lnet/lnet/lib-move.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-move.c
  33  *
  34  * Data movement routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <linux/pagemap.h>
  40
  41 #include <lnet/lib-lnet.h>
  42 #include <linux/nsproxy.h>
  43 #include <net/net_namespace.h>
  44
  45 extern unsigned int lnet_current_net_count;
  46
  47 static int local_nid_dist_zero = 1;
  48 module_param(local_nid_dist_zero, int, 0444);
  49 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  50
  51 struct lnet_send_data {
  52         struct lnet_ni *sd_best_ni;
  53         struct lnet_peer_ni *sd_best_lpni;
  54         struct lnet_peer_ni *sd_final_dst_lpni;
  55         struct lnet_peer *sd_peer;
  56         struct lnet_peer *sd_gw_peer;
  57         struct lnet_peer_ni *sd_gw_lpni;
  58         struct lnet_peer_net *sd_peer_net;
  59         struct lnet_msg *sd_msg;
  60         lnet_nid_t sd_dst_nid;
  61         lnet_nid_t sd_src_nid;
  62         lnet_nid_t sd_rtr_nid;
  63         int sd_cpt;
  64         int sd_md_cpt;
  65         __u32 sd_send_case;
  66 };
  67
  68 static inline struct lnet_comm_count *
  69 get_stats_counts(struct lnet_element_stats *stats,
  70                  enum lnet_stats_type stats_type)
  71 {
  72         switch (stats_type) {
  73         case LNET_STATS_TYPE_SEND:
  74                 return &stats->el_send_stats;
  75         case LNET_STATS_TYPE_RECV:
  76                 return &stats->el_recv_stats;
  77         case LNET_STATS_TYPE_DROP:
  78                 return &stats->el_drop_stats;
  79         default:
  80                 CERROR("Unknown stats type\n");
  81         }
  82
  83         return NULL;
  84 }
  85
  86 void lnet_incr_stats(struct lnet_element_stats *stats,
  87                      enum lnet_msg_type msg_type,
  88                      enum lnet_stats_type stats_type)
  89 {
  90         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
  91         if (!counts)
  92                 return;
  93
  94         switch (msg_type) {
  95         case LNET_MSG_ACK:
  96                 atomic_inc(&counts->co_ack_count);
  97                 break;
  98         case LNET_MSG_PUT:
  99                 atomic_inc(&counts->co_put_count);
 100                 break;
 101         case LNET_MSG_GET:
 102                 atomic_inc(&counts->co_get_count);
 103                 break;
 104         case LNET_MSG_REPLY:
 105                 atomic_inc(&counts->co_reply_count);
 106                 break;
 107         case LNET_MSG_HELLO:
 108                 atomic_inc(&counts->co_hello_count);
 109                 break;
 110         default:
 111                 CERROR("There is a BUG in the code. Unknown message type\n");
 112                 break;
 113         }
 114 }
 115
 116 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 117                      enum lnet_stats_type stats_type)
 118 {
 119         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 120         if (!counts)
 121                 return 0;
 122
 123         return (atomic_read(&counts->co_ack_count) +
 124                 atomic_read(&counts->co_put_count) +
 125                 atomic_read(&counts->co_get_count) +
 126                 atomic_read(&counts->co_reply_count) +
 127                 atomic_read(&counts->co_hello_count));
 128 }
 129
 130 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 131                                 struct lnet_comm_count *counts)
 132 {
 133         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 134         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 135         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 136         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 137         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 138 }
 139
 140 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 141                               struct lnet_element_stats *stats)
 142 {
 143         struct lnet_comm_count *counts;
 144
 145         LASSERT(msg_stats);
 146         LASSERT(stats);
 147
 148         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 149         if (!counts)
 150                 return;
 151         assign_stats(&msg_stats->im_send_stats, counts);
 152
 153         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 154         if (!counts)
 155                 return;
 156         assign_stats(&msg_stats->im_recv_stats, counts);
 157
 158         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 159         if (!counts)
 160                 return;
 161         assign_stats(&msg_stats->im_drop_stats, counts);
 162 }
 163
 164 int
 165 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 166 {
 167         struct lnet_test_peer *tp;
 168         struct list_head *el;
 169         struct list_head *next;
 170         struct list_head  cull;
 171
 172         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 173         if (threshold != 0) {
 174                 /* Adding a new entry */
 175                 LIBCFS_ALLOC(tp, sizeof(*tp));
 176                 if (tp == NULL)
 177                         return -ENOMEM;
 178
 179                 tp->tp_nid = nid;
 180                 tp->tp_threshold = threshold;
 181
 182                 lnet_net_lock(0);
 183                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 184                 lnet_net_unlock(0);
 185                 return 0;
 186         }
 187
 188         /* removing entries */
 189         INIT_LIST_HEAD(&cull);
 190
 191         lnet_net_lock(0);
 192
 193         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 194                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 195
 196                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 197                     nid == LNET_NID_ANY ||      /* removing all entries */
 198                     tp->tp_nid == nid) {        /* matched this one */
 199                         list_del(&tp->tp_list);
 200                         list_add(&tp->tp_list, &cull);
 201                 }
 202         }
 203
 204         lnet_net_unlock(0);
 205
 206         while (!list_empty(&cull)) {
 207                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 208
 209                 list_del(&tp->tp_list);
 210                 LIBCFS_FREE(tp, sizeof(*tp));
 211         }
 212         return 0;
 213 }
 214
 215 static int
 216 fail_peer (lnet_nid_t nid, int outgoing)
 217 {
 218         struct lnet_test_peer *tp;
 219         struct list_head *el;
 220         struct list_head *next;
 221         struct list_head  cull;
 222         int               fail = 0;
 223
 224         INIT_LIST_HEAD(&cull);
 225
 226         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 227         lnet_net_lock(0);
 228
 229         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 230                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 231
 232                 if (tp->tp_threshold == 0) {
 233                         /* zombie entry */
 234                         if (outgoing) {
 235                                 /* only cull zombies on outgoing tests,
 236                                  * since we may be at interrupt priority on
 237                                  * incoming messages. */
 238                                 list_del(&tp->tp_list);
 239                                 list_add(&tp->tp_list, &cull);
 240                         }
 241                         continue;
 242                 }
 243
 244                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 245                     nid == tp->tp_nid) {                /* fail this peer */
 246                         fail = 1;
 247
 248                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 249                                 tp->tp_threshold--;
 250                                 if (outgoing &&
 251                                     tp->tp_threshold == 0) {
 252                                         /* see above */
 253                                         list_del(&tp->tp_list);
 254                                         list_add(&tp->tp_list, &cull);
 255                                 }
 256                         }
 257                         break;
 258                 }
 259         }
 260
 261         lnet_net_unlock(0);
 262
 263         while (!list_empty(&cull)) {
 264                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 265                 list_del(&tp->tp_list);
 266
 267                 LIBCFS_FREE(tp, sizeof(*tp));
 268         }
 269
 270         return fail;
 271 }
 272
 273 unsigned int
 274 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 275 {
 276         unsigned int nob = 0;
 277
 278         LASSERT(niov == 0 || iov != NULL);
 279         while (niov-- > 0)
 280                 nob += (iov++)->iov_len;
 281
 282         return (nob);
 283 }
 284 EXPORT_SYMBOL(lnet_iov_nob);
 285
 286 void
 287 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 288                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 289                   unsigned int nob)
 290 {
 291         /* NB diov, siov are READ-ONLY */
 292         unsigned int  this_nob;
 293
 294         if (nob == 0)
 295                 return;
 296
 297         /* skip complete frags before 'doffset' */
 298         LASSERT(ndiov > 0);
 299         while (doffset >= diov->iov_len) {
 300                 doffset -= diov->iov_len;
 301                 diov++;
 302                 ndiov--;
 303                 LASSERT(ndiov > 0);
 304         }
 305
 306         /* skip complete frags before 'soffset' */
 307         LASSERT(nsiov > 0);
 308         while (soffset >= siov->iov_len) {
 309                 soffset -= siov->iov_len;
 310                 siov++;
 311                 nsiov--;
 312                 LASSERT(nsiov > 0);
 313         }
 314
 315         do {
 316                 LASSERT(ndiov > 0);
 317                 LASSERT(nsiov > 0);
 318                 this_nob = MIN(diov->iov_len - doffset,
 319                                siov->iov_len - soffset);
 320                 this_nob = MIN(this_nob, nob);
 321
 322                 memcpy((char *)diov->iov_base + doffset,
 323                        (char *)siov->iov_base + soffset, this_nob);
 324                 nob -= this_nob;
 325
 326                 if (diov->iov_len > doffset + this_nob) {
 327                         doffset += this_nob;
 328                 } else {
 329                         diov++;
 330                         ndiov--;
 331                         doffset = 0;
 332                 }
 333
 334                 if (siov->iov_len > soffset + this_nob) {
 335                         soffset += this_nob;
 336                 } else {
 337                         siov++;
 338                         nsiov--;
 339                         soffset = 0;
 340                 }
 341         } while (nob > 0);
 342 }
 343 EXPORT_SYMBOL(lnet_copy_iov2iov);
 344
 345 int
 346 lnet_extract_iov(int dst_niov, struct kvec *dst,
 347                  int src_niov, struct kvec *src,
 348                  unsigned int offset, unsigned int len)
 349 {
 350         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 351          * for exactly 'len' bytes, and return the number of entries.
 352          * NB not destructive to 'src' */
 353         unsigned int    frag_len;
 354         unsigned int    niov;
 355
 356         if (len == 0)                           /* no data => */
 357                 return (0);                     /* no frags */
 358
 359         LASSERT(src_niov > 0);
 360         while (offset >= src->iov_len) {      /* skip initial frags */
 361                 offset -= src->iov_len;
 362                 src_niov--;
 363                 src++;
 364                 LASSERT(src_niov > 0);
 365         }
 366
 367         niov = 1;
 368         for (;;) {
 369                 LASSERT(src_niov > 0);
 370                 LASSERT((int)niov <= dst_niov);
 371
 372                 frag_len = src->iov_len - offset;
 373                 dst->iov_base = ((char *)src->iov_base) + offset;
 374
 375                 if (len <= frag_len) {
 376                         dst->iov_len = len;
 377                         return (niov);
 378                 }
 379
 380                 dst->iov_len = frag_len;
 381
 382                 len -= frag_len;
 383                 dst++;
 384                 src++;
 385                 niov++;
 386                 src_niov--;
 387                 offset = 0;
 388         }
 389 }
 390 EXPORT_SYMBOL(lnet_extract_iov);
 391
 392
 393 unsigned int
 394 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 395 {
 396         unsigned int  nob = 0;
 397
 398         LASSERT(niov == 0 || kiov != NULL);
 399         while (niov-- > 0)
 400                 nob += (kiov++)->kiov_len;
 401
 402         return (nob);
 403 }
 404 EXPORT_SYMBOL(lnet_kiov_nob);
 405
 406 void
 407 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 408                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 409                     unsigned int nob)
 410 {
 411         /* NB diov, siov are READ-ONLY */
 412         unsigned int    this_nob;
 413         char           *daddr = NULL;
 414         char           *saddr = NULL;
 415
 416         if (nob == 0)
 417                 return;
 418
 419         LASSERT (!in_interrupt ());
 420
 421         LASSERT (ndiov > 0);
 422         while (doffset >= diov->kiov_len) {
 423                 doffset -= diov->kiov_len;
 424                 diov++;
 425                 ndiov--;
 426                 LASSERT(ndiov > 0);
 427         }
 428
 429         LASSERT(nsiov > 0);
 430         while (soffset >= siov->kiov_len) {
 431                 soffset -= siov->kiov_len;
 432                 siov++;
 433                 nsiov--;
 434                 LASSERT(nsiov > 0);
 435         }
 436
 437         do {
 438                 LASSERT(ndiov > 0);
 439                 LASSERT(nsiov > 0);
 440                 this_nob = MIN(diov->kiov_len - doffset,
 441                                siov->kiov_len - soffset);
 442                 this_nob = MIN(this_nob, nob);
 443
 444                 if (daddr == NULL)
 445                         daddr = ((char *)kmap(diov->kiov_page)) +
 446                                 diov->kiov_offset + doffset;
 447                 if (saddr == NULL)
 448                         saddr = ((char *)kmap(siov->kiov_page)) +
 449                                 siov->kiov_offset + soffset;
 450
 451                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 452                  * However in practice at least one of the kiovs will be mapped
 453                  * kernel pages and the map/unmap will be NOOPs */
 454
 455                 memcpy (daddr, saddr, this_nob);
 456                 nob -= this_nob;
 457
 458                 if (diov->kiov_len > doffset + this_nob) {
 459                         daddr += this_nob;
 460                         doffset += this_nob;
 461                 } else {
 462                         kunmap(diov->kiov_page);
 463                         daddr = NULL;
 464                         diov++;
 465                         ndiov--;
 466                         doffset = 0;
 467                 }
 468
 469                 if (siov->kiov_len > soffset + this_nob) {
 470                         saddr += this_nob;
 471                         soffset += this_nob;
 472                 } else {
 473                         kunmap(siov->kiov_page);
 474                         saddr = NULL;
 475                         siov++;
 476                         nsiov--;
 477                         soffset = 0;
 478                 }
 479         } while (nob > 0);
 480
 481         if (daddr != NULL)
 482                 kunmap(diov->kiov_page);
 483         if (saddr != NULL)
 484                 kunmap(siov->kiov_page);
 485 }
 486 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 487
 488 void
 489 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 490                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 491                     unsigned int nob)
 492 {
 493         /* NB iov, kiov are READ-ONLY */
 494         unsigned int    this_nob;
 495         char           *addr = NULL;
 496
 497         if (nob == 0)
 498                 return;
 499
 500         LASSERT (!in_interrupt ());
 501
 502         LASSERT (niov > 0);
 503         while (iovoffset >= iov->iov_len) {
 504                 iovoffset -= iov->iov_len;
 505                 iov++;
 506                 niov--;
 507                 LASSERT(niov > 0);
 508         }
 509
 510         LASSERT(nkiov > 0);
 511         while (kiovoffset >= kiov->kiov_len) {
 512                 kiovoffset -= kiov->kiov_len;
 513                 kiov++;
 514                 nkiov--;
 515                 LASSERT(nkiov > 0);
 516         }
 517
 518         do {
 519                 LASSERT(niov > 0);
 520                 LASSERT(nkiov > 0);
 521                 this_nob = MIN(iov->iov_len - iovoffset,
 522                                kiov->kiov_len - kiovoffset);
 523                 this_nob = MIN(this_nob, nob);
 524
 525                 if (addr == NULL)
 526                         addr = ((char *)kmap(kiov->kiov_page)) +
 527                                 kiov->kiov_offset + kiovoffset;
 528
 529                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 530                 nob -= this_nob;
 531
 532                 if (iov->iov_len > iovoffset + this_nob) {
 533                         iovoffset += this_nob;
 534                 } else {
 535                         iov++;
 536                         niov--;
 537                         iovoffset = 0;
 538                 }
 539
 540                 if (kiov->kiov_len > kiovoffset + this_nob) {
 541                         addr += this_nob;
 542                         kiovoffset += this_nob;
 543                 } else {
 544                         kunmap(kiov->kiov_page);
 545                         addr = NULL;
 546                         kiov++;
 547                         nkiov--;
 548                         kiovoffset = 0;
 549                 }
 550
 551         } while (nob > 0);
 552
 553         if (addr != NULL)
 554                 kunmap(kiov->kiov_page);
 555 }
 556 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 557
 558 void
 559 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 560                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 561                    unsigned int nob)
 562 {
 563         /* NB kiov, iov are READ-ONLY */
 564         unsigned int    this_nob;
 565         char           *addr = NULL;
 566
 567         if (nob == 0)
 568                 return;
 569
 570         LASSERT (!in_interrupt ());
 571
 572         LASSERT (nkiov > 0);
 573         while (kiovoffset >= kiov->kiov_len) {
 574                 kiovoffset -= kiov->kiov_len;
 575                 kiov++;
 576                 nkiov--;
 577                 LASSERT(nkiov > 0);
 578         }
 579
 580         LASSERT(niov > 0);
 581         while (iovoffset >= iov->iov_len) {
 582                 iovoffset -= iov->iov_len;
 583                 iov++;
 584                 niov--;
 585                 LASSERT(niov > 0);
 586         }
 587
 588         do {
 589                 LASSERT(nkiov > 0);
 590                 LASSERT(niov > 0);
 591                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 592                                iov->iov_len - iovoffset);
 593                 this_nob = MIN(this_nob, nob);
 594
 595                 if (addr == NULL)
 596                         addr = ((char *)kmap(kiov->kiov_page)) +
 597                                 kiov->kiov_offset + kiovoffset;
 598
 599                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 600                 nob -= this_nob;
 601
 602                 if (kiov->kiov_len > kiovoffset + this_nob) {
 603                         addr += this_nob;
 604                         kiovoffset += this_nob;
 605                 } else {
 606                         kunmap(kiov->kiov_page);
 607                         addr = NULL;
 608                         kiov++;
 609                         nkiov--;
 610                         kiovoffset = 0;
 611                 }
 612
 613                 if (iov->iov_len > iovoffset + this_nob) {
 614                         iovoffset += this_nob;
 615                 } else {
 616                         iov++;
 617                         niov--;
 618                         iovoffset = 0;
 619                 }
 620         } while (nob > 0);
 621
 622         if (addr != NULL)
 623                 kunmap(kiov->kiov_page);
 624 }
 625 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 626
 627 int
 628 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 629                   int src_niov, lnet_kiov_t *src,
 630                   unsigned int offset, unsigned int len)
 631 {
 632         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 633          * for exactly 'len' bytes, and return the number of entries.
