROR("peer %s is unreachable\n",
2257                                        libcfs_nid2str(sd->sd_dst_nid));
2258                                 return -EHOSTUNREACH;
2259                         }
2260
2261                         /* We're attempting to round robin over the remote peer
2262                          * NI's so update the final destination we selected
2263                          */
2264                         sd->sd_final_dst_lpni = sd->sd_best_lpni;
2265
2266                         /* Increment the sequence number of the remote lpni so
2267                          * we can round robin over the different interfaces of
2268                          * the remote lpni
2269                          */
2270                         sd->sd_best_lpni->lpni_seq++;
2271                 }
2272
2273                 /*
2274                  * find the best route. Restrict the selection on the net of the
2275                  * local NI if we've already picked the local NI to send from.
2276                  * Otherwise, let's pick any route we can find and then find
2277                  * a local NI we can reach the route's gateway on. Any route we
2278                  * select will be reachable by virtue of the restriction we have
2279                  * when adding a route.
2280                  */
2281                 best_route = lnet_find_route_locked(best_rnet,
2282                                                     LNET_NIDNET(src_nid),
2283                                                     sd->sd_best_lpni,
2284                                                     &last_route, &gwni);
2285
2286                 if (!best_route) {
2287                         CERROR("no route to %s from %s\n",
2288                                libcfs_nid2str(dst_nid),
2289                                libcfs_nid2str(src_nid));
2290                         return -EHOSTUNREACH;
2291                 }
2292
2293                 if (!gwni) {
2294                         CERROR("Internal Error. Route expected to %s from %s\n",
2295                                libcfs_nid2str(dst_nid),
2296                                libcfs_nid2str(src_nid));
2297                         return -EFAULT;
2298                 }
2299
2300                 gw = best_route->lr_gateway;
2301                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2302         }
2303
2304         /*
2305          * Discover this gateway if it hasn't already been discovered.
2306          * This means we might delay the message until discovery has
2307          * completed
2308          */
2309         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2310         if (rc)
2311                 return rc;
2312
2313         if (!sd->sd_best_ni) {
2314                 lpn = gwni->lpni_peer_net;
2315                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2316                                                                sd->sd_msg,
2317                                                                sd->sd_md_cpt);
2318                 if (!sd->sd_best_ni) {
2319                         CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2320                                libcfs_net2str(lpn->lpn_net_id),
2321                                libcfs_nid2str(sd->sd_src_nid));
2322                         return -EFAULT;
2323                 }
2324         }
2325
2326         *gw_lpni = gwni;
2327         *gw_peer = gw;
2328
2329         /*
2330          * increment the sequence numbers since now we're sure we're
2331          * going to use this path
2332          */
2333         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2334                 LASSERT(best_route && last_route);
2335                 best_route->lr_seq = last_route->lr_seq + 1;
2336                 if (best_lpn)
2337                         best_lpn->lpn_seq++;
2338         }
2339
2340         return 0;
2341 }
2342
2343 /*
2344  * Handle two cases:
2345  *
2346  * Case 1:
2347  *  Source specified
2348  *  Remote destination
2349  *  Non-MR destination
2350  *
2351  * Case 2:
2352  *  Source specified
2353  *  Remote destination
2354  *  MR destination
2355  *
2356  * The handling of these two cases is similar. Even though the destination
2357  * can be MR or non-MR, we'll deal directly with the router.
2358  */
2359 static int
2360 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2361 {
2362         int rc;
2363         struct lnet_peer_ni *gw_lpni = NULL;
2364         struct lnet_peer *gw_peer = NULL;
2365
2366         /* find local NI */
2367         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2368         if (!sd->sd_best_ni) {
2369                 CERROR("Can't send to %s: src %s is not a "
2370                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2371                                 libcfs_nid2str(sd->sd_src_nid));
2372                 return -EINVAL;
2373         }
2374
2375         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2376                                      &gw_peer);
2377         if (rc)
2378                 return rc;
2379
2380         if (sd->sd_send_case & NMR_DST)
2381                 /*
2382                  * since the final destination is non-MR let's set its preferred
2383                  * NID before we send
2384                  */
2385                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2386                                          sd->sd_msg);
2387
2388         /*
2389          * We're going to send to the gw found so let's set its
2390          * info
2391          */
2392         sd->sd_peer = gw_peer;
2393         sd->sd_best_lpni = gw_lpni;
2394
2395         return lnet_handle_send(sd);
2396 }
2397
2398 struct lnet_ni *
2399 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2400                                struct lnet_msg *msg, bool discovery)
2401 {
2402         struct lnet_peer_net *lpn = NULL;
2403         struct lnet_peer_net *best_lpn = NULL;
2404         struct lnet_net *net = NULL;
2405         struct lnet_net *best_net = NULL;
2406         struct lnet_ni *best_ni = NULL;
2407         int best_lpn_healthv = 0;
2408         int best_net_healthv = 0;
2409         int net_healthv;
2410         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2411         __u32 lpn_sel_prio;
2412         __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2413         __u32 net_sel_prio;
2414         bool exit = false;
2415
2416         /*
2417          * The peer can have multiple interfaces, some of them can be on
2418          * the local network and others on a routed network. We should
2419          * prefer the local network. However if the local network is not
2420          * available then we need to try the routed network
2421          */
2422
2423         /* go through all the peer nets and find the best_ni */
2424         list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2425                 /*
2426                  * The peer's list of nets can contain non-local nets. We
2427                  * want to only examine the local ones.
2428                  */
2429                 net = lnet_get_net_locked(lpn->lpn_net_id);
2430                 if (!net)
2431                         continue;
2432
2433                 lpn_sel_prio = lpn->lpn_sel_priority;
2434                 net_healthv = lnet_get_net_healthv_locked(net);
2435                 net_sel_prio = net->net_sel_priority;
2436
2437                 /*
2438                  * if this is a discovery message and lp_disc_net_id is
2439                  * specified then use that net to send the discovery on.
2440                  */
2441                 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2442                     discovery) {
2443                         exit = true;
2444                         goto select_lpn;
2445                 }
2446
2447                 if (!best_lpn)
2448                         goto select_lpn;
2449
2450                 /* always select the lpn with the best health */
2451                 if (best_lpn_healthv > lpn->lpn_healthv)
2452                         continue;
2453                 else if (best_lpn_healthv < lpn->lpn_healthv)
2454                         goto select_lpn;
2455
2456                 /* select the preferred peer and local nets */
2457                 if (best_lpn_sel_prio < lpn_sel_prio)
2458                         continue;
2459                 else if (best_lpn_sel_prio > lpn_sel_prio)
2460                         goto select_lpn;
2461
2462                 if (best_net_healthv > net_healthv)
2463                         continue;
2464                 else if (best_net_healthv < net_healthv)
2465                         goto select_lpn;
2466
2467                 if (best_net_sel_prio < net_sel_prio)
2468                         continue;
2469                 else if (best_net_sel_prio > net_sel_prio)
2470                         goto select_lpn;
2471
2472                 if (best_lpn->lpn_seq < lpn->lpn_seq)
2473                         continue;
2474                 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2475                         goto select_lpn;
2476
2477                 /* round robin over the local networks */
2478                 if (best_net->net_seq <= net->net_seq)
2479                         continue;
2480
2481 select_lpn:
2482                 best_net_healthv = net_healthv;
2483                 best_net_sel_prio = net_sel_prio;
2484                 best_lpn_healthv = lpn->lpn_healthv;
2485                 best_lpn_sel_prio = lpn_sel_prio;
2486                 best_lpn = lpn;
2487                 best_net = net;
2488
2489                 if (exit)
2490                         break;
2491         }
2492
2493         if (best_lpn) {
2494                 /* Select the best NI on the same net as best_lpn chosen
2495                  * above
2496                  */
2497                 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2498                                                         msg, md_cpt);
2499         }
2500
2501         return best_ni;
2502 }
2503
2504 static struct lnet_ni *
2505 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2506 {
2507         struct lnet_ni *best_ni = NULL;
2508         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2509         struct lnet_peer_ni *lpni_entry;
2510
2511         /*
2512          * We must use a consistent source address when sending to a
2513          * non-MR peer. However, a non-MR peer can have multiple NIDs
2514          * on multiple networks, and we may even need to talk to this
2515          * peer on multiple networks -- certain types of
2516          * load-balancing configuration do this.
2517          *
2518          * So we need to pick the NI the peer prefers for this
2519          * particular network.
2520          */
2521         LASSERT(peer_net);
2522         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2523                             lpni_peer_nis) {
2524                 if (lpni_entry->lpni_pref_nnids == 0)
2525                         continue;
2526                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2527                 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2528                 break;
2529         }
2530
2531         return best_ni;
2532 }
2533
2534 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2535 static int
2536 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2537 {
2538         struct lnet_ni *best_ni = NULL;
2539         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2540
2541         /*
2542          * We must use a consistent source address when sending to a
2543          * non-MR peer. However, a non-MR peer can have multiple NIDs
2544          * on multiple networks, and we may even need to talk to this
2545          * peer on multiple networks -- certain types of
2546          * load-balancing configuration do this.
2547          *
2548          * So we need to pick the NI the peer prefers for this
2549          * particular network.
2550          */
2551
2552         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2553                                                        sd->sd_cpt);
2554
2555         /* if best_ni is still not set just pick one */
2556         if (!best_ni) {
2557                 best_ni =
2558                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2559                                                 sd->sd_best_lpni->lpni_peer_net,
2560                                                 sd->sd_msg,
2561                                                 sd->sd_md_cpt);
2562                 /* If there is no best_ni we don't have a route */
2563                 if (!best_ni) {
2564                         CERROR("no path to %s from net %s\n",
2565                                 libcfs_nidstr(&best_lpni->lpni_nid),
2566                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2567                         return -EHOSTUNREACH;
2568                 }
2569         }
2570
2571         sd->sd_best_ni = best_ni;
2572
2573         /* Set preferred NI if necessary. */
2574         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2575
2576         return 0;
2577 }
2578
2579
2580 /*
2581  * Source not specified
2582  * Local destination
2583  * Non-MR Peer
2584  *
2585  * always use the same source NID for NMR peers
2586  * If we've talked to that peer before then we already have a preferred
2587  * source NI associated with it. Otherwise, we select a preferred local NI
2588  * and store it in the peer
2589  */
2590 static int
2591 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2592 {
2593         int rc = 0;
2594
2595         /* sd->sd_best_lpni is already set to the final destination */
2596
2597         /*
2598          * At this point we should've created the peer ni and peer. If we
2599          * can't find it, then something went wrong. Instead of assert
2600          * output a relevant message and fail the send
2601          */
2602         if (!sd->sd_best_lpni) {
2603                 CERROR("Internal fault. Unable to send msg %s to %s. "
2604                        "NID not known\n",
2605                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2606                        libcfs_nid2str(sd->sd_dst_nid));
2607                 return -EFAULT;
2608         }
2609
2610         if (sd->sd_msg->msg_routing) {
2611                 /* If I'm forwarding this message then I can choose any NI
2612                  * on the destination peer net
2613                  */
2614                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2615                                                                sd->sd_peer,
2616                                                                sd->sd_best_lpni->lpni_peer_net,
2617                                                                sd->sd_msg,
2618                                                                sd->sd_md_cpt);
2619                 if (!sd->sd_best_ni) {
2620                         CERROR("Unable to forward message to %s. No local NI available\n",
2621                                libcfs_nid2str(sd->sd_dst_nid));
2622                         rc = -EHOSTUNREACH;
2623                 }
2624         } else
2625                 rc = lnet_select_preferred_best_ni(sd);
2626
2627         if (!rc)
2628                 rc = lnet_handle_send(sd);
2629
2630         return rc;
2631 }
2632
2633 static int
2634 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2635 {
2636         /*
2637          * NOTE we've already handled the remote peer case. So we only
2638          * need to worry about the local case here.
2639          *
2640          * if we're sending a response, ACK or reply, we need to send it
2641          * to the destination NID given to us. At this point we already
2642          * have the peer_ni we're suppose to send to, so just find the
2643          * best_ni on the peer net and use that. Since we're sending to an
2644          * MR peer then we can just run the selection algorithm on our
2645          * local NIs and pick the best one.
2646          */
2647         if (sd->sd_send_case & SND_RESP) {
2648                 sd->sd_best_ni =
2649                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2650                                                 sd->sd_best_lpni->lpni_peer_net,
2651                                                 sd->sd_msg,
2652                                                 sd->sd_md_cpt);
2653
2654                 if (!sd->sd_best_ni) {
2655                         /*
2656                          * We're not going to deal with not able to send
2657                          * a response to the provided final destination
2658                          */
2659                         CERROR("Can't send response to %s. "
2660                                "No local NI available\n",
2661                                 libcfs_nid2str(sd->sd_dst_nid));
2662                         return -EHOSTUNREACH;
2663                 }
2664
2665                 return lnet_handle_send(sd);
2666         }
2667
2668         /*
2669          * If we get here that means we're sending a fresh request, PUT or
2670          * GET, so we need to run our standard selection algorithm.