 634          * NB not destructive to 'src' */
 635         unsigned int    frag_len;
 636         unsigned int    niov;
 637
 638         if (len == 0)                           /* no data => */
 639                 return (0);                     /* no frags */
 640
 641         LASSERT(src_niov > 0);
 642         while (offset >= src->kiov_len) {      /* skip initial frags */
 643                 offset -= src->kiov_len;
 644                 src_niov--;
 645                 src++;
 646                 LASSERT(src_niov > 0);
 647         }
 648
 649         niov = 1;
 650         for (;;) {
 651                 LASSERT(src_niov > 0);
 652                 LASSERT((int)niov <= dst_niov);
 653
 654                 frag_len = src->kiov_len - offset;
 655                 dst->kiov_page = src->kiov_page;
 656                 dst->kiov_offset = src->kiov_offset + offset;
 657
 658                 if (len <= frag_len) {
 659                         dst->kiov_len = len;
 660                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 661                         return niov;
 662                 }
 663
 664                 dst->kiov_len = frag_len;
 665                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 666
 667                 len -= frag_len;
 668                 dst++;
 669                 src++;
 670                 niov++;
 671                 src_niov--;
 672                 offset = 0;
 673         }
 674 }
 675 EXPORT_SYMBOL(lnet_extract_kiov);
 676
 677 void
 678 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 679              int delayed, unsigned int offset, unsigned int mlen,
 680              unsigned int rlen)
 681 {
 682         unsigned int  niov = 0;
 683         struct kvec *iov = NULL;
 684         lnet_kiov_t  *kiov = NULL;
 685         int           rc;
 686
 687         LASSERT (!in_interrupt ());
 688         LASSERT (mlen == 0 || msg != NULL);
 689
 690         if (msg != NULL) {
 691                 LASSERT(msg->msg_receiving);
 692                 LASSERT(!msg->msg_sending);
 693                 LASSERT(rlen == msg->msg_len);
 694                 LASSERT(mlen <= msg->msg_len);
 695                 LASSERT(msg->msg_offset == offset);
 696                 LASSERT(msg->msg_wanted == mlen);
 697
 698                 msg->msg_receiving = 0;
 699
 700                 if (mlen != 0) {
 701                         niov = msg->msg_niov;
 702                         iov  = msg->msg_iov;
 703                         kiov = msg->msg_kiov;
 704
 705                         LASSERT (niov > 0);
 706                         LASSERT ((iov == NULL) != (kiov == NULL));
 707                 }
 708         }
 709
 710         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 711                                              niov, iov, kiov, offset, mlen,
 712                                              rlen);
 713         if (rc < 0)
 714                 lnet_finalize(msg, rc);
 715 }
 716
 717 static void
 718 lnet_setpayloadbuffer(struct lnet_msg *msg)
 719 {
 720         struct lnet_libmd *md = msg->msg_md;
 721
 722         LASSERT(msg->msg_len > 0);
 723         LASSERT(!msg->msg_routing);
 724         LASSERT(md != NULL);
 725         LASSERT(msg->msg_niov == 0);
 726         LASSERT(msg->msg_iov == NULL);
 727         LASSERT(msg->msg_kiov == NULL);
 728
 729         msg->msg_niov = md->md_niov;
 730         if ((md->md_options & LNET_MD_KIOV) != 0)
 731                 msg->msg_kiov = md->md_iov.kiov;
 732         else
 733                 msg->msg_iov = md->md_iov.iov;
 734 }
 735
 736 void
 737 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 738                unsigned int offset, unsigned int len)
 739 {
 740         msg->msg_type = type;
 741         msg->msg_target = target;
 742         msg->msg_len = len;
 743         msg->msg_offset = offset;
 744
 745         if (len != 0)
 746                 lnet_setpayloadbuffer(msg);
 747
 748         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 749         msg->msg_hdr.type           = cpu_to_le32(type);
 750         /* dest_nid will be overwritten by lnet_select_pathway() */
 751         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 752         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 753         /* src_nid will be set later */
 754         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 755         msg->msg_hdr.payload_length = cpu_to_le32(len);
 756 }
 757
 758 static void
 759 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 760 {
 761         void   *priv = msg->msg_private;
 762         int rc;
 763
 764         LASSERT (!in_interrupt ());
 765         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 766                  (msg->msg_txcredit && msg->msg_peertxcredit));
 767
 768         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 769         if (rc < 0) {
 770                 msg->msg_no_resend = true;
 771                 lnet_finalize(msg, rc);
 772         }
 773 }
 774
 775 static int
 776 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 777 {
 778         int     rc;
 779
 780         LASSERT(!msg->msg_sending);
 781         LASSERT(msg->msg_receiving);
 782         LASSERT(!msg->msg_rx_ready_delay);
 783         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 784
 785         msg->msg_rx_ready_delay = 1;
 786         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 787                                                   &msg->msg_private);
 788         if (rc != 0) {
 789                 CERROR("recv from %s / send to %s aborted: "
 790                        "eager_recv failed %d\n",
 791                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 792                        libcfs_id2str(msg->msg_target), rc);
 793                 LASSERT(rc < 0); /* required by my callers */
 794         }
 795
 796         return rc;
 797 }
 798
 799 static bool
 800 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 801 {
 802         time64_t deadline;
 803
 804         deadline = lpni->lpni_last_alive +
 805                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 806
 807         /*
 808          * assume peer_ni is alive as long as we're within the configured
 809          * peer timeout
 810          */
 811         if (deadline > now)
 812                 return false;
 813
 814         return true;
 815 }
 816
 817 /* NB: returns 1 when alive, 0 when dead, negative when error;
 818  *     may drop the lnet_net_lock */
 819 static int
 820 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 821                        struct lnet_msg *msg)
 822 {
 823         time64_t now = ktime_get_seconds();
 824
 825         if (!lnet_peer_aliveness_enabled(lpni))
 826                 return -ENODEV;
 827
 828         /*
 829          * If we're resending a message, let's attempt to send it even if
 830          * the peer is down to fulfill our resend quota on the message
 831          */
 832         if (msg->msg_retry_count > 0)
 833                 return 1;
 834
 835         /* try and send recovery messages irregardless */
 836         if (msg->msg_recovery)
 837                 return 1;
 838
 839         /* always send any responses */
 840         if (msg->msg_type == LNET_MSG_ACK ||
 841             msg->msg_type == LNET_MSG_REPLY)
 842                 return 1;
 843
 844         if (!lnet_is_peer_deadline_passed(lpni, now))
 845                 return true;
 846
 847         return lnet_is_peer_ni_alive(lpni);
 848 }
 849
 850 /**
 851  * \param msg The message to be sent.
 852  * \param do_send True if lnet_ni_send() should be called in this function.
 853  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 854  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 855  *
 856  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 857  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 858  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 859  * \retval -ECANCELED If the MD of the message has been unlinked.
 860  */
 861 static int
 862 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 863 {
 864         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 865         struct lnet_ni          *ni = msg->msg_txni;
 866         int                     cpt = msg->msg_tx_cpt;
 867         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 868
 869         /* non-lnet_send() callers have checked before */
 870         LASSERT(!do_send || msg->msg_tx_delayed);
 871         LASSERT(!msg->msg_receiving);
 872         LASSERT(msg->msg_tx_committed);
 873         /* can't get here if we're sending to the loopback interface */
 874         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 875
 876         /* NB 'lp' is always the next hop */
 877         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 878             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 879                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 880                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 881                         msg->msg_len;
 882                 lnet_net_unlock(cpt);
 883                 if (msg->msg_txpeer)
 884                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 885                                         msg->msg_type,
 886                                         LNET_STATS_TYPE_DROP);
 887                 if (msg->msg_txni)
 888                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 889                                         msg->msg_type,
 890                                         LNET_STATS_TYPE_DROP);
 891
 892                 CNETERR("Dropping message for %s: peer not alive\n",
 893                         libcfs_id2str(msg->msg_target));
 894                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 895                 if (do_send)
 896                         lnet_finalize(msg, -EHOSTUNREACH);
 897
 898                 lnet_net_lock(cpt);
 899                 return -EHOSTUNREACH;
 900         }
 901
 902         if (msg->msg_md != NULL &&
 903             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 904                 lnet_net_unlock(cpt);
 905
 906                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 907                         "called on the MD/ME.\n",
 908                         libcfs_id2str(msg->msg_target));
 909                 if (do_send) {
 910                         msg->msg_no_resend = true;
 911                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 912                                msg, libcfs_id2str(msg->msg_target));
 913                         lnet_finalize(msg, -ECANCELED);
 914                 }
 915
 916                 lnet_net_lock(cpt);
 917                 return -ECANCELED;
 918         }
 919
 920         if (!msg->msg_peertxcredit) {
 921                 spin_lock(&lp->lpni_lock);
 922                 LASSERT((lp->lpni_txcredits < 0) ==
 923                         !list_empty(&lp->lpni_txq));
 924
 925                 msg->msg_peertxcredit = 1;
 926                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 927                 lp->lpni_txcredits--;
 928
 929                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 930                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 931
 932                 if (lp->lpni_txcredits < 0) {
 933                         msg->msg_tx_delayed = 1;
 934                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 935                         spin_unlock(&lp->lpni_lock);
 936                         return LNET_CREDIT_WAIT;
 937                 }
 938                 spin_unlock(&lp->lpni_lock);
 939         }
 940
 941         if (!msg->msg_txcredit) {
 942                 LASSERT((tq->tq_credits < 0) ==
 943                         !list_empty(&tq->tq_delayed));
 944
 945                 msg->msg_txcredit = 1;
 946                 tq->tq_credits--;
 947                 atomic_dec(&ni->ni_tx_credits);
 948
 949                 if (tq->tq_credits < tq->tq_credits_min)
 950                         tq->tq_credits_min = tq->tq_credits;
 951
 952                 if (tq->tq_credits < 0) {
 953                         msg->msg_tx_delayed = 1;
 954                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 955                         return LNET_CREDIT_WAIT;
 956                 }
 957         }
 958
 959         /* unset the tx_delay flag as we're going to send it now */
 960         msg->msg_tx_delayed = 0;
 961
 962         if (do_send) {
 963                 lnet_net_unlock(cpt);
 964                 lnet_ni_send(ni, msg);
 965                 lnet_net_lock(cpt);
 966         }
 967         return LNET_CREDIT_OK;
 968 }
 969
 970
 971 static struct lnet_rtrbufpool *
 972 lnet_msg2bufpool(struct lnet_msg *msg)
 973 {
 974         struct lnet_rtrbufpool  *rbp;
 975         int                     cpt;
 976
 977         LASSERT(msg->msg_rx_committed);
 978
 979         cpt = msg->msg_rx_cpt;
 980         rbp = &the_lnet.ln_rtrpools[cpt][0];
 981
 982         LASSERT(msg->msg_len <= LNET_MTU);
 983         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 984                 rbp++;
 985                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 986         }
 987
 988         return rbp;
 989 }
 990
 991 static int
 992 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 993 {
 994         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 995          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 996          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 997          * received or OK to receive */
 998         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 999         struct lnet_peer *lp;
1000         struct lnet_rtrbufpool *rbp;
1001         struct lnet_rtrbuf *rb;
1002
1003         LASSERT(msg->msg_iov == NULL);
1004         LASSERT(msg->msg_kiov == NULL);
1005         LASSERT(msg->msg_niov == 0);
1006         LASSERT(msg->msg_routing);
1007         LASSERT(msg->msg_receiving);
1008         LASSERT(!msg->msg_sending);
1009         LASSERT(lpni->lpni_peer_net);
1010         LASSERT(lpni->lpni_peer_net->lpn_peer);
1011
1012         lp = lpni->lpni_peer_net->lpn_peer;
1013
1014         /* non-lnet_parse callers only receive delayed messages */
1015         LASSERT(!do_recv || msg->msg_rx_delayed);
1016
1017         if (!msg->msg_peerrtrcredit) {
1018                 /* lpni_lock protects the credit manipulation */
1019                 spin_lock(&lpni->lpni_lock);
1020                 /* lp_lock protects the lp_rtrq */
1021                 spin_lock(&lp->lp_lock);
1022
1023                 msg->msg_peerrtrcredit = 1;
1024                 lpni->lpni_rtrcredits--;
1025                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1026                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1027
1028                 if (lpni->lpni_rtrcredits < 0) {
1029                         /* must have checked eager_recv before here */
1030                         LASSERT(msg->msg_rx_ready_delay);
1031                         msg->msg_rx_delayed = 1;
1032                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1033                         spin_unlock(&lp->lp_lock);
1034                         spin_unlock(&lpni->lpni_lock);
1035                         return LNET_CREDIT_WAIT;
1036                 }
1037                 spin_unlock(&lp->lp_lock);
1038                 spin_unlock(&lpni->lpni_lock);
1039         }
1040
1041         rbp = lnet_msg2bufpool(msg);
1042
1043         if (!msg->msg_rtrcredit) {
1044                 msg->msg_rtrcredit = 1;
1045                 rbp->rbp_credits--;
1046                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1047                         rbp->rbp_mincredits = rbp->rbp_credits;
1048
1049                 if (rbp->rbp_credits < 0) {
1050                         /* must have checked eager_recv before here */
1051                         LASSERT(msg->msg_rx_ready_delay);
1052                         msg->msg_rx_delayed = 1;
1053                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1054                         return LNET_CREDIT_WAIT;
1055                 }
1056         }
1057
1058         LASSERT(!list_empty(&rbp->rbp_bufs));
1059         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1060         list_del(&rb->rb_list);
1061
1062         msg->msg_niov = rbp->rbp_npages;
1063         msg->msg_kiov = &rb->rb_kiov[0];
1064
1065         /* unset the msg-rx_delayed flag since we're receiving the message */
1066         msg->msg_rx_delayed = 0;
1067
1068         if (do_recv) {
1069                 int cpt = msg->msg_rx_cpt;
1070
1071                 lnet_net_unlock(cpt);
1072                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1073                              0, msg->msg_len, msg->msg_len);
1074                 lnet_net_lock(cpt);
1075         }
1076         return LNET_CREDIT_OK;
1077 }
1078
1079 void
1080 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1081 {
1082         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1083         struct lnet_ni          *txni = msg->msg_txni;
1084         struct lnet_msg         *msg2;
1085
1086         if (msg->msg_txcredit) {
1087                 struct lnet_ni       *ni = msg->msg_txni;
1088                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1089
1090                 /* give back NI txcredits */
1091                 msg->msg_txcredit = 0;
1092
1093                 LASSERT((tq->tq_credits < 0) ==
1094                         !list_empty(&tq->tq_delayed));
1095
1096                 tq->tq_credits++;
1097                 atomic_inc(&ni->ni_tx_credits);
1098                 if (tq->tq_credits <= 0) {
1099                         msg2 = list_entry(tq->tq_delayed.next,
1100                                           struct lnet_msg, msg_list);
1101                         list_del(&msg2->msg_list);
1102
1103                         LASSERT(msg2->msg_txni == ni);
1104                         LASSERT(msg2->msg_tx_delayed);
1105                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1106
1107                         (void) lnet_post_send_locked(msg2, 1);
1108                 }
1109         }
1110
1111         if (msg->msg_peertxcredit) {
1112                 /* give back peer txcredits */
1113                 msg->msg_peertxcredit = 0;
1114
1115                 spin_lock(&txpeer->lpni_lock);
1116                 LASSERT((txpeer->lpni_txcredits < 0) ==
1117                         !list_empty(&txpeer->lpni_txq));
1118
1119                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1120                 LASSERT(txpeer->lpni_txqnob >= 0);
1121
1122                 txpeer->lpni_txcredits++;
1123                 if (txpeer->lpni_txcredits <= 0) {
1124                         int msg2_cpt;
1125
1126                         msg2 = list_entry(txpeer->lpni_txq.next,
1127                                               struct lnet_msg, msg_list);
1128                         list_del(&msg2->msg_list);
1129                         spin_unlock(&txpeer->lpni_lock);
1130
1131                         LASSERT(msg2->msg_txpeer == txpeer);
1132                         LASSERT(msg2->msg_tx_delayed);
1133
1134                         msg2_cpt = msg2->msg_tx_cpt;
1135
1136                         /*
1137                          * The msg_cpt can be different from the msg2_cpt
1138                          * so we need to make sure we lock the correct cpt
1139                          * for msg2.
1140                          * Once we call lnet_post_send_locked() it is no
1141                          * longer safe to access msg2, since it could've
1142                          * been freed by lnet_finalize(), but we still
1143                          * need to relock the correct cpt, so we cache the
1144                          * msg2_cpt for the purpose of the check that
1145                          * follows the call to lnet_pose_send_locked().
1146                          */
1147                         if (msg2_cpt != msg->msg_tx_cpt) {
1148                                 lnet_net_unlock(msg->msg_tx_cpt);
1149                                 lnet_net_lock(msg2_cpt);
1150                         }
1151                         (void) lnet_post_send_locked(msg2, 1);
1152                         if (msg2_cpt != msg->msg_tx_cpt) {
1153                                 lnet_net_unlock(msg2_cpt);
1154                                 lnet_net_lock(msg->msg_tx_cpt);
1155                         }
1156                 } else {
1157                         spin_unlock(&txpeer->lpni_lock);
1158                 }
1159         }
1160
1161         if (txni != NULL) {
1162                 msg->msg_txni = NULL;
1163                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1164         }
1165
1166         if (txpeer != NULL) {
1167                 msg->msg_txpeer = NULL;
1168                 lnet_peer_ni_decref_locked(txpeer);
1169         }
1170 }
1171
1172 void
1173 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1174 {
1175         struct lnet_msg *msg;
1176
1177         if (list_empty(&rbp->rbp_msgs))
1178                 return;
1179         msg = list_entry(rbp->rbp_msgs.next,
1180                          struct lnet_msg, msg_list);
1181         list_del(&msg->msg_list);
1182
1183         (void)lnet_post_routed_recv_locked(msg, 1);
1184 }
1185
1186 void
1187 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1188 {
1189         struct lnet_msg *msg;
1190         struct lnet_msg *tmp;
1191
1192         lnet_net_unlock(cpt);
1193
1194         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1195                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1196                              0, 0, 0, msg->msg_hdr.payload_length);
1197                 list_del_init(&msg->msg_list);
1198                 msg->msg_no_resend = true;
1199                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1200                 lnet_finalize(msg, -ECANCELED);
1201         }
1202
1203         lnet_net_lock(cpt);
1204 }
1205
1206 void
1207 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1208 {
1209         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1210         struct lnet_peer *lp;
1211         struct lnet_ni *rxni = msg->msg_rxni;
1212         struct lnet_msg *msg2;
1213
1214         if (msg->msg_rtrcredit) {
1215                 /* give back global router credits */
1216                 struct lnet_rtrbuf *rb;
1217                 struct lnet_rtrbufpool *rbp;
1218
1219                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1220                  * there until it gets one allocated, or aborts the wait
1221                  * itself */
1222                 LASSERT(msg->msg_kiov != NULL);
1223
1224                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1225                 rbp = rb->rb_pool;
1226
1227                 msg->msg_kiov = NULL;
1228                 msg->msg_rtrcredit = 0;
1229
1230                 LASSERT(rbp == lnet_msg2bufpool(msg));
1231
1232                 LASSERT((rbp->rbp_credits > 0) ==
1233                         !list_empty(&rbp->rbp_bufs));
1234
1235                 /* If routing is now turned off, we just drop this buffer and
1236                  * don't bother trying to return credits.  */
1237                 if (!the_lnet.ln_routing) {
1238                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1239                         goto routing_off;
1240                 }
1241
1242                 /* It is possible that a user has lowered the desired number of
1243                  * buffers in this pool.  Make sure we never put back
1244                  * more buffers than the stated number. */
1245                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1246                         /* Discard this buffer so we don't have too
1247                          * many. */
1248                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1249                         rbp->rbp_nbuffers--;
1250                 } else {
1251                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1252                         rbp->rbp_credits++;
1253                         if (rbp->rbp_credits <= 0)
1254                                 lnet_schedule_blocked_locked(rbp);
1255                 }
1256         }
1257
1258 routing_off:
1259         if (msg->msg_peerrtrcredit) {
1260                 LASSERT(rxpeerni);
1261                 LASSERT(rxpeerni->lpni_peer_net);
1262                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1263
1264                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1265
1266                 /* give back peer router credits */
1267                 msg->msg_peerrtrcredit = 0;
1268
1269                 spin_lock(&rxpeerni->lpni_lock);
1270                 spin_lock(&lp->lp_lock);
1271
1272                 rxpeerni->lpni_rtrcredits++;
1273
1274                 /* drop all messages which are queued to be routed on that
1275                  * peer. */
1276                 if (!the_lnet.ln_routing) {
1277                         struct list_head drop;
1278                         INIT_LIST_HEAD(&drop);
1279                         list_splice_init(&lp->lp_rtrq, &drop);
1280                         spin_unlock(&lp->lp_lock);
1281                         spin_unlock(&rxpeerni->lpni_lock);
1282                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1283                 } else if (!list_empty(&lp->lp_rtrq)) {
1284                         int msg2_cpt;
1285
1286                         msg2 = list_entry(lp->lp_rtrq.next,
1287                                           struct lnet_msg, msg_list);
1288                         list_del(&msg2->msg_list);
1289                         msg2_cpt = msg2->msg_rx_cpt;
1290                         spin_unlock(&lp->lp_lock);
1291                         spin_unlock(&rxpeerni->lpni_lock);
1292                         /*
1293                          * messages on the lp_rtrq can be from any NID in
1294                          * the peer, which means they might have different
1295                          * cpts. We need to make sure we lock the right
1296                          * one.
1297                          */
1298                         if (msg2_cpt != msg->msg_rx_cpt) {
1299                                 lnet_net_unlock(msg->msg_rx_cpt);
1300                                 lnet_net_lock(msg2_cpt);
1301                         }
1302                         (void) lnet_post_routed_recv_locked(msg2, 1);
1303                         if (msg2_cpt != msg->msg_rx_cpt) {
1304                                 lnet_net_unlock(msg2_cpt);
1305                                 lnet_net_lock(msg->msg_rx_cpt);
1306                         }
1307                 } else {
1308                         spin_unlock(&lp->lp_lock);
1309                         spin_unlock(&rxpeerni->lpni_lock);
1310                 }
1311         }
1312         if (rxni != NULL) {
1313                 msg->msg_rxni = NULL;
1314                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1315         }
1316         if (rxpeerni != NULL) {
1317                 msg->msg_rxpeer = NULL;
1318                 lnet_peer_ni_decref_locked(rxpeerni);
1319         }
1320 }
1321
1322 static int
1323 lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1324 {
1325         if (p1->lpni_txqnob < p2->lpni_txqnob)
1326                 return 1;
1327
1328         if (p1->lpni_txqnob > p2->lpni_txqnob)
1329                 return -1;
1330
1331         if (p1->lpni_txcredits > p2->lpni_txcredits)
1332                 return 1;
1333
1334         if (p1->lpni_txcredits < p2->lpni_txcredits)
1335                 return -1;
1336
1337         return 0;
1338 }
1339
1340 static struct lnet_peer_ni *
1341 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1342                     struct lnet_peer *peer,
1343                     struct lnet_peer_net *peer_net)
1344 {
1345         /*
1346          * Look at the peer NIs for the destination peer that connect
1347          * to the chosen net. If a peer_ni is preferred when using the
1348          * best_ni to communicate, we use that one. If there is no
1349          * preferred peer_ni, or there are multiple preferred peer_ni,
1350          * the available transmit credits are used. If the transmit
1351          * credits are equal, we round-robin over the peer_ni.
1352          */
1353         struct lnet_peer_ni *lpni = NULL;
1354         struct lnet_peer_ni *best_lpni = NULL;
1355         int best_lpni_credits = INT_MIN;
1356         bool preferred = false;
1357         bool ni_is_pref;
1358         int best_lpni_healthv = 0;
1359         int lpni_healthv;
1360
1361         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1362                 /*
1363                  * if the best_ni we've chosen aleady has this lpni
1364                  * preferred, then let's use it
1365                  */
1366                 if (best_ni) {
1367                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1368                                                                 best_ni->ni_nid);
1369                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1370                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1371                 } else {
1372                         ni_is_pref = false;
1373                 }
1374
1375                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1376
1377                 if (best_lpni)
1378                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1379                                 libcfs_nid2str(lpni->lpni_nid),
1380                                 lpni->lpni_txcredits, best_lpni_credits,
1381                                 lpni->lpni_seq, best_lpni->lpni_seq);
1382
1383                 /* pick the healthiest peer ni */
1384                 if (lpni_healthv < best_lpni_healthv) {
1385                         continue;
1386                 } else if (lpni_healthv > best_lpni_healthv) {
1387                         best_lpni_healthv = lpni_healthv;
1388                 /* if this is a preferred peer use it */
1389                 } else if (!preferred && ni_is_pref) {
1390                         preferred = true;
1391                 } else if (preferred && !ni_is_pref) {
1392                         /*
1393                          * this is not the preferred peer so let's ignore
1394                          * it.
1395                          */
1396                         continue;
1397                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1398                         /*
1399                          * We already have a peer that has more credits
1400                          * available than this one. No need to consider
1401                          * this peer further.