2671          * First find the best local interface that's on any of the peer's
2672          * networks.
2673          */
2674         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2675                                         sd->sd_md_cpt,
2676                                         sd->sd_msg,
2677                                         lnet_msg_discovery(sd->sd_msg));
2678         if (sd->sd_best_ni) {
2679                 sd->sd_best_lpni =
2680                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2681                                       sd->sd_peer,
2682                                       sd->sd_best_ni->ni_net->net_id);
2683
2684                 /*
2685                  * if we're successful in selecting a peer_ni on the local
2686                  * network, then send to it. Otherwise fall through and
2687                  * try and see if we can reach it over another routed
2688                  * network
2689                  */
2690                 if (sd->sd_best_lpni &&
2691                     nid_same(&sd->sd_best_lpni->lpni_nid,
2692                              &the_lnet.ln_loni->ni_nid)) {
2693                         /*
2694                          * in case we initially started with a routed
2695                          * destination, let's reset to local
2696                          */
2697                         sd->sd_send_case &= ~REMOTE_DST;
2698                         sd->sd_send_case |= LOCAL_DST;
2699                         return lnet_handle_lo_send(sd);
2700                 } else if (sd->sd_best_lpni) {
2701                         /*
2702                          * in case we initially started with a routed
2703                          * destination, let's reset to local
2704                          */
2705                         sd->sd_send_case &= ~REMOTE_DST;
2706                         sd->sd_send_case |= LOCAL_DST;
2707                         return lnet_handle_send(sd);
2708                 }
2709
2710                 CERROR("Internal Error. Expected to have a best_lpni: "
2711                        "%s -> %s\n",
2712                        libcfs_nid2str(sd->sd_src_nid),
2713                        libcfs_nid2str(sd->sd_dst_nid));
2714
2715                 return -EFAULT;
2716         }
2717
2718         /*
2719          * Peer doesn't have a local network. Let's see if there is
2720          * a remote network we can reach it on.
2721          */
2722         return PASS_THROUGH;
2723 }
2724
2725 /*
2726  * Case 1:
2727  *      Source NID not specified
2728  *      Local destination
2729  *      MR peer
2730  *
2731  * Case 2:
2732  *      Source NID not speified
2733  *      Remote destination
2734  *      MR peer
2735  *
2736  * In both of these cases if we're sending a response, ACK or REPLY, then
2737  * we need to send to the destination NID provided.
2738  *
2739  * In the remote case let's deal with MR routers.
2740  *
2741  */
2742
2743 static int
2744 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2745 {
2746         int rc = 0;
2747         struct lnet_peer *gw_peer = NULL;
2748         struct lnet_peer_ni *gw_lpni = NULL;
2749
2750         /*
2751          * handle sending a response to a remote peer here so we don't
2752          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2753          */
2754         if (sd->sd_send_case & REMOTE_DST &&
2755             sd->sd_send_case & SND_RESP) {
2756                 struct lnet_peer_ni *gw;
2757                 struct lnet_peer *gw_peer;
2758
2759                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2760                                                   &gw_peer);
2761                 if (rc < 0) {
2762                         CERROR("Can't send response to %s. "
2763                                "No route available\n",
2764                                 libcfs_nid2str(sd->sd_dst_nid));
2765                         return -EHOSTUNREACH;
2766                 } else if (rc > 0) {
2767                         return rc;
2768                 }
2769
2770                 sd->sd_best_lpni = gw;
2771                 sd->sd_peer = gw_peer;
2772
2773                 return lnet_handle_send(sd);
2774         }
2775
2776         /*
2777          * Even though the NID for the peer might not be on a local network,
2778          * since the peer is MR there could be other interfaces on the
2779          * local network. In that case we'd still like to prefer the local
2780          * network over the routed network. If we're unable to do that
2781          * then we select the best router among the different routed networks,
2782          * and if the router is MR then we can deal with it as such.
2783          */
2784         rc = lnet_handle_any_mr_dsta(sd);
2785         if (rc != PASS_THROUGH)
2786                 return rc;
2787
2788         /*
2789          * Now that we must route to the destination, we must consider the
2790          * MR case, where the destination has multiple interfaces, some of
2791          * which we can route to and others we do not. For this reason we
2792          * need to select the destination which we can route to and if
2793          * there are multiple, we need to round robin.
2794          */
2795         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2796                                           &gw_peer);
2797         if (rc)
2798                 return rc;
2799
2800         sd->sd_send_case &= ~LOCAL_DST;
2801         sd->sd_send_case |= REMOTE_DST;
2802
2803         sd->sd_peer = gw_peer;
2804         sd->sd_best_lpni = gw_lpni;
2805
2806         return lnet_handle_send(sd);
2807 }
2808
2809 /*
2810  * Source not specified
2811  * Remote destination
2812  * Non-MR peer
2813  *
2814  * Must send to the specified peer NID using the same source NID that
2815  * we've used before. If it's the first time to talk to that peer then
2816  * find the source NI and assign it as preferred to that peer
2817  */
2818 static int
2819 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2820 {
2821         int rc;
2822         struct lnet_peer_ni *gw_lpni = NULL;
2823         struct lnet_peer *gw_peer = NULL;
2824
2825         /*
2826          * Let's see if we have a preferred NI to talk to this NMR peer
2827          */
2828         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2829                                                               sd->sd_cpt);
2830
2831         /*
2832          * find the router and that'll find the best NI if we didn't find
2833          * it already.
2834          */
2835         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2836                                           &gw_peer);
2837         if (rc)
2838                 return rc;
2839
2840         /*
2841          * set the best_ni we've chosen as the preferred one for
2842          * this peer
2843          */
2844         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2845
2846         /* we'll be sending to the gw */
2847         sd->sd_best_lpni = gw_lpni;
2848         sd->sd_peer = gw_peer;
2849
2850         return lnet_handle_send(sd);
2851 }
2852
2853 static int
2854 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2855 {
2856         /*
2857          * turn off the SND_RESP bit.
2858          * It will be checked in the case handling
2859          */
2860         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2861
2862         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2863                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2864                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2865                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2866                 libcfs_nid2str(sd->sd_dst_nid),
2867                 (send_case & LOCAL_DST) ? "local" : "routed");
2868
2869         switch (send_case) {
2870         /*
2871          * For all cases where the source is specified, we should always
2872          * use the destination NID, whether it's an MR destination or not,
2873          * since we're continuing a series of related messages for the
2874          * same RPC
2875          */
2876         case SRC_SPEC_LOCAL_NMR_DST:
2877                 return lnet_handle_spec_local_nmr_dst(sd);
2878         case SRC_SPEC_LOCAL_MR_DST:
2879                 return lnet_handle_spec_local_mr_dst(sd);
2880         case SRC_SPEC_ROUTER_NMR_DST:
2881         case SRC_SPEC_ROUTER_MR_DST:
2882                 return lnet_handle_spec_router_dst(sd);
2883         case SRC_ANY_LOCAL_NMR_DST:
2884                 return lnet_handle_any_local_nmr_dst(sd);
2885         case SRC_ANY_LOCAL_MR_DST:
2886         case SRC_ANY_ROUTER_MR_DST:
2887                 return lnet_handle_any_mr_dst(sd);
2888         case SRC_ANY_ROUTER_NMR_DST:
2889                 return lnet_handle_any_router_nmr_dst(sd);
2890         default:
2891                 CERROR("Unknown send case\n");
2892                 return -1;
2893         }
2894 }
2895
2896 static int
2897 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2898                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2899 {
2900         struct lnet_peer_ni *lpni;
2901         struct lnet_peer *peer;
2902         struct lnet_send_data send_data;
2903         int cpt, rc;
2904         int md_cpt;
2905         __u32 send_case = 0;
2906         bool final_hop;
2907         bool mr_forwarding_allowed;
2908
2909         memset(&send_data, 0, sizeof(send_data));
2910
2911         /*
2912          * get an initial CPT to use for locking. The idea here is not to
2913          * serialize the calls to select_pathway, so that as many
2914          * operations can run concurrently as possible. To do that we use
2915          * the CPT where this call is being executed. Later on when we
2916          * determine the CPT to use in lnet_message_commit, we switch the
2917          * lock and check if there was any configuration change.  If none,
2918          * then we proceed, if there is, then we restart the operation.
2919          */
2920         cpt = lnet_net_lock_current();
2921
2922         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2923         if (md_cpt == CFS_CPT_ANY)
2924                 md_cpt = cpt;
2925
2926 again:
2927
2928         /*
2929          * If we're being asked to send to the loopback interface, there
2930          * is no need to go through any selection. We can just shortcut
2931          * the entire process and send over lolnd
2932          */
2933         send_data.sd_msg = msg;
2934         send_data.sd_cpt = cpt;
2935         if (dst_nid == LNET_NID_LO_0) {
2936                 rc = lnet_handle_lo_send(&send_data);
2937                 lnet_net_unlock(cpt);
2938                 return rc;
2939         }
2940
2941         /*
2942          * find an existing peer_ni, or create one and mark it as having been
2943          * created due to network traffic. This call will create the
2944          * peer->peer_net->peer_ni tree.
2945          */
2946         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2947         if (IS_ERR(lpni)) {
2948                 lnet_net_unlock(cpt);
2949                 return PTR_ERR(lpni);
2950         }
2951
2952         /*
2953          * Cache the original src_nid and rtr_nid. If we need to resend the
2954          * message then we'll need to know whether the src_nid was originally
2955          * specified for this message. If it was originally specified,
2956          * then we need to keep using the same src_nid since it's
2957          * continuing the same sequence of messages. Similarly, rtr_nid will
2958          * affect our choice of next hop.
2959          */
2960         msg->msg_src_nid_param = src_nid;
2961         msg->msg_rtr_nid_param = rtr_nid;
2962
2963         /*
2964          * If necessary, perform discovery on the peer that owns this peer_ni.
2965          * Note, this can result in the ownership of this peer_ni changing
2966          * to another peer object.
2967          */
2968         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2969         if (rc) {
2970                 lnet_peer_ni_decref_locked(lpni);
2971                 lnet_net_unlock(cpt);
2972                 return rc;
2973         }
2974         lnet_peer_ni_decref_locked(lpni);
2975
2976         peer = lpni->lpni_peer_net->lpn_peer;
2977
2978         /*
2979          * Identify the different send cases
2980          */
2981         if (src_nid == LNET_NID_ANY)
2982                 send_case |= SRC_ANY;
2983         else
2984                 send_case |= SRC_SPEC;
2985
2986         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2987                 send_case |= LOCAL_DST;
2988         else
2989                 send_case |= REMOTE_DST;
2990
2991         final_hop = false;
2992         if (msg->msg_routing && (send_case & LOCAL_DST))
2993                 final_hop = true;
2994
2995         /* Determine whether to allow MR forwarding for this message.
2996          * NB: MR forwarding is allowed if the message originator and the
2997          * destination are both MR capable, and the destination lpni that was
2998          * originally chosen by the originator is unhealthy or down.
2999          * We check the MR capability of the destination further below
3000          */
3001         mr_forwarding_allowed = false;
3002         if (final_hop) {
3003                 struct lnet_peer *src_lp;
3004                 struct lnet_peer_ni *src_lpni;
3005
3006                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
3007                                                   LNET_NID_ANY, cpt);
3008                 /* We don't fail the send if we hit any errors here. We'll just
3009                  * try to send it via non-multi-rail criteria
3010                  */
3011                 if (!IS_ERR(src_lpni)) {
3012                         /* Drop ref taken by lnet_nid2peerni_locked() */
3013                         lnet_peer_ni_decref_locked(src_lpni);
3014                         src_lp = lpni->lpni_peer_net->lpn_peer;
3015                         if (lnet_peer_is_multi_rail(src_lp) &&
3016                             !lnet_is_peer_ni_alive(lpni))
3017                                 mr_forwarding_allowed = true;
3018
3019                 }
3020                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3021                        mr_forwarding_allowed ? "allowed" : "not allowed");
3022         }
3023
3024         /*
3025          * Deal with the peer as NMR in the following cases:
3026          * 1. the peer is NMR
3027          * 2. We're trying to recover a specific peer NI
3028          * 3. I'm a router sending to the final destination and MR forwarding is
3029          *    not allowed for this message (as determined above).
3030          *    In this case the source of the message would've
3031          *    already selected the final destination so my job
3032          *    is to honor the selection.
3033          */
3034         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3035             (final_hop && !mr_forwarding_allowed))
3036                 send_case |= NMR_DST;
3037         else
3038                 send_case |= MR_DST;
3039
3040         if (lnet_msg_is_response(msg))
3041                 send_case |= SND_RESP;
3042
3043         /* assign parameters to the send_data */
3044         send_data.sd_rtr_nid = rtr_nid;
3045         send_data.sd_src_nid = src_nid;
3046         send_data.sd_dst_nid = dst_nid;
3047         send_data.sd_best_lpni = lpni;
3048         /*
3049          * keep a pointer to the final destination in case we're going to
3050          * route, so we'll need to access it later
3051          */
3052         send_data.sd_final_dst_lpni = lpni;
3053         send_data.sd_peer = peer;
3054         send_data.sd_md_cpt = md_cpt;
3055         send_data.sd_send_case = send_case;
3056
3057         rc = lnet_handle_send_case_locked(&send_data);
3058
3059         /*
3060          * Update the local cpt since send_data.sd_cpt might've been
3061          * updated as a result of calling lnet_handle_send_case_locked().
3062          */
3063         cpt = send_data.sd_cpt;
3064
3065         if (rc == REPEAT_SEND)
3066                 goto again;
3067
3068         lnet_net_unlock(cpt);
3069
3070         return rc;
3071 }
3072
3073 int
3074 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3075 {
3076         lnet_nid_t              dst_nid = msg->msg_target.nid;
3077         int                     rc;
3078
3079         /*
3080          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3081          * but we might want to use pre-determined router for ACK/REPLY
3082          * in the future
3083          */
3084         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3085         LASSERT(msg->msg_txpeer == NULL);
3086         LASSERT(msg->msg_txni == NULL);
3087         LASSERT(!msg->msg_sending);
3088         LASSERT(!msg->msg_target_is_router);
3089         LASSERT(!msg->msg_receiving);
3090
3091         msg->msg_sending = 1;
3092
3093         LASSERT(!msg->msg_tx_committed);
3094
3095         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3096         if (rc < 0) {
3097                 if (rc == -EHOSTUNREACH)
3098                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3099                 else
3100                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3101                 return rc;
3102         }
3103
3104         if (rc == LNET_CREDIT_OK)
3105                 lnet_ni_send(msg->msg_txni, msg);
3106
3107         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3108         return 0;
3109 }
3110
3111 enum lnet_mt_event_type {
3112         MT_TYPE_LOCAL_NI = 0,
3113         MT_TYPE_PEER_NI
3114 };
3115
3116 struct lnet_mt_event_info {
3117         enum lnet_mt_event_type mt_type;
3118         lnet_nid_t mt_nid;
3119 };
3120
3121 /* called with res_lock held */
3122 void
3123 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3124 {
3125         struct lnet_rsp_tracker *rspt;
3126
3127         /*
3128          * msg has a refcount on the MD so the MD is not going away.
3129          * The rspt queue for the cpt is protected by
3130          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3131          */
3132         if (!md->md_rspt_ptr)
3133                 return;
3134
3135         rspt = md->md_rspt_ptr;
3136
3137         /* debug code */
3138         LASSERT(rspt->rspt_cpt == cpt);
3139
3140         md->md_rspt_ptr = NULL;
3141
3142         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3143                 /*
3144                  * The monitor thread has invalidated this handle because the
3145                  * response timed out, but it failed to lookup the MD. That
3146                  * means this response tracker is on the zombie list. We can
3147                  * safely remove it under the resource lock (held by caller) and
3148                  * free the response tracker block.
3149                  */
3150                 list_del(&rspt->rspt_on_list);
3151                 lnet_rspt_free(rspt, cpt);
3152         } else {
3153                 /*
3154                  * invalidate the handle to indicate that a response has been
3155                  * received, which will then lead the monitor thread to clean up
3156                  * the rspt block.
3157                  */
3158                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3159         }
3160 }
3161
3162 void
3163 lnet_clean_zombie_rstqs(void)
3164 {
3165         struct lnet_rsp_tracker *rspt, *tmp;
3166         int i;
3167
3168         cfs_cpt_for_each(i, lnet_cpt_table()) {
3169                 list_for_each_entry_safe(rspt, tmp,
3170                                          the_lnet.ln_mt_zombie_rstqs[i],
3171                                          rspt_on_list) {
3172                         list_del(&rspt->rspt_on_list);
3173                         lnet_rspt_free(rspt, i);
3174                 }
3175         }
3176
3177         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3178 }
3179
3180 static void
3181 lnet_finalize_expired_responses(void)
3182 {
3183         struct lnet_libmd *md;
3184         struct lnet_rsp_tracker *rspt, *tmp;
3185         ktime_t now;
3186         int i;
3187
3188         if (the_lnet.ln_mt_rstq == NULL)
3189                 return;
3190
3191         cfs_cpt_for_each(i, lnet_cpt_table()) {
3192                 LIST_HEAD(local_queue);
3193
3194                 lnet_net_lock(i);
3195                 if (!the_lnet.ln_mt_rstq[i]) {
3196                         lnet_net_unlock(i);
3197                         continue;
3198                 }
3199                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3200                 lnet_net_unlock(i);
3201
3202                 now = ktime_get();
3203
3204                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3205                         /*
3206                          * The rspt mdh will be invalidated when a response
3207                          * is received or whenever we want to discard the
3208                          * block the monitor thread will walk the queue
3209                          * and clean up any rsts with an invalid mdh.
3210                          * The monitor thread will walk the queue until
3211                          * the first unexpired rspt block. This means that
3212                          * some rspt blocks which received their
3213                          * corresponding responses will linger in the
3214                          * queue until they are cleaned up eventually.
3215                          */
3216                         lnet_res_lock(i);
3217                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3218                                 lnet_res_unlock(i);
3219                                 list_del(&rspt->rspt_on_list);
3220                                 lnet_rspt_free(rspt, i);
3221                                 continue;
3222                         }
3223
3224                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3225                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3226                                 struct lnet_peer_ni *lpni;
3227                                 struct lnet_nid nid;
3228
3229                                 md = lnet_handle2md(&rspt->rspt_mdh);
3230                                 if (!md) {
3231                                         /* MD has been queued for unlink, but
3232                                          * rspt hasn't been detached (Note we've
3233                                          * checked above that the rspt_mdh is
3234                                          * valid). Since we cannot lookup the MD
3235                                          * we're unable to detach the rspt
3236                                          * ourselves. Thus, move the rspt to the
3237                                          * zombie list where we'll wait for
3238                                          * either:
3239                                          *   1. The remaining operations on the
3240                                          *   MD to complete. In this case the
3241                                          *   final operation will result in
3242                                          *   lnet_msg_detach_md()->
3243                                          *   lnet_detach_rsp_tracker() where
3244                                          *   we will clean up this response
3245                                          *   tracker.
3246                                          *   2. LNet to shutdown. In this case
3247                                          *   we'll wait until after all LND Nets
3248                                          *   have shutdown and then we can
3249                                          *   safely free any remaining response
3250                                          *   tracker blocks on the zombie list.
3251                                          * Note: We need to hold the resource
3252                                          * lock when adding to the zombie list
3253                                          * because we may have concurrent access
3254                                          * with lnet_detach_rsp_tracker().
3255                                          */
3256                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
3257                                         list_move(&rspt->rspt_on_list,
3258                                                   the_lnet.ln_mt_zombie_rstqs[i]);
3259                                         lnet_res_unlock(i);
3260                                         continue;
3261                                 }
3262                                 LASSERT(md->md_rspt_ptr == rspt);
3263                                 md->md_rspt_ptr = NULL;
3264                                 lnet_res_unlock(i);
3265
3266                                 LNetMDUnlink(rspt->rspt_mdh);
3267
3268                                 nid = rspt->rspt_next_hop_nid;
3269
3270                                 list_del(&rspt->rspt_on_list);
3271                                 lnet_rspt_free(rspt, i);
3272
3273                                 /* If we're shutting down we just want to clean
3274                                  * up the rspt blocks
3275                                  */
3276                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3277                                         continue;
3278
3279                                 lnet_net_lock(i);
3280                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3281                                 lnet_net_unlock(i);
3282
3283                                 CDEBUG(D_NET,
3284                                        "Response timeout: md = %p: nid = %s\n",
3285                                        md, libcfs_nidstr(&nid));
3286
3287                                 /*
3288                                  * If there is a timeout on the response
3289                                  * from the next hop decrement its health
3290                                  * value so that we don't use it
3291                                  */
3292                                 lnet_net_lock(0);
3293                                 lpni = lnet_peer_ni_find_locked(&nid);
3294                                 if (lpni) {
3295                                         lnet_handle_remote_failure_locked(lpni);
3296                                         lnet_peer_ni_decref_locked(lpni);
3297                                 }
3298                                 lnet_net_unlock(0);
3299                         } else {
3300                                 lnet_res_unlock(i);
3301                                 break;
3302                         }
3303                 }
3304
3305                 if (!list_empty(&local_queue)) {
3306                         lnet_net_lock(i);
3307                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3308                         lnet_net_unlock(i);
3309                 }
3310         }
3311 }
3312
3313 static void
3314 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3315 {
3316         struct lnet_msg *msg;
3317
3318         while (!list_empty(resendq)) {
3319                 struct lnet_peer_ni *lpni;
3320
3321                 msg = list_entry(resendq->next, struct lnet_msg,
3322                                  msg_list);
3323
3324                 list_del_init(&msg->msg_list);
3325
3326                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3327                 if (!lpni) {
3328                         lnet_net_unlock(cpt);
3329                         CERROR("Expected that a peer is already created for %s\n",
3330                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3331                         msg->msg_no_resend = true;
3332                         lnet_finalize(msg, -EFAULT);
3333                         lnet_net_lock(cpt);
3334                 } else {
3335                         int rc;
3336
3337                         lnet_peer_ni_decref_locked(lpni);
3338
3339                         lnet_net_unlock(cpt);
3340                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3341                                libcfs_nid2str(msg->msg_src_nid_param),
3342                                libcfs_id2str(msg->msg_target),
3343                                lnet_msgtyp2str(msg->msg_type),
3344                                msg->msg_recovery,
3345                                msg->msg_retry_count);
3346                         rc = lnet_send(msg->msg_src_nid_param, msg,
3347                                        msg->msg_rtr_nid_param);
3348                         if (rc) {
3349                                 CERROR("Error sending %s to %s: %d\n",
3350                                        lnet_msgtyp2str(msg->msg_type),
3351                                        libcfs_id2str(msg->msg_target), rc);
3352                                 msg->msg_no_resend = true;
3353                                 lnet_finalize(msg, rc);
3354                         }
3355                         lnet_net_lock(cpt);
3356                         if (!rc)
3357                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3358                 }
3359         }
3360 }
3361
3362 static void
3363 lnet_resend_pending_msgs(void)
3364 {
3365         int i;
3366
3367         cfs_cpt_for_each(i, lnet_cpt_table()) {
3368                 lnet_net_lock(i);
3369                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3370                 lnet_net_unlock(i);
3371         }
3372 }
3373
3374 /* called with cpt and ni_lock held */
3375 static void
3376 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3377 {
3378         struct lnet_handle_md recovery_mdh;
3379
3380         LNetInvalidateMDHandle(&recovery_mdh);
3381
3382         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3383             force) {
3384                 recovery_mdh = ni->ni_ping_mdh;
3385                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3386         }
3387         lnet_ni_unlock(ni);
3388         lnet_net_unlock(cpt);
3389         if (!LNetMDHandleIsInvalid(recovery_mdh))
3390                 LNetMDUnlink(recovery_mdh);
3391         lnet_net_lock(cpt);
3392         lnet_ni_lock(ni);
3393 }
3394
3395 static void
3396 lnet_recover_local_nis(void)
3397 {
3398         struct lnet_mt_event_info *ev_info;
3399         LIST_HEAD(processed_list);
3400         LIST_HEAD(local_queue);
3401         struct lnet_handle_md mdh;
3402         struct lnet_ni *tmp;
3403         struct lnet_ni *ni;
3404         lnet_nid_t nid;
3405         int healthv;
3406         int rc;
3407         time64_t now;
3408
3409         /*
3410          * splice the recovery queue on a local queue. We will iterate
3411          * through the local queue and update it as needed. Once we're
3412          * done with the traversal, we'll splice the local queue back on
3413          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3414          * will be traversed in the next iteration.