1402                          */
1403                         continue;
1404                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1405                         /*
1406                          * The best peer found so far and the current peer
1407                          * have the same number of available credits let's
1408                          * make sure to select between them using Round
1409                          * Robin
1410                          */
1411                         if (best_lpni) {
1412                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1413                                         continue;
1414                         }
1415                 }
1416
1417                 best_lpni = lpni;
1418                 best_lpni_credits = lpni->lpni_txcredits;
1419         }
1420
1421         /* if we still can't find a peer ni then we can't reach it */
1422         if (!best_lpni) {
1423                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1424                         LNET_NIDNET(dst_nid);
1425                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1426                                 libcfs_net2str(net_id));
1427                 return NULL;
1428         }
1429
1430         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1431                libcfs_nid2str(best_lpni->lpni_nid));
1432
1433         return best_lpni;
1434 }
1435
1436 /*
1437  * Prerequisite: the best_ni should already be set in the sd
1438  */
1439 static inline struct lnet_peer_ni *
1440 lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
1441                            __u32 net_id)
1442 {
1443         struct lnet_peer_net *peer_net;
1444
1445         /*
1446          * The gateway is Multi-Rail capable so now we must select the
1447          * proper peer_ni
1448          */
1449         peer_net = lnet_peer_get_net_locked(peer, net_id);
1450
1451         if (!peer_net) {
1452                 CERROR("gateway peer %s has no NI on net %s\n",
1453                        libcfs_nid2str(peer->lp_primary_nid),
1454                        libcfs_net2str(net_id));
1455                 return NULL;
1456         }
1457
1458         return lnet_select_peer_ni(sd->sd_best_ni, sd->sd_dst_nid,
1459                                    peer, peer_net);
1460 }
1461
1462 static int
1463 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2,
1464                     struct lnet_peer_ni **best_lpni)
1465 {
1466         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1467         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1468         struct lnet_peer *lp1 = r1->lr_gateway;
1469         struct lnet_peer *lp2 = r2->lr_gateway;
1470         struct lnet_peer_ni *lpni1;
1471         struct lnet_peer_ni *lpni2;
1472         struct lnet_send_data sd;
1473         int rc;
1474
1475         sd.sd_best_ni = NULL;
1476         sd.sd_dst_nid = LNET_NID_ANY;
1477         lpni1 = lnet_find_best_lpni_on_net(&sd, lp1, r1->lr_lnet);
1478         lpni2 = lnet_find_best_lpni_on_net(&sd, lp2, r2->lr_lnet);
1479         LASSERT(lpni1 && lpni2);
1480
1481         if (r1->lr_priority < r2->lr_priority) {
1482                 *best_lpni = lpni1;
1483                 return 1;
1484         }
1485
1486         if (r1->lr_priority > r2->lr_priority) {
1487                 *best_lpni = lpni2;
1488                 return -1;
1489         }
1490
1491         if (r1_hops < r2_hops) {
1492                 *best_lpni = lpni1;
1493                 return 1;
1494         }
1495
1496         if (r1_hops > r2_hops) {
1497                 *best_lpni = lpni2;
1498                 return -1;
1499         }
1500
1501         rc = lnet_compare_peers(lpni1, lpni2);
1502         if (rc == 1) {
1503                 *best_lpni = lpni1;
1504                 return rc;
1505         } else if (rc == -1) {
1506                 *best_lpni = lpni2;
1507                 return rc;
1508         }
1509
1510         if (r1->lr_seq - r2->lr_seq <= 0) {
1511                 *best_lpni = lpni1;
1512                 return 1;
1513         }
1514
1515         *best_lpni = lpni2;
1516         return -1;
1517 }
1518
1519 static struct lnet_route *
1520 lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
1521                        lnet_nid_t rtr_nid, struct lnet_route **prev_route,
1522                        struct lnet_peer_ni **gwni)
1523 {
1524         struct lnet_peer_ni *best_gw_ni = NULL;
1525         struct lnet_route *best_route;
1526         struct lnet_route *last_route;
1527         struct lnet_remotenet *rnet;
1528         struct lnet_peer *lp_best;
1529         struct lnet_route *route;
1530         struct lnet_peer *lp;
1531         int rc;
1532
1533         /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
1534          * rtr_nid nid, otherwise find the best gateway I can use */
1535
1536         rnet = lnet_find_rnet_locked(remote_net);
1537         if (rnet == NULL)
1538                 return NULL;
1539
1540         lp_best = NULL;
1541         best_route = last_route = NULL;
1542         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1543                 lp = route->lr_gateway;
1544
1545                 if (!lnet_is_route_alive(route))
1546                         continue;
1547
1548                 if (lp_best == NULL) {
1549                         best_route = last_route = route;
1550                         lp_best = lp;
1551                 }
1552
1553                 /* no protection on below fields, but it's harmless */
1554                 if (last_route->lr_seq - route->lr_seq < 0)
1555                         last_route = route;
1556
1557                 rc = lnet_compare_routes(route, best_route, &best_gw_ni);
1558                 if (rc < 0)
1559                         continue;
1560
1561                 best_route = route;
1562                 lp_best = lp;
1563         }
1564
1565         *prev_route = last_route;
1566         *gwni = best_gw_ni;
1567
1568         return best_route;
1569 }
1570
1571 static struct lnet_ni *
1572 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1573                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1574                  int md_cpt)
1575 {
1576         struct lnet_ni *ni = NULL;
1577         unsigned int shortest_distance;
1578         int best_credits;
1579         int best_healthv;
1580
1581         /*
1582          * If there is no peer_ni that we can send to on this network,
1583          * then there is no point in looking for a new best_ni here.
1584         */
1585         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1586                 return best_ni;
1587
1588         if (best_ni == NULL) {
1589                 shortest_distance = UINT_MAX;
1590                 best_credits = INT_MIN;
1591                 best_healthv = 0;
1592         } else {
1593                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1594                                                      best_ni->ni_dev_cpt);
1595                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1596                 best_healthv = atomic_read(&best_ni->ni_healthv);
1597         }
1598
1599         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1600                 unsigned int distance;
1601                 int ni_credits;
1602                 int ni_healthv;
1603                 int ni_fatal;
1604
1605                 ni_credits = atomic_read(&ni->ni_tx_credits);
1606                 ni_healthv = atomic_read(&ni->ni_healthv);
1607                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1608
1609                 /*
1610                  * calculate the distance from the CPT on which
1611                  * the message memory is allocated to the CPT of
1612                  * the NI's physical device
1613                  */
1614                 distance = cfs_cpt_distance(lnet_cpt_table(),
1615                                             md_cpt,
1616                                             ni->ni_dev_cpt);
1617
1618                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1619                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1620                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1621                         : "not seleced", best_credits, shortest_distance,
1622                         (best_ni) ? best_ni->ni_seq : 0);
1623
1624                 /*
1625                  * All distances smaller than the NUMA range
1626                  * are treated equally.
1627                  */
1628                 if (distance < lnet_numa_range)
1629                         distance = lnet_numa_range;
1630
1631                 /*
1632                  * Select on health, shorter distance, available
1633                  * credits, then round-robin.
1634                  */
1635                 if (ni_fatal) {
1636                         continue;
1637                 } else if (ni_healthv < best_healthv) {
1638                         continue;
1639                 } else if (ni_healthv > best_healthv) {
1640                         best_healthv = ni_healthv;
1641                         /*
1642                          * If we're going to prefer this ni because it's
1643                          * the healthiest, then we should set the
1644                          * shortest_distance in the algorithm in case
1645                          * there are multiple NIs with the same health but
1646                          * different distances.
1647                          */
1648                         if (distance < shortest_distance)
1649                                 shortest_distance = distance;
1650                 } else if (distance > shortest_distance) {
1651                         continue;
1652                 } else if (distance < shortest_distance) {
1653                         shortest_distance = distance;
1654                 } else if (ni_credits < best_credits) {
1655                         continue;
1656                 } else if (ni_credits == best_credits) {
1657                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1658                                 continue;
1659                 }
1660                 best_ni = ni;
1661                 best_credits = ni_credits;
1662         }
1663
1664         CDEBUG(D_NET, "selected best_ni %s\n",
1665                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1666
1667         return best_ni;
1668 }
1669
1670 /*
1671  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1672  * because such traffic is required to perform discovery. We therefore
1673  * exclude all GET and PUT on that portal. We also exclude all ACK and
1674  * REPLY traffic, but that is because the portal is not tracked in the
1675  * message structure for these message types. We could restrict this
1676  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1677  */
1678 static bool
1679 lnet_msg_discovery(struct lnet_msg *msg)
1680 {
1681         if (msg->msg_type == LNET_MSG_PUT) {
1682                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1683                         return true;
1684         } else if (msg->msg_type == LNET_MSG_GET) {
1685                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1686                         return true;
1687         }
1688         return false;
1689 }
1690
1691 #define SRC_SPEC        0x0001
1692 #define SRC_ANY         0x0002
1693 #define LOCAL_DST       0x0004
1694 #define REMOTE_DST      0x0008
1695 #define MR_DST          0x0010
1696 #define NMR_DST         0x0020
1697 #define SND_RESP        0x0040
1698
1699 /* The following to defines are used for return codes */
1700 #define REPEAT_SEND     0x1000
1701 #define PASS_THROUGH    0x2000
1702
1703 /* The different cases lnet_select pathway needs to handle */
1704 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1705 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1706 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1707 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1708 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1709 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1710 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1711 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1712
1713 static int
1714 lnet_handle_lo_send(struct lnet_send_data *sd)
1715 {
1716         struct lnet_msg *msg = sd->sd_msg;
1717         int cpt = sd->sd_cpt;
1718
1719         /* No send credit hassles with LOLND */
1720         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1721         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1722         if (!msg->msg_routing)
1723                 msg->msg_hdr.src_nid =
1724                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1725         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1726         lnet_msg_commit(msg, cpt);
1727         msg->msg_txni = the_lnet.ln_loni;
1728
1729         return LNET_CREDIT_OK;
1730 }
1731
1732 static int
1733 lnet_handle_send(struct lnet_send_data *sd)
1734 {
1735         struct lnet_ni *best_ni = sd->sd_best_ni;
1736         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1737         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1738         struct lnet_msg *msg = sd->sd_msg;
1739         int cpt2;
1740         __u32 send_case = sd->sd_send_case;
1741         int rc;
1742         __u32 routing = send_case & REMOTE_DST;
1743          struct lnet_rsp_tracker *rspt;
1744
1745         /*
1746          * Increment sequence number of the selected peer so that we
1747          * pick the next one in Round Robin.
1748          */
1749         best_lpni->lpni_seq++;
1750
1751         /*
1752          * grab a reference on the peer_ni so it sticks around even if
1753          * we need to drop and relock the lnet_net_lock below.
1754          */
1755         lnet_peer_ni_addref_locked(best_lpni);
1756
1757         /*
1758          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1759          * message. This ensures that we get a CPT that is correct for
1760          * the NI when the NI has been restricted to a subset of all CPTs.
1761          * If the selected CPT differs from the one currently locked, we
1762          * must unlock and relock the lnet_net_lock(), and then check whether
1763          * the configuration has changed. We don't have a hold on the best_ni
1764          * yet, and it may have vanished.
1765          */
1766         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1767         if (sd->sd_cpt != cpt2) {
1768                 __u32 seq = lnet_get_dlc_seq_locked();
1769                 lnet_net_unlock(sd->sd_cpt);
1770                 sd->sd_cpt = cpt2;
1771                 lnet_net_lock(sd->sd_cpt);
1772                 if (seq != lnet_get_dlc_seq_locked()) {
1773                         lnet_peer_ni_decref_locked(best_lpni);
1774                         return REPEAT_SEND;
1775                 }
1776         }
1777
1778         /*
1779          * store the best_lpni in the message right away to avoid having
1780          * to do the same operation under different conditions
1781          */
1782         msg->msg_txpeer = best_lpni;
1783         msg->msg_txni = best_ni;
1784
1785         /*
1786          * grab a reference for the best_ni since now it's in use in this
1787          * send. The reference will be dropped in lnet_finalize()
1788          */
1789         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1790
1791         /*
1792          * Always set the target.nid to the best peer picked. Either the
1793          * NID will be one of the peer NIDs selected, or the same NID as
1794          * what was originally set in the target or it will be the NID of
1795          * a router if this message should be routed
1796          */
1797         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1798
1799         /*
1800          * lnet_msg_commit assigns the correct cpt to the message, which
1801          * is used to decrement the correct refcount on the ni when it's
1802          * time to return the credits
1803          */
1804         lnet_msg_commit(msg, sd->sd_cpt);
1805
1806         /*
1807          * If we are routing the message then we keep the src_nid that was
1808          * set by the originator. If we are not routing then we are the
1809          * originator and set it here.
1810          */
1811         if (!msg->msg_routing)
1812                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1813
1814         if (routing) {
1815                 msg->msg_target_is_router = 1;
1816                 msg->msg_target.pid = LNET_PID_LUSTRE;
1817                 /*
1818                  * since we're routing we want to ensure that the
1819                  * msg_hdr.dest_nid is set to the final destination. When
1820                  * the router receives this message it knows how to route
1821                  * it.
1822                  *
1823                  * final_dst_lpni is set at the beginning of the
1824                  * lnet_select_pathway() function and is never changed.
1825                  * It's safe to use it here.
1826                  */
1827                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1828         } else {
1829                 /*
1830                  * if we're not routing set the dest_nid to the best peer
1831                  * ni NID that we picked earlier in the algorithm.