3415          */
3416         lnet_net_lock(0);
3417         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3418                          &local_queue);
3419         lnet_net_unlock(0);
3420
3421         now = ktime_get_seconds();
3422
3423         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3424                 /*
3425                  * if an NI is being deleted or it is now healthy, there
3426                  * is no need to keep it around in the recovery queue.
3427                  * The monitor thread is the only thread responsible for
3428                  * removing the NI from the recovery queue.
3429                  * Multiple threads can be adding NIs to the recovery
3430                  * queue.
3431                  */
3432                 healthv = atomic_read(&ni->ni_healthv);
3433
3434                 lnet_net_lock(0);
3435                 lnet_ni_lock(ni);
3436                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3437                     healthv == LNET_MAX_HEALTH_VALUE) {
3438                         list_del_init(&ni->ni_recovery);
3439                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3440                         lnet_ni_unlock(ni);
3441                         lnet_ni_decref_locked(ni, 0);
3442                         lnet_net_unlock(0);
3443                         continue;
3444                 }
3445
3446                 /*
3447                  * if the local NI failed recovery we must unlink the md.
3448                  * But we want to keep the local_ni on the recovery queue
3449                  * so we can continue the attempts to recover it.
3450                  */
3451                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3452                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3453                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3454                 }
3455
3456
3457                 lnet_ni_unlock(ni);
3458
3459                 if (now < ni->ni_next_ping) {
3460                         lnet_net_unlock(0);
3461                         continue;
3462                 }
3463
3464                 lnet_net_unlock(0);
3465
3466                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3467                        libcfs_nidstr(&ni->ni_nid));
3468
3469                 lnet_ni_lock(ni);
3470                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3471                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3472                         lnet_ni_unlock(ni);
3473
3474                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3475                         if (!ev_info) {
3476                                 CERROR("out of memory. Can't recover %s\n",
3477                                        libcfs_nidstr(&ni->ni_nid));
3478                                 lnet_ni_lock(ni);
3479                                 ni->ni_recovery_state &=
3480                                   ~LNET_NI_RECOVERY_PENDING;
3481                                 lnet_ni_unlock(ni);
3482                                 continue;
3483                         }
3484
3485                         mdh = ni->ni_ping_mdh;
3486                         /*
3487                          * Invalidate the ni mdh in case it's deleted.
3488                          * We'll unlink the mdh in this case below.
3489                          */
3490                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3491                         /* FIXME need to handle large-addr nid */
3492                         nid = lnet_nid_to_nid4(&ni->ni_nid);
3493
3494                         /*
3495                          * remove the NI from the local queue and drop the
3496                          * reference count to it while we're recovering
3497                          * it. The reason for that, is that the NI could
3498                          * be deleted, and the way the code is structured
3499                          * is if we don't drop the NI, then the deletion
3500                          * code will enter a loop waiting for the
3501                          * reference count to be removed while holding the
3502                          * ln_mutex_lock(). When we look up the peer to
3503                          * send to in lnet_select_pathway() we will try to
3504                          * lock the ln_mutex_lock() as well, leading to
3505                          * a deadlock. By dropping the refcount and
3506                          * removing it from the list, we allow for the NI
3507                          * to be removed, then we use the cached NID to
3508                          * look it up again. If it's gone, then we just
3509                          * continue examining the rest of the queue.
3510                          */
3511                         lnet_net_lock(0);
3512                         list_del_init(&ni->ni_recovery);
3513                         lnet_ni_decref_locked(ni, 0);
3514                         lnet_net_unlock(0);
3515
3516                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3517                         ev_info->mt_nid = nid;
3518                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3519                                             ev_info, the_lnet.ln_mt_handler,
3520                                             true);
3521                         /* lookup the nid again */
3522                         lnet_net_lock(0);
3523                         ni = lnet_nid2ni_locked(nid, 0);
3524                         if (!ni) {
3525                                 /*
3526                                  * the NI has been deleted when we dropped
3527                                  * the ref count
3528                                  */
3529                                 lnet_net_unlock(0);
3530                                 LNetMDUnlink(mdh);
3531                                 continue;
3532                         }
3533                         ni->ni_ping_count++;
3534
3535                         ni->ni_ping_mdh = mdh;
3536                         lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3537                                                         now);
3538
3539                         if (rc) {
3540                                 lnet_ni_lock(ni);
3541                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3542                                 lnet_ni_unlock(ni);
3543                         }
3544                         lnet_net_unlock(0);
3545                 } else
3546                         lnet_ni_unlock(ni);
3547         }
3548
3549         /*
3550          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3551          * reexamined in the next iteration.
3552          */
3553         list_splice_init(&processed_list, &local_queue);
3554         lnet_net_lock(0);
3555         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3556         lnet_net_unlock(0);
3557 }
3558
3559 static int
3560 lnet_resendqs_create(void)
3561 {
3562         struct list_head **resendqs;
3563         resendqs = lnet_create_array_of_queues();
3564
3565         if (!resendqs)
3566                 return -ENOMEM;
3567
3568         lnet_net_lock(LNET_LOCK_EX);
3569         the_lnet.ln_mt_resendqs = resendqs;
3570         lnet_net_unlock(LNET_LOCK_EX);
3571
3572         return 0;
3573 }
3574
3575 static void
3576 lnet_clean_local_ni_recoveryq(void)
3577 {
3578         struct lnet_ni *ni;
3579
3580         /* This is only called when the monitor thread has stopped */
3581         lnet_net_lock(0);
3582
3583         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3584                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3585                                 struct lnet_ni, ni_recovery);
3586                 list_del_init(&ni->ni_recovery);
3587                 lnet_ni_lock(ni);
3588                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3589                 lnet_ni_unlock(ni);
3590                 lnet_ni_decref_locked(ni, 0);
3591         }
3592
3593         lnet_net_unlock(0);
3594 }
3595
3596 static void
3597 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3598                                      bool force)
3599 {
3600         struct lnet_handle_md recovery_mdh;
3601
3602         LNetInvalidateMDHandle(&recovery_mdh);
3603
3604         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3605                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3606                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3607         }
3608         spin_unlock(&lpni->lpni_lock);
3609         lnet_net_unlock(cpt);
3610         if (!LNetMDHandleIsInvalid(recovery_mdh))
3611                 LNetMDUnlink(recovery_mdh);
3612         lnet_net_lock(cpt);
3613         spin_lock(&lpni->lpni_lock);
3614 }
3615
3616 static void
3617 lnet_clean_peer_ni_recoveryq(void)
3618 {
3619         struct lnet_peer_ni *lpni, *tmp;
3620
3621         lnet_net_lock(LNET_LOCK_EX);
3622
3623         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3624                                  lpni_recovery) {
3625                 list_del_init(&lpni->lpni_recovery);
3626                 spin_lock(&lpni->lpni_lock);
3627                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3628                 spin_unlock(&lpni->lpni_lock);
3629                 lnet_peer_ni_decref_locked(lpni);
3630         }
3631
3632         lnet_net_unlock(LNET_LOCK_EX);
3633 }
3634
3635 static void
3636 lnet_clean_resendqs(void)
3637 {
3638         struct lnet_msg *msg, *tmp;
3639         LIST_HEAD(msgs);
3640         int i;
3641
3642         cfs_cpt_for_each(i, lnet_cpt_table()) {
3643                 lnet_net_lock(i);
3644                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3645                 lnet_net_unlock(i);
3646                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3647                         list_del_init(&msg->msg_list);
3648                         msg->msg_no_resend = true;
3649                         lnet_finalize(msg, -ESHUTDOWN);
3650                 }
3651         }
3652
3653         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3654 }
3655
3656 static void
3657 lnet_recover_peer_nis(void)
3658 {
3659         struct lnet_mt_event_info *ev_info;
3660         LIST_HEAD(processed_list);
3661         LIST_HEAD(local_queue);
3662         struct lnet_handle_md mdh;
3663         struct lnet_peer_ni *lpni;
3664         struct lnet_peer_ni *tmp;
3665         lnet_nid_t nid;
3666         int healthv;
3667         int rc;
3668         time64_t now;
3669
3670         /*
3671          * Always use cpt 0 for locking across all interactions with
3672          * ln_mt_peerNIRecovq
3673          */
3674         lnet_net_lock(0);
3675         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3676                          &local_queue);
3677         lnet_net_unlock(0);
3678
3679         now = ktime_get_seconds();
3680
3681         list_for_each_entry_safe(lpni, tmp, &local_queue,
3682                                  lpni_recovery) {
3683                 /*
3684                  * The same protection strategy is used here as is in the
3685                  * local recovery case.
3686                  */
3687                 lnet_net_lock(0);
3688                 healthv = atomic_read(&lpni->lpni_healthv);
3689                 spin_lock(&lpni->lpni_lock);
3690                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3691                     healthv == LNET_MAX_HEALTH_VALUE) {
3692                         list_del_init(&lpni->lpni_recovery);
3693                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3694                         spin_unlock(&lpni->lpni_lock);
3695                         lnet_peer_ni_decref_locked(lpni);
3696                         lnet_net_unlock(0);
3697                         continue;
3698                 }
3699
3700                 /*
3701                  * If the peer NI has failed recovery we must unlink the
3702                  * md. But we want to keep the peer ni on the recovery
3703                  * queue so we can try to continue recovering it
3704                  */
3705                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3706                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3707                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3708                 }
3709
3710                 spin_unlock(&lpni->lpni_lock);
3711
3712                 if (now < lpni->lpni_next_ping) {
3713                         lnet_net_unlock(0);
3714                         continue;
3715                 }
3716
3717                 lnet_net_unlock(0);
3718
3719                 /*
3720                  * NOTE: we're racing with peer deletion from user space.
3721                  * It's possible that a peer is deleted after we check its
3722                  * state. In this case the recovery can create a new peer
3723                  */
3724                 spin_lock(&lpni->lpni_lock);
3725                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3726                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3727                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3728                         spin_unlock(&lpni->lpni_lock);
3729
3730                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3731                         if (!ev_info) {
3732                                 CERROR("out of memory. Can't recover %s\n",
3733                                        libcfs_nidstr(&lpni->lpni_nid));
3734                                 spin_lock(&lpni->lpni_lock);
3735                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3736                                 spin_unlock(&lpni->lpni_lock);
3737                                 continue;
3738                         }
3739
3740                         /* look at the comments in lnet_recover_local_nis() */
3741                         mdh = lpni->lpni_recovery_ping_mdh;
3742                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3743                         /* FIXME handle large-addr nid */
3744                         nid = lnet_nid_to_nid4(&lpni->lpni_nid);
3745                         lnet_net_lock(0);
3746                         list_del_init(&lpni->lpni_recovery);
3747                         lnet_peer_ni_decref_locked(lpni);
3748                         lnet_net_unlock(0);
3749
3750                         ev_info->mt_type = MT_TYPE_PEER_NI;
3751                         ev_info->mt_nid = nid;
3752                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3753                                             ev_info, the_lnet.ln_mt_handler,
3754                                             true);
3755                         lnet_net_lock(0);
3756                         /*
3757                          * lnet_find_peer_ni_locked() grabs a refcount for
3758                          * us. No need to take it explicitly.
3759                          */
3760                         lpni = lnet_find_peer_ni_locked(nid);
3761                         if (!lpni) {
3762                                 lnet_net_unlock(0);
3763                                 LNetMDUnlink(mdh);
3764                                 continue;
3765                         }
3766
3767                         lpni->lpni_ping_count++;
3768
3769                         lpni->lpni_recovery_ping_mdh = mdh;
3770
3771                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
3772                                                              &processed_list,
3773                                                              now);
3774                         if (rc) {
3775                                 spin_lock(&lpni->lpni_lock);
3776                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3777                                 spin_unlock(&lpni->lpni_lock);
3778                         }
3779
3780                         /* Drop the ref taken by lnet_find_peer_ni_locked() */
3781                         lnet_peer_ni_decref_locked(lpni);
3782                         lnet_net_unlock(0);
3783                 } else
3784                         spin_unlock(&lpni->lpni_lock);
3785         }
3786
3787         list_splice_init(&processed_list, &local_queue);
3788         lnet_net_lock(0);
3789         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3790         lnet_net_unlock(0);
3791 }
3792
3793 static int
3794 lnet_monitor_thread(void *arg)
3795 {
3796         time64_t rsp_timeout = 0;
3797         time64_t now;
3798
3799         wait_for_completion(&the_lnet.ln_started);
3800         /*
3801          * The monitor thread takes care of the following:
3802          *  1. Checks the aliveness of routers
3803          *  2. Checks if there are messages on the resend queue to resend
3804          *     them.
3805          *  3. Check if there are any NIs on the local recovery queue and
3806          *     pings them
3807          *  4. Checks if there are any NIs on the remote recovery queue
3808          *     and pings them.
3809          */
3810         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3811                 now = ktime_get_real_seconds();
3812
3813                 if (lnet_router_checker_active())
3814                         lnet_check_routers();
3815
3816                 lnet_resend_pending_msgs();
3817
3818                 if (now >= rsp_timeout) {
3819                         lnet_finalize_expired_responses();
3820                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3821                 }
3822
3823                 lnet_recover_local_nis();
3824                 lnet_recover_peer_nis();
3825
3826                 /*
3827                  * TODO do we need to check if we should sleep without
3828                  * timeout?  Technically, an active system will always
3829                  * have messages in flight so this check will always
3830                  * evaluate to false. And on an idle system do we care
3831                  * if we wake up every 1 second? Although, we've seen
3832                  * cases where we get a complaint that an idle thread
3833                  * is waking up unnecessarily.
3834                  */
3835                 wait_for_completion_interruptible_timeout(
3836                         &the_lnet.ln_mt_wait_complete,
3837                         cfs_time_seconds(1));
3838                 /* Must re-init the completion before testing anything,
3839                  * including ln_mt_state.
3840                  */
3841                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3842         }
3843
3844         /* Shutting down */
3845         lnet_net_lock(LNET_LOCK_EX);
3846         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3847         lnet_net_unlock(LNET_LOCK_EX);
3848
3849         /* signal that the monitor thread is exiting */
3850         up(&the_lnet.ln_mt_signal);
3851
3852         return 0;
3853 }
3854
3855 /*
3856  * lnet_send_ping
3857  * Sends a ping.
3858  * Returns == 0 if success
3859  * Returns > 0 if LNetMDBind or prior fails
3860  * Returns < 0 if LNetGet fails
3861  */
3862 int
3863 lnet_send_ping(lnet_nid_t dest_nid,
3864                struct lnet_handle_md *mdh, int nnis,
3865                void *user_data, lnet_handler_t handler, bool recovery)
3866 {
3867         struct lnet_md md = { NULL };
3868         struct lnet_process_id id;
3869         struct lnet_ping_buffer *pbuf;
3870         int rc;
3871
3872         if (dest_nid == LNET_NID_ANY) {
3873                 rc = -EHOSTUNREACH;
3874                 goto fail_error;
3875         }
3876
3877         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3878         if (!pbuf) {
3879                 rc = ENOMEM;
3880                 goto fail_error;
3881         }
3882
3883         /* initialize md content */
3884         md.start     = &pbuf->pb_info;
3885         md.length    = LNET_PING_INFO_SIZE(nnis);
3886         md.threshold = 2; /* GET/REPLY */
3887         md.max_size  = 0;
3888         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3889         md.user_ptr  = user_data;
3890         md.handler   = handler;
3891
3892         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3893         if (rc) {
3894                 lnet_ping_buffer_decref(pbuf);
3895                 CERROR("Can't bind MD: %d\n", rc);
3896                 rc = -rc; /* change the rc to positive */
3897                 goto fail_error;
3898         }
3899         id.pid = LNET_PID_LUSTRE;
3900         id.nid = dest_nid;
3901
3902         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3903                      LNET_RESERVED_PORTAL,
3904                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3905
3906         if (rc)
3907                 goto fail_unlink_md;
3908
3909         return 0;
3910
3911 fail_unlink_md:
3912         LNetMDUnlink(*mdh);
3913         LNetInvalidateMDHandle(mdh);
3914 fail_error:
3915         return rc;
3916 }
3917
3918 static void
3919 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3920                            int status, bool send, bool unlink_event)
3921 {
3922         lnet_nid_t nid = ev_info->mt_nid;
3923
3924         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3925                 struct lnet_ni *ni;
3926
3927                 lnet_net_lock(0);
3928                 ni = lnet_nid2ni_locked(nid, 0);
3929                 if (!ni) {
3930                         lnet_net_unlock(0);
3931                         return;
3932                 }
3933                 lnet_ni_lock(ni);
3934                 if (!send || (send && status != 0))
3935                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3936                 if (status)
3937                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3938                 lnet_ni_unlock(ni);
3939                 lnet_net_unlock(0);
3940
3941                 if (status != 0) {
3942                         CERROR("local NI (%s) recovery failed with %d\n",
3943                                libcfs_nid2str(nid), status);
3944                         return;
3945                 }
3946                 /*
3947                  * need to increment healthv for the ni here, because in
3948                  * the lnet_finalize() path we don't have access to this
3949                  * NI. And in order to get access to it, we'll need to
3950                  * carry forward too much information.
3951                  * In the peer case, it'll naturally be incremented
3952                  */
3953                 if (!unlink_event)
3954                         lnet_inc_healthv(&ni->ni_healthv,
3955                                          lnet_health_sensitivity);
3956         } else {
3957                 struct lnet_peer_ni *lpni;
3958                 int cpt;
3959
3960                 cpt = lnet_net_lock_current();
3961                 lpni = lnet_find_peer_ni_locked(nid);
3962                 if (!lpni) {
3963                         lnet_net_unlock(cpt);
3964                         return;
3965                 }
3966                 spin_lock(&lpni->lpni_lock);
3967                 if (!send || (send && status != 0))
3968                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3969                 if (status)
3970                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3971                 spin_unlock(&lpni->lpni_lock);
3972                 lnet_peer_ni_decref_locked(lpni);
3973                 lnet_net_unlock(cpt);
3974
3975                 if (status != 0)
3976                         CERROR("peer NI (%s) recovery failed with %d\n",
3977                                libcfs_nid2str(nid), status);
3978         }
3979 }
3980
3981 void
3982 lnet_mt_event_handler(struct lnet_event *event)
3983 {
3984         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3985         struct lnet_ping_buffer *pbuf;
3986
3987         /* TODO: remove assert */
3988         LASSERT(event->type == LNET_EVENT_REPLY ||
3989                 event->type == LNET_EVENT_SEND ||
3990                 event->type == LNET_EVENT_UNLINK);
3991
3992         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3993                event->status);
3994
3995         switch (event->type) {
3996         case LNET_EVENT_UNLINK:
3997                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3998                        libcfs_nid2str(ev_info->mt_nid));
3999                 /* fallthrough */
4000         case LNET_EVENT_REPLY:
4001                 lnet_handle_recovery_reply(ev_info, event->status, false,
4002                                            event->type == LNET_EVENT_UNLINK);
4003                 break;
4004         case LNET_EVENT_SEND:
4005                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4006                                libcfs_nid2str(ev_info->mt_nid),
4007                                (event->status) ? "unsuccessfully" :
4008                                "successfully", event->status);
4009                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4010                 break;
4011         default:
4012                 CERROR("Unexpected event: %d\n", event->type);
4013                 break;
4014         }
4015         if (event->unlinked) {
4016                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4017                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4018                 lnet_ping_buffer_decref(pbuf);
4019         }
4020 }
4021
4022 static int
4023 lnet_rsp_tracker_create(void)
4024 {
4025         struct list_head **rstqs;
4026         rstqs = lnet_create_array_of_queues();
4027
4028         if (!rstqs)
4029                 return -ENOMEM;
4030
4031         the_lnet.ln_mt_rstq = rstqs;
4032
4033         return 0;
4034 }
4035
4036 static void
4037 lnet_rsp_tracker_clean(void)
4038 {
4039         lnet_finalize_expired_responses();
4040
4041         cfs_percpt_free(the_lnet.ln_mt_rstq);
4042         the_lnet.ln_mt_rstq = NULL;
4043 }
4044
4045 int lnet_monitor_thr_start(void)
4046 {
4047         int rc = 0;
4048         struct task_struct *task;
4049
4050         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4051                 return -EALREADY;
4052
4053         rc = lnet_resendqs_create();
4054         if (rc)
4055                 return rc;
4056
4057         rc = lnet_rsp_tracker_create();
4058         if (rc)
4059                 goto clean_queues;
4060
4061         sema_init(&the_lnet.ln_mt_signal, 0);
4062
4063         lnet_net_lock(LNET_LOCK_EX);
4064         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4065         lnet_net_unlock(LNET_LOCK_EX);
4066         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4067         if (IS_ERR(task)) {
4068                 rc = PTR_ERR(task);
4069                 CERROR("Can't start monitor thread: %d\n", rc);
4070                 goto clean_thread;
4071         }
4072
4073         return 0;
4074
4075 clean_thread:
4076         lnet_net_lock(LNET_LOCK_EX);
4077         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4078         lnet_net_unlock(LNET_LOCK_EX);
4079         /* block until event callback signals exit */
4080         down(&the_lnet.ln_mt_signal);
4081         /* clean up */
4082         lnet_net_lock(LNET_LOCK_EX);
4083         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4084         lnet_net_unlock(LNET_LOCK_EX);
4085         lnet_rsp_tracker_clean();
4086         lnet_clean_local_ni_recoveryq();
4087         lnet_clean_peer_ni_recoveryq();
4088         lnet_clean_resendqs();
4089         the_lnet.ln_mt_handler = NULL;
4090         return rc;
4091 clean_queues:
4092         lnet_rsp_tracker_clean();
4093         lnet_clean_local_ni_recoveryq();
4094         lnet_clean_peer_ni_recoveryq();
4095         lnet_clean_resendqs();
4096         return rc;
4097 }
4098
4099 void lnet_monitor_thr_stop(void)
4100 {
4101         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4102                 return;
4103
4104         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4105         lnet_net_lock(LNET_LOCK_EX);
4106         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4107         lnet_net_unlock(LNET_LOCK_EX);
4108
4109         /* tell the monitor thread that we're shutting down */
4110         complete(&the_lnet.ln_mt_wait_complete);
4111
4112         /* block until monitor thread signals that it's done */
4113         down(&the_lnet.ln_mt_signal);
4114         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4115
4116         /* perform cleanup tasks */
4117         lnet_rsp_tracker_clean();
4118         lnet_clean_local_ni_recoveryq();
4119         lnet_clean_peer_ni_recoveryq();
4120         lnet_clean_resendqs();
4121 }
4122
4123 void
4124 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4125                   __u32 msg_type)
4126 {
4127         lnet_net_lock(cpt);
4128         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4129         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4130         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4131         lnet_net_unlock(cpt);
4132
4133         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4134 }
4135
4136 static void
4137 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4138 {
4139         struct lnet_hdr *hdr = &msg->msg_hdr;
4140
4141         if (msg->msg_wanted != 0)
4142                 lnet_setpayloadbuffer(msg);
4143
4144         lnet_build_msg_event(msg, LNET_EVENT_PUT);
4145
4146         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
4147          * it back into the ACK during lnet_finalize() */
4148         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4149                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4150
4151         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4152                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4153 }
4154
4155 static int
4156 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4157 {
4158         struct lnet_hdr         *hdr = &msg->msg_hdr;
4159         struct lnet_match_info  info;
4160         int                     rc;
4161         bool                    ready_delay;
4162
4163         /* Convert put fields to host byte order */
4164         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4165         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
4166         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
4167
4168         /* Primary peer NID. */
4169         info.mi_id.nid  = msg->msg_initiator;
4170         info.mi_id.pid  = hdr->src_pid;
4171         info.mi_opc     = LNET_MD_OP_PUT;
4172         info.mi_portal  = hdr->msg.put.ptl_index;
4173         info.mi_rlength = hdr->payload_length;
4174         info.mi_roffset = hdr->msg.put.offset;
4175         info.mi_mbits   = hdr->msg.put.match_bits;
4176         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
4177
4178         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4179         ready_delay = msg->msg_rx_ready_delay;
4180
4181  again:
4182         rc = lnet_ptl_match_md(&info, msg);
4183         switch (rc) {
4184         default:
4185                 LBUG();
4186
4187         case LNET_MATCHMD_OK:
4188                 lnet_recv_put(ni, msg);
4189                 return 0;
4190
4191         case LNET_MATCHMD_NONE:
4192                 if (ready_delay)
4193                         /* no eager_recv or has already called it, should
4194                          * have been attached on delayed list */
4195                         return 0;
4196
4197                 rc = lnet_ni_eager_recv(ni, msg);
4198                 if (rc == 0) {
4199                         ready_delay = true;
4200                         goto again;
4201                 }
4202                 /* fall through */
4203
4204         case LNET_MATCHMD_DROP:
4205                 CNETERR("Dropping PUT from %s portal %d match %llu"
4206                         " offset %d length %d: %d\n",
4207                         libcfs_id2str(info.mi_id), info.mi_portal,
4208                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4209
4210                 return -ENOENT; /* -ve: OK but no match */
4211         }
4212 }
4213
4214 static int
4215 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4216 {
4217         struct lnet_match_info info;
4218         struct lnet_hdr *hdr = &msg->msg_hdr;
4219         struct lnet_process_id source_id;
4220         struct lnet_handle_wire reply_wmd;
4221         int rc;
4222
4223         /* Convert get fields to host byte order */
4224         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
4225         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
4226         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
4227         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
4228
4229         source_id.nid = hdr->src_nid;
4230         source_id.pid = hdr->src_pid;
4231         /* Primary peer NID */
4232         info.mi_id.nid  = msg->msg_initiator;
4233         info.mi_id.pid  = hdr->src_pid;
4234         info.mi_opc     = LNET_MD_OP_GET;
4235         info.mi_portal  = hdr->msg.get.ptl_index;
4236         info.mi_rlength = hdr->msg.get.sink_length;
4237         info.mi_roffset = hdr->msg.get.src_offset;
4238         info.mi_mbits   = hdr->msg.get.match_bits;
4239         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
4240
4241         rc = lnet_ptl_match_md(&info, msg);
4242         if (rc == LNET_MATCHMD_DROP) {
4243                 CNETERR("Dropping GET from %s portal %d match %llu"
4244                         " offset %d length %d\n",
4245                         libcfs_id2str(info.mi_id), info.mi_portal,
4246                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
4247                 return -ENOENT; /* -ve: OK but no match */
4248         }
4249
4250         LASSERT(rc == LNET_MATCHMD_OK);
4251
4252         lnet_build_msg_event(msg, LNET_EVENT_GET);
4253
4254         reply_wmd = hdr->msg.get.return_wmd;
4255
4256         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4257                        msg->msg_offset, msg->msg_wanted);
4258
4259         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4260
4261         if (rdma_get) {
4262                 /* The LND completes the REPLY from her recv procedure */
4263                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4264                              msg->msg_offset, msg->msg_len, msg->msg_len);
4265                 return 0;
4266         }
4267
4268         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4269         msg->msg_receiving = 0;
4270
4271         /* FIXME need to handle large-addr nid */
4272         rc = lnet_send(lnet_nid_to_nid4(&ni->ni_nid), msg, msg->msg_from);