1832                  */
1833                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1834         }
1835
1836         /*
1837          * if we have response tracker block update it with the next hop
1838          * nid
1839          */
1840         if (msg->msg_md) {
1841                 rspt = msg->msg_md->md_rspt_ptr;
1842                 if (rspt) {
1843                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1844                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1845                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1846                 }
1847         }
1848
1849         rc = lnet_post_send_locked(msg, 0);
1850
1851         if (!rc)
1852                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
1853                        libcfs_nid2str(msg->msg_hdr.src_nid),
1854                        libcfs_nid2str(msg->msg_txni->ni_nid),
1855                        libcfs_nid2str(sd->sd_src_nid),
1856                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1857                        libcfs_nid2str(sd->sd_dst_nid),
1858                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1859                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1860
1861         return rc;
1862 }
1863
1864 static inline void
1865 lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
1866 {
1867         if (sd->sd_send_case & NMR_DST &&
1868             sd->sd_msg->msg_type != LNET_MSG_REPLY &&
1869             sd->sd_msg->msg_type != LNET_MSG_ACK &&
1870             sd->sd_best_lpni->lpni_pref_nnids == 0) {
1871                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1872                        libcfs_nid2str(sd->sd_best_ni->ni_nid),
1873                        libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
1874                 lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
1875                                                  sd->sd_best_ni->ni_nid);
1876         }
1877 }
1878
1879 /*
1880  * Source Specified
1881  * Local Destination
1882  * non-mr peer
1883  *
1884  * use the source and destination NIDs as the pathway
1885  */
1886 static int
1887 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1888 {
1889         /* the destination lpni is set before we get here. */
1890
1891         /* find local NI */
1892         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1893         if (!sd->sd_best_ni) {
1894                 CERROR("Can't send to %s: src %s is not a "
1895                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1896                                 libcfs_nid2str(sd->sd_src_nid));
1897                 return -EINVAL;
1898         }
1899
1900         /*
1901          * the preferred NID will only be set for NMR peers
1902          */
1903         lnet_set_non_mr_pref_nid(sd);
1904
1905         return lnet_handle_send(sd);
1906 }
1907
1908 /*
1909  * Source Specified
1910  * Local Destination
1911  * MR Peer
1912  *
1913  * Run the selection algorithm on the peer NIs unless we're sending
1914  * a response, in this case just send to the destination
1915  */
1916 static int
1917 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1918 {
1919         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1920         if (!sd->sd_best_ni) {
1921                 CERROR("Can't send to %s: src %s is not a "
1922                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1923                                 libcfs_nid2str(sd->sd_src_nid));
1924                 return -EINVAL;
1925         }
1926
1927         /*
1928          * only run the selection algorithm to pick the peer_ni if we're
1929          * sending a GET or a PUT. Responses are sent to the same
1930          * destination NID provided.
1931          */
1932         if (!(sd->sd_send_case & SND_RESP)) {
1933                 sd->sd_best_lpni =
1934                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
1935                                              sd->sd_best_ni->ni_net->net_id);
1936         }
1937
1938         if (sd->sd_best_lpni &&
1939             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1940                 return lnet_handle_lo_send(sd);
1941         else if (sd->sd_best_lpni)
1942                 return lnet_handle_send(sd);
1943
1944         CERROR("can't send to %s. no NI on %s\n",
1945                libcfs_nid2str(sd->sd_dst_nid),
1946                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1947
1948         return -EHOSTUNREACH;
1949 }
1950
1951 struct lnet_ni *
1952 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1953                               struct lnet_peer *peer,
1954                               struct lnet_peer_net *peer_net,
1955                               int cpt,
1956                               bool incr_seq)
1957 {
1958         struct lnet_net *local_net;
1959         struct lnet_ni *best_ni;
1960
1961         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1962         if (!local_net)
1963                 return NULL;
1964
1965         /*
1966          * Iterate through the NIs in this local Net and select
1967          * the NI to send from. The selection is determined by
1968          * these 3 criterion in the following priority:
1969          *      1. NUMA
1970          *      2. NI available credits
1971          *      3. Round Robin
1972          */
1973         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1974                                    peer, peer_net, cpt);
1975
1976         if (incr_seq && best_ni)
1977                 best_ni->ni_seq++;
1978
1979         return best_ni;
1980 }
1981
1982 static int
1983 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
1984                              struct lnet_msg *msg, lnet_nid_t rtr_nid,
1985                              int cpt)
1986 {
1987         struct lnet_peer *peer;
1988         lnet_nid_t primary_nid;
1989         int rc;
1990
1991         lnet_peer_ni_addref_locked(lpni);
1992
1993         peer = lpni->lpni_peer_net->lpn_peer;
1994
1995         if (lnet_peer_gw_discovery(peer)) {
1996                 lnet_peer_ni_decref_locked(lpni);
1997                 return 0;
1998         }
1999
2000         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2001                 lnet_peer_ni_decref_locked(lpni);
2002                 return 0;
2003         }
2004
2005         rc = lnet_discover_peer_locked(lpni, cpt, false);
2006         if (rc) {
2007                 lnet_peer_ni_decref_locked(lpni);
2008                 return rc;
2009         }
2010         /* The peer may have changed. */
2011         peer = lpni->lpni_peer_net->lpn_peer;
2012         /* queue message and return */
2013         msg->msg_rtr_nid_param = rtr_nid;
2014         msg->msg_sending = 0;
2015         msg->msg_txpeer = NULL;
2016         spin_lock(&peer->lp_lock);
2017         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
2018         spin_unlock(&peer->lp_lock);
2019         lnet_peer_ni_decref_locked(lpni);
2020         primary_nid = peer->lp_primary_nid;
2021
2022         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2023                 msg, libcfs_nid2str(primary_nid));
2024
2025         return LNET_DC_WAIT;
2026 }
2027
2028 static int
2029 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2030                              lnet_nid_t dst_nid,
2031                              struct lnet_peer_ni **gw_lpni,
2032                              struct lnet_peer **gw_peer)
2033 {
2034         int rc;
2035         struct lnet_peer *gw;
2036         struct lnet_peer *lp;
2037         struct lnet_peer_net *lpn;
2038         struct lnet_peer_net *best_lpn = NULL;
2039         struct lnet_remotenet *rnet;
2040         struct lnet_route *best_route;
2041         struct lnet_route *last_route;
2042         struct lnet_peer_ni *lpni = NULL;
2043         struct lnet_peer_ni *gwni = NULL;
2044         lnet_nid_t src_nid = sd->sd_src_nid;
2045
2046         /* we've already looked up the initial lpni using dst_nid */
2047         lpni = sd->sd_best_lpni;
2048         /* the peer tree must be in existence */
2049         LASSERT(lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
2050         lp = lpni->lpni_peer_net->lpn_peer;
2051
2052         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2053                 /* is this remote network reachable?  */
2054                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2055                 if (!rnet)
2056                         continue;
2057
2058                 if (!best_lpn)
2059                         best_lpn = lpn;
2060
2061                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2062                         continue;
2063
2064                 best_lpn = lpn;
2065         }
2066
2067         if (!best_lpn) {
2068                 CERROR("peer %s has no available nets \n",
2069                        libcfs_nid2str(sd->sd_dst_nid));
2070                 return -EHOSTUNREACH;
2071         }
2072
2073         sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd, lp, best_lpn->lpn_net_id);
2074         if (!sd->sd_best_lpni) {
2075                 CERROR("peer %s down\n", libcfs_nid2str(sd->sd_dst_nid));
2076                 return -EHOSTUNREACH;
2077         }
2078
2079         best_route = lnet_find_route_locked(NULL, best_lpn->lpn_net_id,
2080                                             sd->sd_rtr_nid, &last_route,
2081                                             &gwni);
2082         if (!best_route) {
2083                 CERROR("no route to %s from %s\n",
2084                        libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
2085                 return -EHOSTUNREACH;
2086         }
2087
2088         if (!gwni) {
2089                 CERROR("Internal Error. Route expected to %s from %s\n",
2090                         libcfs_nid2str(dst_nid),
2091                         libcfs_nid2str(src_nid));
2092                 return -EFAULT;
2093         }
2094
2095         gw = best_route->lr_gateway;
2096         LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2097
2098         /*
2099          * Discover this gateway if it hasn't already been discovered.
2100          * This means we might delay the message until discovery has
2101          * completed
2102          */
2103         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2104         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_rtr_nid,
2105                                           sd->sd_cpt);
2106         if (rc)
2107                 return rc;
2108
2109         if (!sd->sd_best_ni)
2110                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2111                                         lnet_peer_get_net_locked(gw,
2112                                                 best_route->lr_lnet),
2113                                         sd->sd_md_cpt,
2114                                         true);
2115
2116         if (!sd->sd_best_ni) {
2117                 CERROR("Internal Error. Expected local ni on %s "
2118                        "but non found :%s\n",
2119                        libcfs_net2str(best_route->lr_lnet),
2120                        libcfs_nid2str(sd->sd_src_nid));
2121                 return -EFAULT;
2122         }
2123
2124         *gw_lpni = gwni;
2125         *gw_peer = gw;
2126
2127         /*
2128          * increment the sequence numbers since now we're sure we're
2129          * going to use this path
2130          */
2131         LASSERT(best_route && last_route);
2132         best_route->lr_seq = last_route->lr_seq + 1;
2133         best_lpn->lpn_seq++;
2134
2135         return 0;
2136 }
2137
2138 /*
2139  * Handle two cases:
2140  *
2141  * Case 1:
2142  *  Source specified
2143  *  Remote destination
2144  *  Non-MR destination
2145  *
2146  * Case 2:
2147  *  Source specified
2148  *  Remote destination
2149  *  MR destination
2150  *
2151  * The handling of these two cases is similar. Even though the destination
2152  * can be MR or non-MR, we'll deal directly with the router.
2153  */
2154 static int
2155 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2156 {
2157         int rc;
2158         struct lnet_peer_ni *gw_lpni = NULL;
2159         struct lnet_peer *gw_peer = NULL;
2160
2161         /* find local NI */
2162         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2163         if (!sd->sd_best_ni) {
2164                 CERROR("Can't send to %s: src %s is not a "
2165                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2166                                 libcfs_nid2str(sd->sd_src_nid));
2167                 return -EINVAL;
2168         }
2169
2170         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2171                                      &gw_peer);
2172         if (rc)
2173                 return rc;
2174
2175         if (sd->sd_send_case & NMR_DST)
2176                 /*
2177                 * since the final destination is non-MR let's set its preferred
2178                 * NID before we send
2179                 */
2180                 lnet_set_non_mr_pref_nid(sd);
2181
2182         /*
2183          * We're going to send to the gw found so let's set its
2184          * info
2185          */
2186         sd->sd_peer = gw_peer;
2187         sd->sd_best_lpni = gw_lpni;
2188
2189         return lnet_handle_send(sd);
2190 }
2191
2192 struct lnet_ni *
2193 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2194                                bool discovery)
2195 {
2196         struct lnet_peer_net *peer_net = NULL;
2197         struct lnet_ni *best_ni = NULL;
2198
2199         /*
2200          * The peer can have multiple interfaces, some of them can be on
2201          * the local network and others on a routed network. We should
2202          * prefer the local network. However if the local network is not
2203          * available then we need to try the routed network
2204          */
2205
2206         /* go through all the peer nets and find the best_ni */
2207         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2208                 /*
2209                  * The peer's list of nets can contain non-local nets. We
2210                  * want to only examine the local ones.
2211                  */
2212                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2213                         continue;
2214                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
2215                                                    peer_net, md_cpt, false);
2216
2217                 /*
2218                  * if this is a discovery message and lp_disc_net_id is
2219                  * specified then use that net to send the discovery on.
2220                  */
2221                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2222                     discovery)
2223                         break;
2224         }
2225
2226         if (best_ni)
2227                 /* increment sequence number so we can round robin */
2228                 best_ni->ni_seq++;
2229
2230         return best_ni;
2231 }
2232
2233 static struct lnet_ni *
2234 lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
2235 {
2236         struct lnet_ni *best_ni = NULL;
2237         struct lnet_peer_net *peer_net;
2238         struct lnet_peer *peer = sd->sd_peer;
2239         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2240         struct lnet_peer_ni *lpni;
2241         int cpt = sd->sd_cpt;
2242
2243         /*
2244          * We must use a consistent source address when sending to a
2245          * non-MR peer. However, a non-MR peer can have multiple NIDs
2246          * on multiple networks, and we may even need to talk to this
2247          * peer on multiple networks -- certain types of
2248          * load-balancing configuration do this.
2249          *
2250          * So we need to pick the NI the peer prefers for this
2251          * particular network.
2252          */
2253
2254         /* Get the target peer_ni */
2255         peer_net = lnet_peer_get_net_locked(peer,
2256                         LNET_NIDNET(best_lpni->lpni_nid));
2257         LASSERT(peer_net != NULL);
2258         list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
2259                                 lpni_peer_nis) {
2260                 if (lpni->lpni_pref_nnids == 0)
2261                         continue;
2262                 LASSERT(lpni->lpni_pref_nnids == 1);
2263                 best_ni = lnet_nid2ni_locked(
2264                                 lpni->lpni_pref.nid, cpt);
2265                 break;
2266         }
2267
2268         return best_ni;
2269 }
2270
2271 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2272 static int
2273 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2274 {
2275         struct lnet_ni *best_ni = NULL;
2276         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2277
2278         /*
2279          * We must use a consistent source address when sending to a
2280          * non-MR peer. However, a non-MR peer can have multiple NIDs
2281          * on multiple networks, and we may even need to talk to this
2282          * peer on multiple networks -- certain types of
2283          * load-balancing configuration do this.
2284          *
2285          * So we need to pick the NI the peer prefers for this
2286          * particular network.
2287          */
2288
2289         best_ni = lnet_find_existing_preferred_best_ni(sd);
2290
2291         /* if best_ni is still not set just pick one */
2292         if (!best_ni) {
2293                 best_ni =
2294                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2295                                                 sd->sd_best_lpni->lpni_peer_net,
2296                                                 sd->sd_md_cpt, true);
2297                 /* If there is no best_ni we don't have a route */
2298                 if (!best_ni) {
2299                         CERROR("no path to %s from net %s\n",
2300                                 libcfs_nid2str(best_lpni->lpni_nid),
2301                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2302                         return -EHOSTUNREACH;
2303                 }
2304         }
2305
2306         sd->sd_best_ni = best_ni;
2307
2308         /* Set preferred NI if necessary. */
2309         lnet_set_non_mr_pref_nid(sd);
2310
2311         return 0;
2312 }
2313
2314
2315 /*
2316  * Source not specified
2317  * Local destination
2318  * Non-MR Peer
2319  *
2320  * always use the same source NID for NMR peers
2321  * If we've talked to that peer before then we already have a preferred
2322  * source NI associated with it. Otherwise, we select a preferred local NI
2323  * and store it in the peer
2324  */
2325 static int
2326 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2327 {
2328         int rc;
2329
2330         /* sd->sd_best_lpni is already set to the final destination */
2331
2332         /*
2333          * At this point we should've created the peer ni and peer. If we
2334          * can't find it, then something went wrong. Instead of assert
2335          * output a relevant message and fail the send
2336          */
2337         if (!sd->sd_best_lpni) {
2338                 CERROR("Internal fault. Unable to send msg %s to %s. "
2339                        "NID not known\n",
2340                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2341                        libcfs_nid2str(sd->sd_dst_nid));
2342                 return -EFAULT;
2343         }
2344
2345         rc = lnet_select_preferred_best_ni(sd);
2346         if (!rc)
2347                 rc = lnet_handle_send(sd);
2348
2349         return rc;
2350 }
2351
2352 static int
2353 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2354 {
2355         /*
2356          * NOTE we've already handled the remote peer case. So we only
2357          * need to worry about the local case here.