4273         if (rc < 0) {
4274                 /* didn't get as far as lnet_ni_send() */
4275                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4276                        libcfs_nidstr(&ni->ni_nid),
4277                        libcfs_id2str(info.mi_id), rc);
4278
4279                 lnet_finalize(msg, rc);
4280         }
4281
4282         return 0;
4283 }
4284
4285 static int
4286 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4287 {
4288         void *private = msg->msg_private;
4289         struct lnet_hdr *hdr = &msg->msg_hdr;
4290         struct lnet_process_id src = {0};
4291         struct lnet_libmd *md;
4292         unsigned int rlength;
4293         unsigned int mlength;
4294         int cpt;
4295
4296         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4297         lnet_res_lock(cpt);
4298
4299         src.nid = hdr->src_nid;
4300         src.pid = hdr->src_pid;
4301
4302         /* NB handles only looked up by creator (no flips) */
4303         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4304         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4305                 CNETERR("%s: Dropping REPLY from %s for %s "
4306                         "MD %#llx.%#llx\n",
4307                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4308                         (md == NULL) ? "invalid" : "inactive",
4309                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4310                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4311                 if (md != NULL && md->md_me != NULL)
4312                         CERROR("REPLY MD also attached to portal %d\n",
4313                                md->md_me->me_portal);
4314
4315                 lnet_res_unlock(cpt);
4316                 return -ENOENT; /* -ve: OK but no match */
4317         }
4318
4319         LASSERT(md->md_offset == 0);
4320
4321         rlength = hdr->payload_length;
4322         mlength = min(rlength, md->md_length);
4323
4324         if (mlength < rlength &&
4325             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4326                 CNETERR("%s: Dropping REPLY from %s length %d "
4327                         "for MD %#llx would overflow (%d)\n",
4328                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4329                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4330                         mlength);
4331                 lnet_res_unlock(cpt);
4332                 return -ENOENT; /* -ve: OK but no match */
4333         }
4334
4335         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4336                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4337                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4338
4339         lnet_msg_attach_md(msg, md, 0, mlength);
4340
4341         if (mlength != 0)
4342                 lnet_setpayloadbuffer(msg);
4343
4344         lnet_res_unlock(cpt);
4345
4346         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4347
4348         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4349         return 0;
4350 }
4351
4352 static int
4353 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4354 {
4355         struct lnet_hdr *hdr = &msg->msg_hdr;
4356         struct lnet_process_id src = {0};
4357         struct lnet_libmd *md;
4358         int cpt;
4359
4360         src.nid = hdr->src_nid;
4361         src.pid = hdr->src_pid;
4362
4363         /* Convert ack fields to host byte order */
4364         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4365         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4366
4367         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4368         lnet_res_lock(cpt);
4369
4370         /* NB handles only looked up by creator (no flips) */
4371         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4372         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4373                 /* Don't moan; this is expected */
4374                 CDEBUG(D_NET,
4375                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4376                        libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4377                        (md == NULL) ? "invalid" : "inactive",
4378                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4379                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4380                 if (md != NULL && md->md_me != NULL)
4381                         CERROR("Source MD also attached to portal %d\n",
4382                                md->md_me->me_portal);
4383
4384                 lnet_res_unlock(cpt);
4385                 return -ENOENT;                  /* -ve! */
4386         }
4387
4388         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4389                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4390                hdr->msg.ack.dst_wmd.wh_object_cookie);
4391
4392         lnet_msg_attach_md(msg, md, 0, 0);
4393
4394         lnet_res_unlock(cpt);
4395
4396         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4397
4398         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4399         return 0;
4400 }
4401
4402 /**
4403  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4404  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4405  * \retval -ve                  error code
4406  */
4407 int
4408 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4409 {
4410         int     rc = 0;
4411
4412         if (!the_lnet.ln_routing)
4413                 return -ECANCELED;
4414
4415         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4416             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4417                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4418                         msg->msg_rx_ready_delay = 1;
4419                 } else {
4420                         lnet_net_unlock(msg->msg_rx_cpt);
4421                         rc = lnet_ni_eager_recv(ni, msg);
4422                         lnet_net_lock(msg->msg_rx_cpt);
4423                 }
4424         }
4425
4426         if (rc == 0)
4427                 rc = lnet_post_routed_recv_locked(msg, 0);
4428         return rc;
4429 }
4430
4431 int
4432 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4433 {
4434         int     rc;
4435
4436         switch (msg->msg_type) {
4437         case LNET_MSG_ACK:
4438                 rc = lnet_parse_ack(ni, msg);
4439                 break;
4440         case LNET_MSG_PUT:
4441                 rc = lnet_parse_put(ni, msg);
4442                 break;
4443         case LNET_MSG_GET:
4444                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4445                 break;
4446         case LNET_MSG_REPLY:
4447                 rc = lnet_parse_reply(ni, msg);
4448                 break;
4449         default: /* prevent an unused label if !kernel */
4450                 LASSERT(0);
4451                 return -EPROTO;
4452         }
4453
4454         LASSERT(rc == 0 || rc == -ENOENT);
4455         return rc;
4456 }
4457
4458 char *
4459 lnet_msgtyp2str (int type)
4460 {
4461         switch (type) {
4462         case LNET_MSG_ACK:
4463                 return ("ACK");
4464         case LNET_MSG_PUT:
4465                 return ("PUT");
4466         case LNET_MSG_GET:
4467                 return ("GET");
4468         case LNET_MSG_REPLY:
4469                 return ("REPLY");
4470         case LNET_MSG_HELLO:
4471                 return ("HELLO");
4472         default:
4473                 return ("<UNKNOWN>");
4474         }
4475 }
4476
4477 int
4478 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4479            void *private, int rdma_req)
4480 {
4481         struct lnet_peer_ni *lpni;
4482         struct lnet_msg *msg;
4483         __u32 payload_length;
4484         lnet_pid_t dest_pid;
4485         lnet_nid_t dest_nid;
4486         lnet_nid_t src_nid;
4487         bool push = false;
4488         int for_me;
4489         __u32 type;
4490         int rc = 0;
4491         int cpt;
4492
4493         LASSERT (!in_interrupt ());
4494
4495         type = le32_to_cpu(hdr->type);
4496         src_nid = le64_to_cpu(hdr->src_nid);
4497         dest_nid = le64_to_cpu(hdr->dest_nid);
4498         dest_pid = le32_to_cpu(hdr->dest_pid);
4499         payload_length = le32_to_cpu(hdr->payload_length);
4500
4501         /* FIXME handle large-addr nids */
4502         for_me = (lnet_nid_to_nid4(&ni->ni_nid) == dest_nid);
4503         cpt = lnet_cpt_of_nid(from_nid, ni);
4504
4505         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4506                 libcfs_nid2str(dest_nid),
4507                 libcfs_nidstr(&ni->ni_nid),
4508                 libcfs_nid2str(src_nid),
4509                 lnet_msgtyp2str(type),
4510                 (for_me) ? "for me" : "routed");
4511
4512         switch (type) {
4513         case LNET_MSG_ACK:
4514         case LNET_MSG_GET:
4515                 if (payload_length > 0) {
4516                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4517                                libcfs_nid2str(from_nid),
4518                                libcfs_nid2str(src_nid),
4519                                lnet_msgtyp2str(type), payload_length);
4520                         return -EPROTO;
4521                 }
4522                 break;
4523
4524         case LNET_MSG_PUT:
4525         case LNET_MSG_REPLY:
4526                 if (payload_length >
4527                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4528                         CERROR("%s, src %s: bad %s payload %d "
4529                                "(%d max expected)\n",
4530                                libcfs_nid2str(from_nid),
4531                                libcfs_nid2str(src_nid),
4532                                lnet_msgtyp2str(type),
4533                                payload_length,
4534                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4535                         return -EPROTO;
4536                 }
4537                 break;
4538
4539         default:
4540                 CERROR("%s, src %s: Bad message type 0x%x\n",
4541                        libcfs_nid2str(from_nid),
4542                        libcfs_nid2str(src_nid), type);
4543                 return -EPROTO;
4544         }
4545
4546         if (the_lnet.ln_routing &&
4547             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4548                 lnet_ni_lock(ni);
4549                 spin_lock(&ni->ni_net->net_lock);
4550                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4551                 spin_unlock(&ni->ni_net->net_lock);
4552                 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4553                 lnet_ni_unlock(ni);
4554         }
4555
4556         if (push)
4557                 lnet_push_update_to_peers(1);
4558
4559         /* Regard a bad destination NID as a protocol error.  Senders should
4560          * know what they're doing; if they don't they're misconfigured, buggy
4561          * or malicious so we chop them off at the knees :) */
4562
4563         if (!for_me) {
4564                 if (LNET_NIDNET(dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4565                         /* should have gone direct */
4566                         CERROR("%s, src %s: Bad dest nid %s "
4567                                "(should have been sent direct)\n",
4568                                 libcfs_nid2str(from_nid),
4569                                 libcfs_nid2str(src_nid),
4570                                 libcfs_nid2str(dest_nid));
4571                         return -EPROTO;
4572                 }
4573
4574                 if (lnet_islocalnid4(dest_nid)) {
4575                         /* dest is another local NI; sender should have used
4576                          * this node's NID on its own network */
4577                         CERROR("%s, src %s: Bad dest nid %s "
4578                                "(it's my nid but on a different network)\n",
4579                                 libcfs_nid2str(from_nid),
4580                                 libcfs_nid2str(src_nid),
4581                                 libcfs_nid2str(dest_nid));
4582                         return -EPROTO;
4583                 }
4584
4585                 if (rdma_req && type == LNET_MSG_GET) {
4586                         CERROR("%s, src %s: Bad optimized GET for %s "
4587                                "(final destination must be me)\n",
4588                                 libcfs_nid2str(from_nid),
4589                                 libcfs_nid2str(src_nid),
4590                                 libcfs_nid2str(dest_nid));
4591                         return -EPROTO;
4592                 }
4593
4594                 if (!the_lnet.ln_routing) {
4595                         CERROR("%s, src %s: Dropping message for %s "
4596                                "(routing not enabled)\n",
4597                                 libcfs_nid2str(from_nid),
4598                                 libcfs_nid2str(src_nid),
4599                                 libcfs_nid2str(dest_nid));
4600                         goto drop;
4601                 }
4602         }
4603
4604         /* Message looks OK; we're not going to return an error, so we MUST
4605          * call back lnd_recv() come what may... */
4606
4607         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4608             fail_peer(src_nid, 0)) {                    /* shall we now? */
4609                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4610                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4611                        lnet_msgtyp2str(type));
4612                 goto drop;
4613         }
4614
4615         /* FIXME need to support large-addr nid */
4616         if (!list_empty(&the_lnet.ln_drop_rules) &&
4617             lnet_drop_rule_match(hdr, lnet_nid_to_nid4(&ni->ni_nid), NULL)) {
4618                 CDEBUG(D_NET,
4619                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4620                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4621                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4622                 goto drop;
4623         }
4624
4625         msg = lnet_msg_alloc();
4626         if (msg == NULL) {
4627                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4628                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4629                        lnet_msgtyp2str(type));
4630                 goto drop;
4631         }
4632
4633         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4634          * pointers NULL etc */
4635
4636         msg->msg_type = type;
4637         msg->msg_private = private;
4638         msg->msg_receiving = 1;
4639         msg->msg_rdma_get = rdma_req;
4640         msg->msg_len = msg->msg_wanted = payload_length;
4641         msg->msg_offset = 0;
4642         msg->msg_hdr = *hdr;
4643         /* for building message event */
4644         msg->msg_from = from_nid;
4645         if (!for_me) {
4646                 msg->msg_target.pid     = dest_pid;
4647                 msg->msg_target.nid     = dest_nid;
4648                 msg->msg_routing        = 1;
4649
4650         } else {
4651                 /* convert common msg->hdr fields to host byteorder */
4652                 msg->msg_hdr.type       = type;
4653                 msg->msg_hdr.src_nid    = src_nid;
4654                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4655                 msg->msg_hdr.dest_nid   = dest_nid;
4656                 msg->msg_hdr.dest_pid   = dest_pid;
4657                 msg->msg_hdr.payload_length = payload_length;
4658         }
4659
4660         lnet_net_lock(cpt);
4661         /* FIXME support large-addr nid */
4662         lpni = lnet_nid2peerni_locked(from_nid, lnet_nid_to_nid4(&ni->ni_nid),
4663                                       cpt);
4664         if (IS_ERR(lpni)) {
4665                 lnet_net_unlock(cpt);
4666                 rc = PTR_ERR(lpni);
4667                 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4668                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4669                        lnet_msgtyp2str(type), rc);
4670                 lnet_msg_free(msg);
4671                 if (rc == -ESHUTDOWN)
4672                         /* We are shutting down.  Don't do anything more */
4673                         return 0;
4674                 goto drop;
4675         }
4676
4677         /* If this message was forwarded to us from a router then we may need
4678          * to update router aliveness or check for an asymmetrical route
4679          * (or both)
4680          */
4681         if (((lnet_drop_asym_route && for_me) ||
4682              !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4683             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4684                 __u32 src_net_id = LNET_NIDNET(src_nid);
4685                 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4686                 struct lnet_route *route;
4687                 bool found = false;
4688
4689                 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4690                         if (route->lr_net == src_net_id) {
4691                                 found = true;
4692                                 /* If we're transitioning the gateway from
4693                                  * dead -> alive, and discovery is disabled
4694                                  * locally or on the gateway, then we need to
4695                                  * update the cached route aliveness for each
4696                                  * route to the src_nid's net.
4697                                  *
4698                                  * Otherwise, we're only checking for
4699                                  * symmetrical route, and we can break the
4700                                  * loop
4701                                  */
4702                                 if (!gw->lp_alive &&
4703                                     lnet_is_discovery_disabled(gw))
4704                                         lnet_set_route_aliveness(route, true);
4705                                 else
4706                                         break;
4707                         }
4708                 }
4709                 if (lnet_drop_asym_route && for_me && !found) {
4710                         /* Drop ref taken by lnet_nid2peerni_locked() */
4711                         lnet_peer_ni_decref_locked(lpni);
4712                         lnet_net_unlock(cpt);
4713                         /* we would not use from_nid to route a message to
4714                          * src_nid
4715                          * => asymmetric routing detected but forbidden
4716                          */
4717                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4718                                libcfs_nid2str(from_nid),
4719                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4720                         lnet_msg_free(msg);
4721                         goto drop;
4722                 }
4723                 if (!gw->lp_alive) {
4724                         struct lnet_peer_net *lpn;
4725                         struct lnet_peer_ni *lpni2;
4726
4727                         gw->lp_alive = true;
4728                         /* Mark all remote NIs on src_nid's net UP */
4729                         lpn = lnet_peer_get_net_locked(gw, src_net_id);
4730                         if (lpn)
4731                                 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4732                                                     lpni_peer_nis)
4733                                         lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4734                 }
4735         }
4736
4737         lpni->lpni_last_alive = ktime_get_seconds();
4738
4739         msg->msg_rxpeer = lpni;
4740         msg->msg_rxni = ni;
4741         lnet_ni_addref_locked(ni, cpt);
4742         /* Multi-Rail: Primary NID of source. */
4743         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4744
4745         /*
4746          * mark the status of this lpni as UP since we received a message
4747          * from it. The ping response reports back the ns_status which is
4748          * marked on the remote as up or down and we cache it here.
4749          */
4750         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4751
4752         lnet_msg_commit(msg, cpt);
4753
4754         /* message delay simulation */
4755         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4756                      lnet_delay_rule_match_locked(hdr, msg))) {
4757                 lnet_net_unlock(cpt);
4758                 return 0;
4759         }
4760
4761         if (!for_me) {
4762                 rc = lnet_parse_forward_locked(ni, msg);
4763                 lnet_net_unlock(cpt);
4764
4765                 if (rc < 0)
4766                         goto free_drop;
4767
4768                 if (rc == LNET_CREDIT_OK) {
4769                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4770                                      0, payload_length, payload_length);
4771                 }
4772                 return 0;
4773         }
4774
4775         lnet_net_unlock(cpt);
4776
4777         rc = lnet_parse_local(ni, msg);
4778         if (rc != 0)
4779                 goto free_drop;
4780         return 0;
4781
4782  free_drop:
4783         LASSERT(msg->msg_md == NULL);
4784         lnet_finalize(msg, rc);
4785
4786  drop:
4787         lnet_drop_message(ni, cpt, private, payload_length, type);
4788         return 0;
4789 }
4790 EXPORT_SYMBOL(lnet_parse);
4791
4792 void
4793 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4794 {
4795         while (!list_empty(head)) {
4796                 struct lnet_process_id id = {0};
4797                 struct lnet_msg *msg;
4798
4799                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4800                 list_del(&msg->msg_list);
4801
4802                 id.nid = msg->msg_hdr.src_nid;
4803                 id.pid = msg->msg_hdr.src_pid;
4804
4805                 LASSERT(msg->msg_md == NULL);
4806                 LASSERT(msg->msg_rx_delayed);
4807                 LASSERT(msg->msg_rxpeer != NULL);
4808                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4809
4810                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4811                       " offset %d length %d: %s\n",
4812                       libcfs_id2str(id),
4813                       msg->msg_hdr.msg.put.ptl_index,
4814                       msg->msg_hdr.msg.put.match_bits,
4815                       msg->msg_hdr.msg.put.offset,
4816                       msg->msg_hdr.payload_length, reason);
4817
4818                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4819                  * called lnet_drop_message(), so I just hang onto msg as well
4820                  * until that's done */
4821
4822                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4823                                   msg->msg_private, msg->msg_len,
4824                                   msg->msg_type);
4825
4826                 msg->msg_no_resend = true;
4827                 /*
4828                  * NB: message will not generate event because w/o attached MD,
4829                  * but we still should give error code so lnet_msg_decommit()
4830                  * can skip counters operations and other checks.
4831                  */
4832                 lnet_finalize(msg, -ENOENT);
4833         }
4834 }
4835
4836 void
4837 lnet_recv_delayed_msg_list(struct list_head *head)
4838 {
4839         while (!list_empty(head)) {
4840                 struct lnet_msg *msg;
4841                 struct lnet_process_id id;
4842
4843                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4844                 list_del(&msg->msg_list);
4845
4846                 /* md won't disappear under me, since each msg
4847                  * holds a ref on it */
4848
4849                 id.nid = msg->msg_hdr.src_nid;
4850                 id.pid = msg->msg_hdr.src_pid;
4851
4852                 LASSERT(msg->msg_rx_delayed);
4853                 LASSERT(msg->msg_md != NULL);
4854                 LASSERT(msg->msg_rxpeer != NULL);
4855                 LASSERT(msg->msg_rxni != NULL);
4856                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4857
4858                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4859                        "match %llu offset %d length %d.\n",
4860                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4861                         msg->msg_hdr.msg.put.match_bits,
4862                         msg->msg_hdr.msg.put.offset,
4863                         msg->msg_hdr.payload_length);
4864
4865                 lnet_recv_put(msg->msg_rxni, msg);
4866         }
4867 }
4868
4869 static void
4870 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4871                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4872 {
4873         s64 timeout_ns;
4874         struct lnet_rsp_tracker *local_rspt;
4875
4876         /*
4877          * MD has a refcount taken by message so it's not going away.
4878          * The MD however can be looked up. We need to secure the access
4879          * to the md_rspt_ptr by taking the res_lock.
4880          * The rspt can be accessed without protection up to when it gets
4881          * added to the list.
4882          */
4883
4884         lnet_res_lock(cpt);
4885         local_rspt = md->md_rspt_ptr;
4886         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4887         if (local_rspt != NULL) {
4888                 /*
4889                  * we already have an rspt attached to the md, so we'll
4890                  * update the deadline on that one.
4891                  */
4892                 lnet_rspt_free(rspt, cpt);
4893         } else {
4894                 /* new md */
4895                 rspt->rspt_mdh = mdh;
4896                 rspt->rspt_cpt = cpt;
4897                 /* store the rspt so we can access it when we get the REPLY */
4898                 md->md_rspt_ptr = rspt;
4899                 local_rspt = rspt;
4900         }
4901         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4902
4903         /*
4904          * add to the list of tracked responses. It's added to tail of the
4905          * list in order to expire all the older entries first.
4906          */
4907         lnet_net_lock(cpt);
4908         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4909         lnet_net_unlock(cpt);
4910         lnet_res_unlock(cpt);
4911 }
4912
4913 /**
4914  * Initiate an asynchronous PUT operation.
4915  *
4916  * There are several events associated with a PUT: completion of the send on
4917  * the initiator node (LNET_EVENT_SEND), and when the send completes
4918  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4919  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4920  * used at the target node to indicate the completion of incoming data
4921  * delivery.
4922  *
4923  * The local events will be logged in the EQ associated with the MD pointed to
4924  * by \a mdh handle. Using a MD without an associated EQ results in these
4925  * events being discarded. In this case, the caller must have another
4926  * mechanism (e.g., a higher level protocol) for determining when it is safe
4927  * to modify the memory region associated with the MD.
4928  *
4929  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4930  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4931  *
4932  * \param self Indicates the NID of a local interface through which to send
4933  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4934  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4935  * must be "free floating" (See LNetMDBind()).
4936  * \param ack Controls whether an acknowledgment is requested.
4937  * Acknowledgments are only sent when they are requested by the initiating
4938  * process and the target MD enables them.
4939  * \param target A process identifier for the target process.
4940  * \param portal The index in the \a target's portal table.
4941  * \param match_bits The match bits to use for MD selection at the target
4942  * process.
4943  * \param offset The offset into the target MD (only used when the target
4944  * MD has the LNET_MD_MANAGE_REMOTE option set).
4945  * \param hdr_data 64 bits of user data that can be included in the message
4946  * header. This data is written to an event queue entry at the target if an
4947  * EQ is present on the matching MD.
4948  *
4949  * \retval  0      Success, and only in this case events will be generated
4950  * and logged to EQ (if it exists).
4951  * \retval -EIO    Simulated failure.
4952  * \retval -ENOMEM Memory allocation failure.
4953  * \retval -ENOENT Invalid MD object.