2358          *
2359          * if we're sending a response, ACK or reply, we need to send it
2360          * to the destination NID given to us. At this point we already
2361          * have the peer_ni we're suppose to send to, so just find the
2362          * best_ni on the peer net and use that. Since we're sending to an
2363          * MR peer then we can just run the selection algorithm on our
2364          * local NIs and pick the best one.
2365          */
2366         if (sd->sd_send_case & SND_RESP) {
2367                 sd->sd_best_ni =
2368                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2369                                                 sd->sd_best_lpni->lpni_peer_net,
2370                                                 sd->sd_md_cpt, true);
2371
2372                 if (!sd->sd_best_ni) {
2373                         /*
2374                          * We're not going to deal with not able to send
2375                          * a response to the provided final destination
2376                          */
2377                         CERROR("Can't send response to %s. "
2378                                "No local NI available\n",
2379                                 libcfs_nid2str(sd->sd_dst_nid));
2380                         return -EHOSTUNREACH;
2381                 }
2382
2383                 return lnet_handle_send(sd);
2384         }
2385
2386         /*
2387          * If we get here that means we're sending a fresh request, PUT or
2388          * GET, so we need to run our standard selection algorithm.
2389          * First find the best local interface that's on any of the peer's
2390          * networks.
2391          */
2392         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2393                                         sd->sd_md_cpt,
2394                                         lnet_msg_discovery(sd->sd_msg));
2395         if (sd->sd_best_ni) {
2396                 sd->sd_best_lpni =
2397                   lnet_find_best_lpni_on_net(sd, sd->sd_peer,
2398                                              sd->sd_best_ni->ni_net->net_id);
2399
2400                 /*
2401                  * if we're successful in selecting a peer_ni on the local
2402                  * network, then send to it. Otherwise fall through and
2403                  * try and see if we can reach it over another routed
2404                  * network
2405                  */
2406                 if (sd->sd_best_lpni &&
2407                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2408                         /*
2409                          * in case we initially started with a routed
2410                          * destination, let's reset to local
2411                          */
2412                         sd->sd_send_case &= ~REMOTE_DST;
2413                         sd->sd_send_case |= LOCAL_DST;
2414                         return lnet_handle_lo_send(sd);
2415                 } else if (sd->sd_best_lpni) {
2416                         /*
2417                          * in case we initially started with a routed
2418                          * destination, let's reset to local
2419                          */
2420                         sd->sd_send_case &= ~REMOTE_DST;
2421                         sd->sd_send_case |= LOCAL_DST;
2422                         return lnet_handle_send(sd);
2423                 }
2424
2425                 CERROR("Internal Error. Expected to have a best_lpni: "
2426                        "%s -> %s\n",
2427                        libcfs_nid2str(sd->sd_src_nid),
2428                        libcfs_nid2str(sd->sd_dst_nid));
2429
2430                 return -EFAULT;
2431         }
2432
2433         /*
2434          * Peer doesn't have a local network. Let's see if there is
2435          * a remote network we can reach it on.
2436          */
2437         return PASS_THROUGH;
2438 }
2439
2440 /*
2441  * Case 1:
2442  *      Source NID not specified
2443  *      Local destination
2444  *      MR peer
2445  *
2446  * Case 2:
2447  *      Source NID not speified
2448  *      Remote destination
2449  *      MR peer
2450  *
2451  * In both of these cases if we're sending a response, ACK or REPLY, then
2452  * we need to send to the destination NID provided.
2453  *
2454  * In the remote case let's deal with MR routers.
2455  *
2456  */
2457
2458 static int
2459 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2460 {
2461         int rc = 0;
2462         struct lnet_peer *gw_peer = NULL;
2463         struct lnet_peer_ni *gw_lpni = NULL;
2464
2465         /*
2466          * handle sending a response to a remote peer here so we don't
2467          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2468          */
2469         if (sd->sd_send_case & REMOTE_DST &&
2470             sd->sd_send_case & SND_RESP) {
2471                 struct lnet_peer_ni *gw;
2472                 struct lnet_peer *gw_peer;
2473
2474                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2475                                                   &gw_peer);
2476                 if (rc < 0) {
2477                         CERROR("Can't send response to %s. "
2478                                "No route available\n",
2479                                 libcfs_nid2str(sd->sd_dst_nid));
2480                         return -EHOSTUNREACH;
2481                 } else if (rc > 0) {
2482                         return rc;
2483                 }
2484
2485                 sd->sd_best_lpni = gw;
2486                 sd->sd_peer = gw_peer;
2487
2488                 return lnet_handle_send(sd);
2489         }
2490
2491         /*
2492          * Even though the NID for the peer might not be on a local network,
2493          * since the peer is MR there could be other interfaces on the
2494          * local network. In that case we'd still like to prefer the local
2495          * network over the routed network. If we're unable to do that
2496          * then we select the best router among the different routed networks,
2497          * and if the router is MR then we can deal with it as such.
2498          */
2499         rc = lnet_handle_any_mr_dsta(sd);
2500         if (rc != PASS_THROUGH)
2501                 return rc;
2502
2503         /*
2504          * Now that we must route to the destination, we must consider the
2505          * MR case, where the destination has multiple interfaces, some of
2506          * which we can route to and others we do not. For this reason we
2507          * need to select the destination which we can route to and if
2508          * there are multiple, we need to round robin.
2509          */
2510         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2511                                           &gw_peer);
2512         if (rc)
2513                 return rc;
2514
2515         sd->sd_send_case &= ~LOCAL_DST;
2516         sd->sd_send_case |= REMOTE_DST;
2517
2518         sd->sd_peer = gw_peer;
2519         sd->sd_best_lpni = gw_lpni;
2520
2521         return lnet_handle_send(sd);
2522 }
2523
2524 /*
2525  * Source not specified
2526  * Remote destination
2527  * Non-MR peer
2528  *
2529  * Must send to the specified peer NID using the same source NID that
2530  * we've used before. If it's the first time to talk to that peer then
2531  * find the source NI and assign it as preferred to that peer
2532  */
2533 static int
2534 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2535 {
2536         int rc;
2537         struct lnet_peer_ni *gw_lpni = NULL;
2538         struct lnet_peer *gw_peer = NULL;
2539
2540         /*
2541          * Let's set if we have a preferred NI to talk to this NMR peer
2542          */
2543         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
2544
2545         /*
2546          * find the router and that'll find the best NI if we didn't find
2547          * it already.
2548          */
2549         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2550                                           &gw_peer);
2551         if (rc)
2552                 return rc;
2553
2554         /*
2555          * set the best_ni we've chosen as the preferred one for
2556          * this peer
2557          */
2558         lnet_set_non_mr_pref_nid(sd);
2559
2560         /* we'll be sending to the gw */
2561         sd->sd_best_lpni = gw_lpni;
2562         sd->sd_peer = gw_peer;
2563
2564         return lnet_handle_send(sd);
2565 }
2566
2567 static int
2568 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2569 {
2570         /*
2571          * turn off the SND_RESP bit.
2572          * It will be checked in the case handling
2573          */
2574         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2575
2576         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2577                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2578                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2579                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2580                 libcfs_nid2str(sd->sd_dst_nid),
2581                 (send_case & LOCAL_DST) ? "local" : "routed");
2582
2583         switch (send_case) {
2584         /*
2585          * For all cases where the source is specified, we should always
2586          * use the destination NID, whether it's an MR destination or not,
2587          * since we're continuing a series of related messages for the
2588          * same RPC
2589          */
2590         case SRC_SPEC_LOCAL_NMR_DST:
2591                 return lnet_handle_spec_local_nmr_dst(sd);
2592         case SRC_SPEC_LOCAL_MR_DST:
2593                 return lnet_handle_spec_local_mr_dst(sd);
2594         case SRC_SPEC_ROUTER_NMR_DST:
2595         case SRC_SPEC_ROUTER_MR_DST:
2596                 return lnet_handle_spec_router_dst(sd);
2597         case SRC_ANY_LOCAL_NMR_DST:
2598                 return lnet_handle_any_local_nmr_dst(sd);
2599         case SRC_ANY_LOCAL_MR_DST:
2600         case SRC_ANY_ROUTER_MR_DST:
2601                 return lnet_handle_any_mr_dst(sd);
2602         case SRC_ANY_ROUTER_NMR_DST:
2603                 return lnet_handle_any_router_nmr_dst(sd);
2604         default:
2605                 CERROR("Unknown send case\n");
2606                 return -1;
2607         }
2608 }
2609
2610 static int
2611 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2612                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2613 {
2614         struct lnet_peer_ni     *lpni;
2615         struct lnet_peer        *peer;
2616         struct lnet_send_data   send_data;
2617         int                     cpt, rc;
2618         int                     md_cpt;
2619         __u32                   send_case = 0;
2620
2621         memset(&send_data, 0, sizeof(send_data));
2622
2623         /*
2624          * get an initial CPT to use for locking. The idea here is not to
2625          * serialize the calls to select_pathway, so that as many
2626          * operations can run concurrently as possible. To do that we use
2627          * the CPT where this call is being executed. Later on when we
2628          * determine the CPT to use in lnet_message_commit, we switch the
2629          * lock and check if there was any configuration change.  If none,
2630          * then we proceed, if there is, then we restart the operation.
2631          */
2632         cpt = lnet_net_lock_current();
2633
2634         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2635         if (md_cpt == CFS_CPT_ANY)
2636                 md_cpt = cpt;
2637
2638 again:
2639
2640         /*
2641          * If we're being asked to send to the loopback interface, there
2642          * is no need to go through any selection. We can just shortcut
2643          * the entire process and send over lolnd
2644          */
2645         send_data.sd_msg = msg;
2646         send_data.sd_cpt = cpt;
2647         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2648                 rc = lnet_handle_lo_send(&send_data);
2649                 lnet_net_unlock(cpt);
2650                 return rc;
2651         }
2652
2653         /*
2654          * find an existing peer_ni, or create one and mark it as having been
2655          * created due to network traffic. This call will create the
2656          * peer->peer_net->peer_ni tree.
2657          */
2658         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2659         if (IS_ERR(lpni)) {
2660                 lnet_net_unlock(cpt);
2661                 return PTR_ERR(lpni);
2662         }
2663
2664         /*
2665          * Cache the original src_nid. If we need to resend the message
2666          * then we'll need to know whether the src_nid was originally
2667          * specified for this message. If it was originally specified,
2668          * then we need to keep using the same src_nid since it's
2669          * continuing the same sequence of messages.
2670          */
2671         msg->msg_src_nid_param = src_nid;
2672
2673         /*
2674          * Now that we have a peer_ni, check if we want to discover
2675          * the peer. Traffic to the LNET_RESERVED_PORTAL should not
2676          * trigger discovery.
2677          */
2678         peer = lpni->lpni_peer_net->lpn_peer;
2679         rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
2680         if (rc) {
2681                 lnet_peer_ni_decref_locked(lpni);
2682                 lnet_net_unlock(cpt);
2683                 return rc;
2684         }
2685         lnet_peer_ni_decref_locked(lpni);
2686
2687         /*
2688          * Identify the different send cases
2689          */
2690         if (src_nid == LNET_NID_ANY)
2691                 send_case |= SRC_ANY;
2692         else
2693                 send_case |= SRC_SPEC;
2694
2695         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2696                 send_case |= LOCAL_DST;
2697         else
2698                 send_case |= REMOTE_DST;
2699
2700         /*
2701          * if this is a non-MR peer or if we're recovering a peer ni then
2702          * let's consider this an NMR case so we can hit the destination
2703          * NID.
2704          */
2705         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
2706                 send_case |= NMR_DST;
2707         else
2708                 send_case |= MR_DST;
2709
2710         if (msg->msg_type == LNET_MSG_REPLY ||
2711             msg->msg_type == LNET_MSG_ACK)
2712                 send_case |= SND_RESP;
2713
2714         /* assign parameters to the send_data */
2715         send_data.sd_rtr_nid = rtr_nid;
2716         send_data.sd_src_nid = src_nid;
2717         send_data.sd_dst_nid = dst_nid;
2718         send_data.sd_best_lpni = lpni;
2719         /*
2720          * keep a pointer to the final destination in case we're going to
2721          * route, so we'll need to access it later
2722          */
2723         send_data.sd_final_dst_lpni = lpni;
2724         send_data.sd_peer = peer;
2725         send_data.sd_md_cpt = md_cpt;
2726         send_data.sd_send_case = send_case;
2727
2728         rc = lnet_handle_send_case_locked(&send_data);
2729
2730         /*
2731          * Update the local cpt since send_data.sd_cpt might've been
2732          * updated as a result of calling lnet_handle_send_case_locked().
2733          */
2734         cpt = send_data.sd_cpt;
2735
2736         if (rc == REPEAT_SEND)
2737                 goto again;
2738
2739         lnet_net_unlock(cpt);
2740
2741         return rc;
2742 }
2743
2744 int
2745 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2746 {
2747         lnet_nid_t              dst_nid = msg->msg_target.nid;
2748         int                     rc;
2749
2750         /*
2751          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2752          * but we might want to use pre-determined router for ACK/REPLY
2753          * in the future
2754          */
2755         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2756         LASSERT(msg->msg_txpeer == NULL);
2757         LASSERT(msg->msg_txni == NULL);
2758         LASSERT(!msg->msg_sending);
2759         LASSERT(!msg->msg_target_is_router);
2760         LASSERT(!msg->msg_receiving);
2761
2762         msg->msg_sending = 1;
2763
2764         LASSERT(!msg->msg_tx_committed);
2765
2766         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2767         if (rc < 0) {
2768                 if (rc == -EHOSTUNREACH)
2769                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2770                 else
2771                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2772                 return rc;
2773         }
2774
2775         if (rc == LNET_CREDIT_OK)
2776                 lnet_ni_send(msg->msg_txni, msg);
2777
2778         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2779         return 0;
2780 }
2781
2782 enum lnet_mt_event_type {
2783         MT_TYPE_LOCAL_NI = 0,
2784         MT_TYPE_PEER_NI
2785 };
2786
2787 struct lnet_mt_event_info {
2788         enum lnet_mt_event_type mt_type;
2789         lnet_nid_t mt_nid;
2790 };
2791
2792 /* called with res_lock held */
2793 void
2794 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2795 {
2796         struct lnet_rsp_tracker *rspt;
2797
2798         /*
2799          * msg has a refcount on the MD so the MD is not going away.