4954  *
4955  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4956  */
4957 int
4958 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4959         struct lnet_process_id target, unsigned int portal,
4960         __u64 match_bits, unsigned int offset,
4961         __u64 hdr_data)
4962 {
4963         struct lnet_msg *msg;
4964         struct lnet_libmd *md;
4965         int cpt;
4966         int rc;
4967         struct lnet_rsp_tracker *rspt = NULL;
4968
4969         LASSERT(the_lnet.ln_refcount > 0);
4970
4971         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4972             fail_peer(target.nid, 1)) {                 /* shall we now? */
4973                 CERROR("Dropping PUT to %s: simulated failure\n",
4974                        libcfs_id2str(target));
4975                 return -EIO;
4976         }
4977
4978         msg = lnet_msg_alloc();
4979         if (msg == NULL) {
4980                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4981                        libcfs_id2str(target));
4982                 return -ENOMEM;
4983         }
4984         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4985
4986         cpt = lnet_cpt_of_cookie(mdh.cookie);
4987
4988         if (ack == LNET_ACK_REQ) {
4989                 rspt = lnet_rspt_alloc(cpt);
4990                 if (!rspt) {
4991                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4992                                 libcfs_id2str(target));
4993                         return -ENOMEM;
4994                 }
4995                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4996         }
4997
4998         lnet_res_lock(cpt);
4999
5000         md = lnet_handle2md(&mdh);
5001         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5002                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5003                        match_bits, portal, libcfs_id2str(target),
5004                        md == NULL ? -1 : md->md_threshold);
5005                 if (md != NULL && md->md_me != NULL)
5006                         CERROR("Source MD also attached to portal %d\n",
5007                                md->md_me->me_portal);
5008                 lnet_res_unlock(cpt);
5009
5010                 if (rspt)
5011                         lnet_rspt_free(rspt, cpt);
5012
5013                 lnet_msg_free(msg);
5014                 return -ENOENT;
5015         }
5016
5017         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
5018
5019         lnet_msg_attach_md(msg, md, 0, 0);
5020
5021         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5022
5023         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5024         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5025         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5026         msg->msg_hdr.msg.put.hdr_data = hdr_data;
5027
5028         /* NB handles only looked up by creator (no flips) */
5029         if (ack == LNET_ACK_REQ) {
5030                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5031                         the_lnet.ln_interface_cookie;
5032                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5033                         md->md_lh.lh_cookie;
5034         } else {
5035                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5036                         LNET_WIRE_HANDLE_COOKIE_NONE;
5037                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5038                         LNET_WIRE_HANDLE_COOKIE_NONE;
5039         }
5040
5041         lnet_res_unlock(cpt);
5042
5043         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5044
5045         if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5046                                                    md->md_options))
5047                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5048         else if (rspt)
5049                 lnet_rspt_free(rspt, cpt);
5050
5051         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5052                                  CFS_FAIL_ONCE))
5053                 rc = -EIO;
5054         else
5055                 rc = lnet_send(self, msg, LNET_NID_ANY);
5056
5057         if (rc != 0) {
5058                 CNETERR("Error sending PUT to %s: %d\n",
5059                         libcfs_id2str(target), rc);
5060                 msg->msg_no_resend = true;
5061                 lnet_finalize(msg, rc);
5062         }
5063
5064         /* completion will be signalled by an event */
5065         return 0;
5066 }
5067 EXPORT_SYMBOL(LNetPut);
5068
5069 /*
5070  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
5071  * returns a msg for the LND to pass to lnet_finalize() when the sink
5072  * data has been received.
5073  *
5074  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5075  * lnet_finalize() is called on it, so the LND must call this first
5076  */
5077 struct lnet_msg *
5078 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5079 {
5080         struct lnet_msg *msg = lnet_msg_alloc();
5081         struct lnet_libmd *getmd = getmsg->msg_md;
5082         struct lnet_process_id peer_id = getmsg->msg_target;
5083         int cpt;
5084
5085         LASSERT(!getmsg->msg_target_is_router);
5086         LASSERT(!getmsg->msg_routing);
5087
5088         if (msg == NULL) {
5089                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5090                        libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id));
5091                 goto drop;
5092         }
5093
5094         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5095         lnet_res_lock(cpt);
5096
5097         LASSERT(getmd->md_refcount > 0);
5098
5099         if (getmd->md_threshold == 0) {
5100                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5101                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id),
5102                         getmd);
5103                 lnet_res_unlock(cpt);
5104                 goto drop;
5105         }
5106
5107         LASSERT(getmd->md_offset == 0);
5108
5109         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5110                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id), getmd);
5111
5112         /* setup information for lnet_build_msg_event */
5113         msg->msg_initiator =
5114                 lnet_nid_to_nid4(&getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid);
5115         msg->msg_from = peer_id.nid;
5116         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5117         msg->msg_hdr.src_nid = peer_id.nid;
5118         msg->msg_hdr.payload_length = getmd->md_length;
5119         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5120
5121         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5122         lnet_res_unlock(cpt);
5123
5124         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5125
5126         lnet_net_lock(cpt);
5127         lnet_msg_commit(msg, cpt);
5128         lnet_net_unlock(cpt);
5129
5130         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5131
5132         return msg;
5133
5134  drop:
5135         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5136
5137         lnet_net_lock(cpt);
5138         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5139         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5140         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5141                 getmd->md_length;
5142         lnet_net_unlock(cpt);
5143
5144         if (msg != NULL)
5145                 lnet_msg_free(msg);
5146
5147         return NULL;
5148 }
5149 EXPORT_SYMBOL(lnet_create_reply_msg);
5150
5151 void
5152 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5153                        unsigned int len)
5154 {
5155         /* Set the REPLY length, now the RDMA that elides the REPLY message has
5156          * completed and I know it. */
5157         LASSERT(reply != NULL);
5158         LASSERT(reply->msg_type == LNET_MSG_GET);
5159         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5160
5161         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
5162          * the end of my buffer, I might as well be dead. */
5163         LASSERT(len <= reply->msg_ev.mlength);
5164
5165         reply->msg_ev.mlength = len;
5166 }
5167 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5168
5169 /**
5170  * Initiate an asynchronous GET operation.
5171  *
5172  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5173  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5174  * the target node in the REPLY has been written to local MD.
5175  *
5176  * On the target node, an LNET_EVENT_GET is logged when the GET request
5177  * arrives and is accepted into a MD.
5178  *
5179  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5180  * \param mdh A handle for the MD that describes the memory into which the
5181  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5182  *
5183  * \retval  0      Success, and only in this case events will be generated
5184  * and logged to EQ (if it exists) of the MD.
5185  * \retval -EIO    Simulated failure.
5186  * \retval -ENOMEM Memory allocation failure.
5187  * \retval -ENOENT Invalid MD object.
5188  */
5189 int
5190 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
5191         struct lnet_process_id target, unsigned int portal,
5192         __u64 match_bits, unsigned int offset, bool recovery)
5193 {
5194         struct lnet_msg *msg;
5195         struct lnet_libmd *md;
5196         struct lnet_rsp_tracker *rspt;
5197         int cpt;
5198         int rc;
5199
5200         LASSERT(the_lnet.ln_refcount > 0);
5201
5202         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
5203             fail_peer(target.nid, 1))                   /* shall we now? */
5204         {
5205                 CERROR("Dropping GET to %s: simulated failure\n",
5206                        libcfs_id2str(target));
5207                 return -EIO;
5208         }
5209
5210         msg = lnet_msg_alloc();
5211         if (!msg) {
5212                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5213                        libcfs_id2str(target));
5214                 return -ENOMEM;
5215         }
5216
5217         cpt = lnet_cpt_of_cookie(mdh.cookie);
5218
5219         rspt = lnet_rspt_alloc(cpt);
5220         if (!rspt) {
5221                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5222                        libcfs_id2str(target));
5223                 return -ENOMEM;
5224         }
5225         INIT_LIST_HEAD(&rspt->rspt_on_list);
5226
5227         msg->msg_recovery = recovery;
5228
5229         lnet_res_lock(cpt);
5230
5231         md = lnet_handle2md(&mdh);
5232         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5233                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5234                        match_bits, portal, libcfs_id2str(target),
5235                        md == NULL ? -1 : md->md_threshold);
5236                 if (md != NULL && md->md_me != NULL)
5237                         CERROR("REPLY MD also attached to portal %d\n",
5238                                md->md_me->me_portal);
5239
5240                 lnet_res_unlock(cpt);
5241
5242                 lnet_msg_free(msg);
5243                 lnet_rspt_free(rspt, cpt);
5244                 return -ENOENT;
5245         }
5246
5247         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5248
5249         lnet_msg_attach_md(msg, md, 0, 0);
5250
5251         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5252
5253         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5254         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5255         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5256         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5257
5258         /* NB handles only looked up by creator (no flips) */
5259         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5260                 the_lnet.ln_interface_cookie;
5261         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5262                 md->md_lh.lh_cookie;
5263
5264         lnet_res_unlock(cpt);
5265
5266         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5267
5268         if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5269                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5270         else
5271                 lnet_rspt_free(rspt, cpt);
5272
5273         rc = lnet_send(self, msg, LNET_NID_ANY);
5274         if (rc < 0) {
5275                 CNETERR("Error sending GET to %s: %d\n",
5276                         libcfs_id2str(target), rc);
5277                 msg->msg_no_resend = true;
5278                 lnet_finalize(msg, rc);
5279         }
5280
5281         /* completion will be signalled by an event */
5282         return 0;
5283 }
5284 EXPORT_SYMBOL(LNetGet);
5285
5286 /**
5287  * Calculate distance to node at \a dstnid.
5288  *
5289  * \param dstnid Target NID.
5290  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5291  * is saved here.
5292  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5293  * here.
5294  *
5295  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5296  * local_nid_dist_zero is set, which is the default.
5297  * \retval positives Distance to target NID, i.e. number of hops plus one.
5298  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5299  */
5300 int
5301 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5302 {
5303         struct list_head *e;
5304         struct lnet_ni *ni = NULL;
5305         struct lnet_remotenet *rnet;
5306         __u32 dstnet = LNET_NIDNET(dstnid);
5307         int hops;
5308         int cpt;
5309         __u32 order = 2;
5310         struct list_head *rn_list;
5311         bool matched_dstnet = false;
5312
5313         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5314          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5315          * keep order 0 free for 0@lo and order 1 free for a local NID
5316          * match */
5317
5318         LASSERT(the_lnet.ln_refcount > 0);
5319
5320         cpt = lnet_net_lock_current();
5321
5322         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5323                 /* FIXME support large-addr nid */
5324                 if (lnet_nid_to_nid4(&ni->ni_nid) == dstnid) {
5325                         if (srcnidp != NULL)
5326                                 *srcnidp = dstnid;
5327                         if (orderp != NULL) {
5328                                 if (dstnid == LNET_NID_LO_0)
5329                                         *orderp = 0;
5330                                 else
5331                                         *orderp = 1;
5332                         }
5333                         lnet_net_unlock(cpt);
5334
5335                         return local_nid_dist_zero ? 0 : 1;
5336                 }
5337
5338                 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5339                         matched_dstnet = true;
5340                         /* We matched the destination net, but we may have
5341                          * additional local NIs to inspect.
5342                          *
5343                          * We record the nid and order as appropriate, but
5344                          * they may be overwritten if we match local NI above.
5345                          */
5346                         if (srcnidp)
5347                                 /* FIXME support large-addr nids */
5348                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5349
5350                         if (orderp) {
5351                                 /* Check if ni was originally created in
5352                                  * current net namespace.
5353                                  * If not, assign order above 0xffff0000,
5354                                  * to make this ni not a priority.
5355                                  */
5356                                 if (current->nsproxy &&
5357                                     !net_eq(ni->ni_net_ns,
5358                                             current->nsproxy->net_ns))
5359                                         *orderp = order + 0xffff0000;
5360                                 else
5361                                         *orderp = order;
5362                         }
5363                 }
5364
5365                 order++;
5366         }
5367
5368         if (matched_dstnet) {
5369                 lnet_net_unlock(cpt);
5370                 return 1;
5371         }
5372
5373         rn_list = lnet_net2rnethash(dstnet);
5374         list_for_each(e, rn_list) {
5375                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5376
5377                 if (rnet->lrn_net == dstnet) {
5378                         struct lnet_route *route;
5379                         struct lnet_route *shortest = NULL;
5380                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5381                         __u32 route_hops;
5382
5383                         LASSERT(!list_empty(&rnet->lrn_routes));
5384
5385                         list_for_each_entry(route, &rnet->lrn_routes,
5386                                             lr_list) {
5387                                 route_hops = route->lr_hops;
5388                                 if (route_hops == LNET_UNDEFINED_HOPS)
5389                                         route_hops = 1;
5390                                 if (shortest == NULL ||
5391                                     route_hops < shortest_hops) {
5392                                         shortest = route;
5393                                         shortest_hops = route_hops;
5394                                 }
5395                         }
5396
5397                         LASSERT(shortest != NULL);
5398                         hops = shortest_hops;
5399                         if (srcnidp != NULL) {
5400                                 struct lnet_net *net;
5401                                 net = lnet_get_net_locked(shortest->lr_lnet);
5402                                 LASSERT(net);
5403                                 ni = lnet_get_next_ni_locked(net, NULL);
5404                                 /* FIXME support large-addr nids */
5405                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5406                         }
5407                         if (orderp != NULL)
5408                                 *orderp = order;
5409                         lnet_net_unlock(cpt);
5410                         return hops + 1;
5411                 }
5412                 order++;
5413         }
5414
5415         lnet_net_unlock(cpt);
5416         return -EHOSTUNREACH;
5417 }
5418 EXPORT_SYMBOL(LNetDist);