2800          * The rspt queue for the cpt is protected by
2801          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2802          */
2803         if (!md->md_rspt_ptr)
2804                 return;
2805
2806         rspt = md->md_rspt_ptr;
2807
2808         /* debug code */
2809         LASSERT(rspt->rspt_cpt == cpt);
2810
2811         md->md_rspt_ptr = NULL;
2812
2813         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2814                 /*
2815                  * The monitor thread has invalidated this handle because the
2816                  * response timed out, but it failed to lookup the MD. That
2817                  * means this response tracker is on the zombie list. We can
2818                  * safely remove it under the resource lock (held by caller) and
2819                  * free the response tracker block.
2820                  */
2821                 list_del(&rspt->rspt_on_list);
2822                 lnet_rspt_free(rspt, cpt);
2823         } else {
2824                 /*
2825                  * invalidate the handle to indicate that a response has been
2826                  * received, which will then lead the monitor thread to clean up
2827                  * the rspt block.
2828                  */
2829                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2830         }
2831 }
2832
2833 void
2834 lnet_clean_zombie_rstqs(void)
2835 {
2836         struct lnet_rsp_tracker *rspt, *tmp;
2837         int i;
2838
2839         cfs_cpt_for_each(i, lnet_cpt_table()) {
2840                 list_for_each_entry_safe(rspt, tmp,
2841                                          the_lnet.ln_mt_zombie_rstqs[i],
2842                                          rspt_on_list) {
2843                         list_del(&rspt->rspt_on_list);
2844                         lnet_rspt_free(rspt, i);
2845                 }
2846         }
2847
2848         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2849 }
2850
2851 static void
2852 lnet_finalize_expired_responses(void)
2853 {
2854         struct lnet_libmd *md;
2855         struct list_head local_queue;
2856         struct lnet_rsp_tracker *rspt, *tmp;
2857         ktime_t now;
2858         int i;
2859
2860         if (the_lnet.ln_mt_rstq == NULL)
2861                 return;
2862
2863         cfs_cpt_for_each(i, lnet_cpt_table()) {
2864                 INIT_LIST_HEAD(&local_queue);
2865
2866                 lnet_net_lock(i);
2867                 if (!the_lnet.ln_mt_rstq[i]) {
2868                         lnet_net_unlock(i);
2869                         continue;
2870                 }
2871                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2872                 lnet_net_unlock(i);
2873
2874                 now = ktime_get();
2875
2876                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2877                         /*
2878                          * The rspt mdh will be invalidated when a response
2879                          * is received or whenever we want to discard the
2880                          * block the monitor thread will walk the queue
2881                          * and clean up any rsts with an invalid mdh.
2882                          * The monitor thread will walk the queue until
2883                          * the first unexpired rspt block. This means that
2884                          * some rspt blocks which received their
2885                          * corresponding responses will linger in the
2886                          * queue until they are cleaned up eventually.
2887                          */
2888                         lnet_res_lock(i);
2889                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2890                                 lnet_res_unlock(i);
2891                                 list_del(&rspt->rspt_on_list);
2892                                 lnet_rspt_free(rspt, i);
2893                                 continue;
2894                         }
2895
2896                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2897                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2898                                 struct lnet_peer_ni *lpni;
2899                                 lnet_nid_t nid;
2900
2901                                 md = lnet_handle2md(&rspt->rspt_mdh);
2902                                 if (!md) {
2903                                         /* MD has been queued for unlink, but
2904                                          * rspt hasn't been detached (Note we've
2905                                          * checked above that the rspt_mdh is
2906                                          * valid). Since we cannot lookup the MD
2907                                          * we're unable to detach the rspt
2908                                          * ourselves. Thus, move the rspt to the
2909                                          * zombie list where we'll wait for
2910                                          * either:
2911                                          *   1. The remaining operations on the
2912                                          *   MD to complete. In this case the
2913                                          *   final operation will result in
2914                                          *   lnet_msg_detach_md()->
2915                                          *   lnet_detach_rsp_tracker() where
2916                                          *   we will clean up this response
2917                                          *   tracker.
2918                                          *   2. LNet to shutdown. In this case
2919                                          *   we'll wait until after all LND Nets
2920                                          *   have shutdown and then we can
2921                                          *   safely free any remaining response
2922                                          *   tracker blocks on the zombie list.
2923                                          * Note: We need to hold the resource
2924                                          * lock when adding to the zombie list
2925                                          * because we may have concurrent access
2926                                          * with lnet_detach_rsp_tracker().
2927                                          */
2928                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2929                                         list_move(&rspt->rspt_on_list,
2930                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2931                                         lnet_res_unlock(i);
2932                                         continue;
2933                                 }
2934                                 LASSERT(md->md_rspt_ptr == rspt);
2935                                 md->md_rspt_ptr = NULL;
2936                                 lnet_res_unlock(i);
2937
2938                                 LNetMDUnlink(rspt->rspt_mdh);
2939
2940                                 nid = rspt->rspt_next_hop_nid;
2941
2942                                 list_del(&rspt->rspt_on_list);
2943                                 lnet_rspt_free(rspt, i);
2944
2945                                 /* If we're shutting down we just want to clean
2946                                  * up the rspt blocks
2947                                  */
2948                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2949                                         continue;
2950
2951                                 lnet_net_lock(i);
2952                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2953                                 lnet_net_unlock(i);
2954
2955                                 CDEBUG(D_NET,
2956                                        "Response timeout: md = %p: nid = %s\n",
2957                                        md, libcfs_nid2str(nid));
2958
2959                                 /*
2960                                  * If there is a timeout on the response
2961                                  * from the next hop decrement its health
2962                                  * value so that we don't use it
2963                                  */
2964                                 lnet_net_lock(0);
2965                                 lpni = lnet_find_peer_ni_locked(nid);
2966                                 if (lpni) {
2967                                         lnet_handle_remote_failure_locked(lpni);
2968                                         lnet_peer_ni_decref_locked(lpni);
2969                                 }
2970                                 lnet_net_unlock(0);
2971                         } else {
2972                                 lnet_res_unlock(i);
2973                                 break;
2974                         }
2975                 }
2976
2977                 if (!list_empty(&local_queue)) {
2978                         lnet_net_lock(i);
2979                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2980                         lnet_net_unlock(i);
2981                 }
2982         }
2983 }
2984
2985 static void
2986 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2987 {
2988         struct lnet_msg *msg;
2989
2990         while (!list_empty(resendq)) {
2991                 struct lnet_peer_ni *lpni;
2992
2993                 msg = list_entry(resendq->next, struct lnet_msg,
2994                                  msg_list);
2995
2996                 list_del_init(&msg->msg_list);
2997
2998                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2999                 if (!lpni) {
3000                         lnet_net_unlock(cpt);
3001                         CERROR("Expected that a peer is already created for %s\n",
3002                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3003                         msg->msg_no_resend = true;
3004                         lnet_finalize(msg, -EFAULT);
3005                         lnet_net_lock(cpt);
3006                 } else {
3007                         struct lnet_peer *peer;
3008                         int rc;
3009                         lnet_nid_t src_nid = LNET_NID_ANY;
3010
3011                         /*
3012                          * if this message is not being routed and the
3013                          * peer is non-MR then we must use the same
3014                          * src_nid that was used in the original send.
3015                          * Otherwise if we're routing the message (IE
3016                          * we're a router) then we can use any of our
3017                          * local interfaces. It doesn't matter to the
3018                          * final destination.
3019                          */
3020                         peer = lpni->lpni_peer_net->lpn_peer;
3021                         if (!msg->msg_routing &&
3022                             !lnet_peer_is_multi_rail(peer))
3023                                 src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
3024
3025                         /*
3026                          * If we originally specified a src NID, then we
3027                          * must attempt to reuse it in the resend as well.
3028                          */
3029                         if (msg->msg_src_nid_param != LNET_NID_ANY)
3030                                 src_nid = msg->msg_src_nid_param;
3031                         lnet_peer_ni_decref_locked(lpni);
3032
3033                         lnet_net_unlock(cpt);
3034                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3035                                libcfs_nid2str(src_nid),
3036                                libcfs_id2str(msg->msg_target),
3037                                lnet_msgtyp2str(msg->msg_type),
3038                                msg->msg_recovery,
3039                                msg->msg_retry_count);
3040                         rc = lnet_send(src_nid, msg, LNET_NID_ANY);
3041                         if (rc) {
3042                                 CERROR("Error sending %s to %s: %d\n",
3043                                        lnet_msgtyp2str(msg->msg_type),
3044                                        libcfs_id2str(msg->msg_target), rc);
3045                                 msg->msg_no_resend = true;
3046                                 lnet_finalize(msg, rc);
3047                         }
3048                         lnet_net_lock(cpt);
3049                         if (!rc)
3050                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3051                 }
3052         }
3053 }
3054
3055 static void
3056 lnet_resend_pending_msgs(void)
3057 {
3058         int i;
3059
3060         cfs_cpt_for_each(i, lnet_cpt_table()) {
3061                 lnet_net_lock(i);
3062                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3063                 lnet_net_unlock(i);
3064         }
3065 }
3066
3067 /* called with cpt and ni_lock held */
3068 static void
3069 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3070 {
3071         struct lnet_handle_md recovery_mdh;
3072
3073         LNetInvalidateMDHandle(&recovery_mdh);
3074
3075         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3076             force) {
3077                 recovery_mdh = ni->ni_ping_mdh;
3078                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3079         }
3080         lnet_ni_unlock(ni);
3081         lnet_net_unlock(cpt);
3082         if (!LNetMDHandleIsInvalid(recovery_mdh))
3083                 LNetMDUnlink(recovery_mdh);
3084         lnet_net_lock(cpt);
3085         lnet_ni_lock(ni);
3086 }
3087
3088 static void
3089 lnet_recover_local_nis(void)
3090 {
3091         struct lnet_mt_event_info *ev_info;
3092         struct list_head processed_list;
3093         struct list_head local_queue;
3094         struct lnet_handle_md mdh;
3095         struct lnet_ni *tmp;
3096         struct lnet_ni *ni;
3097         lnet_nid_t nid;
3098         int healthv;
3099         int rc;
3100
3101         INIT_LIST_HEAD(&local_queue);
3102         INIT_LIST_HEAD(&processed_list);
3103
3104         /*
3105          * splice the recovery queue on a local queue. We will iterate
3106          * through the local queue and update it as needed. Once we're
3107          * done with the traversal, we'll splice the local queue back on
3108          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3109          * will be traversed in the next iteration.
3110          */
3111         lnet_net_lock(0);
3112         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3113                          &local_queue);
3114         lnet_net_unlock(0);
3115
3116         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3117                 /*
3118                  * if an NI is being deleted or it is now healthy, there
3119                  * is no need to keep it around in the recovery queue.
3120                  * The monitor thread is the only thread responsible for
3121                  * removing the NI from the recovery queue.
3122                  * Multiple threads can be adding NIs to the recovery
3123                  * queue.
3124                  */
3125                 healthv = atomic_read(&ni->ni_healthv);
3126
3127                 lnet_net_lock(0);
3128                 lnet_ni_lock(ni);
3129                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3130                     healthv == LNET_MAX_HEALTH_VALUE) {
3131                         list_del_init(&ni->ni_recovery);
3132                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3133                         lnet_ni_unlock(ni);
3134                         lnet_ni_decref_locked(ni, 0);
3135                         lnet_net_unlock(0);
3136                         continue;
3137                 }
3138
3139                 /*
3140                  * if the local NI failed recovery we must unlink the md.
3141                  * But we want to keep the local_ni on the recovery queue
3142                  * so we can continue the attempts to recover it.
3143                  */
3144                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3145                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3146                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3147                 }
3148
3149                 lnet_ni_unlock(ni);
3150                 lnet_net_unlock(0);
3151
3152
3153                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3154                        libcfs_nid2str(ni->ni_nid));
3155
3156                 lnet_ni_lock(ni);
3157                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3158                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3159                         lnet_ni_unlock(ni);
3160
3161                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3162                         if (!ev_info) {
3163                                 CERROR("out of memory. Can't recover %s\n",
3164                                        libcfs_nid2str(ni->ni_nid));
3165                                 lnet_ni_lock(ni);
3166                                 ni->ni_recovery_state &=
3167                                   ~LNET_NI_RECOVERY_PENDING;
3168                                 lnet_ni_unlock(ni);
3169                                 continue;
3170                         }
3171
3172                         mdh = ni->ni_ping_mdh;
3173                         /*
3174                          * Invalidate the ni mdh in case it's deleted.
3175                          * We'll unlink the mdh in this case below.
3176                          */
3177                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3178                         nid = ni->ni_nid;
3179
3180                         /*
3181                          * remove the NI from the local queue and drop the
3182                          * reference count to it while we're recovering
3183                          * it. The reason for that, is that the NI could
3184                          * be deleted, and the way the code is structured
3185                          * is if we don't drop the NI, then the deletion
3186                          * code will enter a loop waiting for the
3187                          * reference count to be removed while holding the
3188                          * ln_mutex_lock(). When we look up the peer to
3189                          * send to in lnet_select_pathway() we will try to
3190                          * lock the ln_mutex_lock() as well, leading to
3191                          * a deadlock. By dropping the refcount and
3192                          * removing it from the list, we allow for the NI
3193                          * to be removed, then we use the cached NID to
3194                          * look it up again. If it's gone, then we just
3195                          * continue examining the rest of the queue.
3196                          */
3197                         lnet_net_lock(0);
3198                         list_del_init(&ni->ni_recovery);
3199                         lnet_ni_decref_locked(ni, 0);
3200                         lnet_net_unlock(0);
3201
3202                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3203                         ev_info->mt_nid = nid;
3204                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3205                                             ev_info, the_lnet.ln_mt_eqh, true);
3206                         /* lookup the nid again */
3207                         lnet_net_lock(0);
3208                         ni = lnet_nid2ni_locked(nid, 0);
3209                         if (!ni) {
3210                                 /*
3211                                  * the NI has been deleted when we dropped
3212                                  * the ref count
3213                                  */
3214                                 lnet_net_unlock(0);
3215                                 LNetMDUnlink(mdh);
3216                                 continue;
3217                         }
3218                         /*
3219                          * Same note as in lnet_recover_peer_nis(). When
3220                          * we're sending the ping, the NI is free to be
3221                          * deleted or manipulated. By this point it
3222                          * could've been added back on the recovery queue,
3223                          * and a refcount taken on it.
3224                          * So we can't just add it blindly again or we'll
3225                          * corrupt the queue. We must check under lock if
3226                          * it's not on any list and if not then add it
3227                          * to the processed list, which will eventually be
3228                          * spliced back on to the recovery queue.
3229                          */
3230                         ni->ni_ping_mdh = mdh;
3231                         if (list_empty(&ni->ni_recovery)) {
3232                                 list_add_tail(&ni->ni_recovery, &processed_list);
3233                                 lnet_ni_addref_locked(ni, 0);
3234                         }
3235                         lnet_net_unlock(0);
3236
3237                         lnet_ni_lock(ni);
3238                         if (rc)
3239                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3240                 }
3241                 lnet_ni_unlock(ni);
3242         }
3243
3244         /*
3245          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3246          * reexamined in the next iteration.
3247          */
3248         list_splice_init(&processed_list, &local_queue);
3249         lnet_net_lock(0);
3250         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3251         lnet_net_unlock(0);
3252 }
3253
3254 static int
3255 lnet_resendqs_create(void)
3256 {
3257         struct list_head **resendqs;
3258         resendqs = lnet_create_array_of_queues();
3259
3260         if (!resendqs)
3261                 return -ENOMEM;
3262
3263         lnet_net_lock(LNET_LOCK_EX);
3264         the_lnet.ln_mt_resendqs = resendqs;
3265         lnet_net_unlock(LNET_LOCK_EX);
3266
3267         return 0;
3268 }
3269
3270 static void
3271 lnet_clean_local_ni_recoveryq(void)
3272 {
3273         struct lnet_ni *ni;
3274
3275         /* This is only called when the monitor thread has stopped */
3276         lnet_net_lock(0);
3277
3278         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3279                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3280                                 struct lnet_ni, ni_recovery);
3281                 list_del_init(&ni->ni_recovery);
3282                 lnet_ni_lock(ni);
3283                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3284                 lnet_ni_unlock(ni);
3285                 lnet_ni_decref_locked(ni, 0);
3286         }
3287
3288         lnet_net_unlock(0);
3289 }
3290
3291 static void
3292 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3293                                      bool force)
3294 {
3295         struct lnet_handle_md recovery_mdh;
3296
3297         LNetInvalidateMDHandle(&recovery_mdh);
3298
3299         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3300                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3301                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3302         }
3303         spin_unlock(&lpni->lpni_lock);
3304         lnet_net_unlock(cpt);
3305         if (!LNetMDHandleIsInvalid(recovery_mdh))
3306                 LNetMDUnlink(recovery_mdh);
3307         lnet_net_lock(cpt);
3308         spin_lock(&lpni->lpni_lock);
3309 }
3310
3311 static void
3312 lnet_clean_peer_ni_recoveryq(void)
3313 {
3314         struct lnet_peer_ni *lpni, *tmp;
3315
3316         lnet_net_lock(LNET_LOCK_EX);
3317
3318         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3319                                  lpni_recovery) {
3320                 list_del_init(&lpni->lpni_recovery);
3321                 spin_lock(&lpni->lpni_lock);
3322                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3323                 spin_unlock(&lpni->lpni_lock);
3324                 lnet_peer_ni_decref_locked(lpni);
3325         }
3326
3327         lnet_net_unlock(LNET_LOCK_EX);
3328 }
3329
3330 static void
3331 lnet_clean_resendqs(void)
3332 {
3333         struct lnet_msg *msg, *tmp;
3334         struct list_head msgs;
3335         int i;
3336
3337         INIT_LIST_HEAD(&msgs);
3338
3339         cfs_cpt_for_each(i, lnet_cpt_table()) {
3340                 lnet_net_lock(i);
3341                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3342                 lnet_net_unlock(i);
3343                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3344                         list_del_init(&msg->msg_list);
3345                         msg->msg_no_resend = true;
3346                         lnet_finalize(msg, -ESHUTDOWN);
3347                 }
3348         }
3349
3350         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3351 }
3352
3353 static void
3354 lnet_recover_peer_nis(void)
3355 {
3356         struct lnet_mt_event_info *ev_info;
3357         struct list_head processed_list;
3358         struct list_head local_queue;
3359         struct lnet_handle_md mdh;
3360         struct lnet_peer_ni *lpni;
3361         struct lnet_peer_ni *tmp;
3362         lnet_nid_t nid;
3363         int healthv;
3364         int rc;
3365
3366         INIT_LIST_HEAD(&local_queue);
3367         INIT_LIST_HEAD(&processed_list);
3368
3369         /*
3370          * Always use cpt 0 for locking across all interactions with
3371          * ln_mt_peerNIRecovq
3372          */
3373         lnet_net_lock(0);
3374         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3375                          &local_queue);
3376         lnet_net_unlock(0);
3377
3378         list_for_each_entry_safe(lpni, tmp, &local_queue,
3379                                  lpni_recovery) {
3380                 /*
3381                  * The same protection strategy is used here as is in the
3382                  * local recovery case.
3383                  */
3384                 lnet_net_lock(0);
3385                 healthv = atomic_read(&lpni->lpni_healthv);
3386                 spin_lock(&lpni->lpni_lock);
3387                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3388                     healthv == LNET_MAX_HEALTH_VALUE) {
3389                         list_del_init(&lpni->lpni_recovery);
3390                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3391                         spin_unlock(&lpni->lpni_lock);
3392                         lnet_peer_ni_decref_locked(lpni);
3393                         lnet_net_unlock(0);
3394                         continue;
3395                 }
3396
3397                 /*
3398                  * If the peer NI has failed recovery we must unlink the
3399                  * md. But we want to keep the peer ni on the recovery
3400                  * queue so we can try to continue recovering it
3401                  */
3402                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3403                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3404                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3405                 }
3406
3407                 spin_unlock(&lpni->lpni_lock);
3408                 lnet_net_unlock(0);
3409
3410                 /*
3411                  * NOTE: we're racing with peer deletion from user space.
3412                  * It's possible that a peer is deleted after we check its
3413                  * state. In this case the recovery can create a new peer
3414                  */
3415                 spin_lock(&lpni->lpni_lock);
3416                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3417                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3418                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3419                         spin_unlock(&lpni->lpni_lock);
3420
3421                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3422                         if (!ev_info) {
3423                                 CERROR("out of memory. Can't recover %s\n",
3424                                        libcfs_nid2str(lpni->lpni_nid));
3425                                 spin_lock(&lpni->lpni_lock);
3426                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3427                                 spin_unlock(&lpni->lpni_lock);
3428                                 continue;
3429                         }
3430
3431                         /* look at the comments in lnet_recover_local_nis() */
3432                         mdh = lpni->lpni_recovery_ping_mdh;
3433                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3434                         nid = lpni->lpni_nid;
3435                         lnet_net_lock(0);
3436                         list_del_init(&lpni->lpni_recovery);
3437                         lnet_peer_ni_decref_locked(lpni);
3438                         lnet_net_unlock(0);
3439
3440                         ev_info->mt_type = MT_TYPE_PEER_NI;
3441                         ev_info->mt_nid = nid;
3442                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3443                                             ev_info, the_lnet.ln_mt_eqh, true);
3444                         lnet_net_lock(0);
3445                         /*
3446                          * lnet_find_peer_ni_locked() grabs a refcount for
3447                          * us. No need to take it explicitly.
3448                          */
3449                         lpni = lnet_find_peer_ni_locked(nid);
3450                         if (!lpni) {
3451                                 lnet_net_unlock(0);
3452                                 LNetMDUnlink(mdh);
3453                                 continue;
3454                         }
3455
3456                         lpni->lpni_recovery_ping_mdh = mdh;
3457                         /*
3458                          * While we're unlocked the lpni could've been
3459                          * readded on the recovery queue. In this case we
3460                          * don't need to add it to the local queue, since
3461                          * it's already on there and the thread that added
3462                          * it would've incremented the refcount on the
3463                          * peer, which means we need to decref the refcount
3464                          * that was implicitly grabbed by find_peer_ni_locked.
3465                          * Otherwise, if the lpni is still not on
3466                          * the recovery queue, then we'll add it to the
3467                          * processed list.
3468                          */
3469                         if (list_empty(&lpni->lpni_recovery))
3470                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3471                         else
3472                                 lnet_peer_ni_decref_locked(lpni);
3473                         lnet_net_unlock(0);
3474
3475                         spin_lock(&lpni->lpni_lock);
3476                         if (rc)
3477                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3478                 }
3479                 spin_unlock(&lpni->lpni_lock);
3480         }
3481
3482         list_splice_init(&processed_list, &local_queue);
3483         lnet_net_lock(0);
3484         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3485         lnet_net_unlock(0);
3486 }
3487
3488 static int
3489 lnet_monitor_thread(void *arg)
3490 {
3491         time64_t recovery_timeout = 0;
3492         time64_t rsp_timeout = 0;
3493         int interval;
3494         time64_t now;
3495
3496         wait_for_completion(&the_lnet.ln_started);
3497         /*
3498          * The monitor thread takes care of the following:
3499          *  1. Checks the aliveness of routers
3500          *  2. Checks if there are messages on the resend queue to resend
3501          *     them.
3502          *  3. Check if there are any NIs on the local recovery queue and
3503          *     pings them
3504          *  4. Checks if there are any NIs on the remote recovery queue
3505          *     and pings them.
3506          */
3507         cfs_block_allsigs();
3508
3509         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3510                 now = ktime_get_real_seconds();
3511
3512                 if (lnet_router_checker_active())
3513                         lnet_check_routers();
3514
3515                 lnet_resend_pending_msgs();
3516
3517                 if (now >= rsp_timeout) {
3518                         lnet_finalize_expired_responses();
3519                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3520                 }
3521
3522                 if (now >= recovery_timeout) {
3523                         lnet_recover_local_nis();
3524                         lnet_recover_peer_nis();
3525                         recovery_timeout = now + lnet_recovery_interval;
3526                 }
3527
3528                 /*
3529                  * TODO do we need to check if we should sleep without
3530                  * timeout?  Technically, an active system will always
3531                  * have messages in flight so this check will always
3532                  * evaluate to false. And on an idle system do we care
3533                  * if we wake up every 1 second? Although, we've seen
3534                  * cases where we get a complaint that an idle thread
3535                  * is waking up unnecessarily.
3536                  *
3537                  * Take into account the current net_count when you wake
3538                  * up for alive router checking, since we need to check
3539                  * possibly as many networks as we have configured.
3540                  */
3541                 interval = min(lnet_recovery_interval,
3542                                min((unsigned int) alive_router_check_interval /
3543                                         lnet_current_net_count,
3544                                    lnet_transaction_timeout / 2));
3545                 wait_for_completion_interruptible_timeout(
3546                         &the_lnet.ln_mt_wait_complete,
3547                         cfs_time_seconds(interval));
3548                 /* Must re-init the completion before testing anything,
3549                  * including ln_mt_state.
3550                  */
3551                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3552         }
3553
3554         /* Shutting down */
3555         lnet_net_lock(LNET_LOCK_EX);
3556         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3557         lnet_net_unlock(LNET_LOCK_EX);
3558
3559         /* signal that the monitor thread is exiting */
3560         up(&the_lnet.ln_mt_signal);
3561
3562         return 0;
3563 }
3564
3565 /*
3566  * lnet_send_ping
3567  * Sends a ping.
3568  * Returns == 0 if success
3569  * Returns > 0 if LNetMDBind or prior fails
3570  * Returns < 0 if LNetGet fails
3571  */
3572 int
3573 lnet_send_ping(lnet_nid_t dest_nid,
3574                struct lnet_handle_md *mdh, int nnis,
3575                void *user_data, struct lnet_handle_eq eqh, bool recovery)
3576 {
3577         struct lnet_md md = { NULL };
3578         struct lnet_process_id id;
3579         struct lnet_ping_buffer *pbuf;
3580         int rc;
3581
3582         if (dest_nid == LNET_NID_ANY) {
3583                 rc = -EHOSTUNREACH;
3584                 goto fail_error;
3585         }
3586
3587         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3588         if (!pbuf) {
3589                 rc = ENOMEM;
3590                 goto fail_error;
3591         }
3592
3593         /* initialize md content */
3594         md.start     = &pbuf->pb_info;
3595         md.length    = LNET_PING_INFO_SIZE(nnis);
3596         md.threshold = 2; /* GET/REPLY */
3597         md.max_size  = 0;
3598         md.options   = LNET_MD_TRUNCATE;
3599         md.user_ptr  = user_data;
3600         md.eq_handle = eqh;
3601
3602         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3603         if (rc) {
3604                 lnet_ping_buffer_decref(pbuf);
3605                 CERROR("Can't bind MD: %d\n", rc);
3606                 rc = -rc; /* change the rc to positive */
3607                 goto fail_error;
3608         }
3609         id.pid = LNET_PID_LUSTRE;
3610         id.nid = dest_nid;
3611
3612         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3613                      LNET_RESERVED_PORTAL,
3614                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3615
3616         if (rc)
3617                 goto fail_unlink_md;
3618
3619         return 0;
3620
3621 fail_unlink_md:
3622         LNetMDUnlink(*mdh);
3623         LNetInvalidateMDHandle(mdh);
3624 fail_error:
3625         return rc;
3626 }
3627
3628 static void
3629 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3630                            int status, bool unlink_event)
3631 {
3632         lnet_nid_t nid = ev_info->mt_nid;
3633
3634         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3635                 struct lnet_ni *ni;
3636
3637                 lnet_net_lock(0);
3638                 ni = lnet_nid2ni_locked(nid, 0);
3639                 if (!ni) {
3640                         lnet_net_unlock(0);
3641                         return;
3642                 }
3643                 lnet_ni_lock(ni);
3644                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3645                 if (status)
3646                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3647                 lnet_ni_unlock(ni);
3648                 lnet_net_unlock(0);
3649
3650                 if (status != 0) {
3651                         CERROR("local NI (%s) recovery failed with %d\n",
3652                                libcfs_nid2str(nid), status);
3653                         return;
3654                 }
3655                 /*
3656                  * need to increment healthv for the ni here, because in
3657                  * the lnet_finalize() path we don't have access to this
3658                  * NI. And in order to get access to it, we'll need to
3659                  * carry forward too much information.
3660                  * In the peer case, it'll naturally be incremented
3661                  */
3662                 if (!unlink_event)
3663                         lnet_inc_healthv(&ni->ni_healthv);
3664         } else {
3665                 struct lnet_peer_ni *lpni;
3666                 int cpt;
3667
3668                 cpt = lnet_net_lock_current();
3669                 lpni = lnet_find_peer_ni_locked(nid);
3670                 if (!lpni) {
3671                         lnet_net_unlock(cpt);
3672                         return;
3673                 }
3674                 spin_lock(&lpni->lpni_lock);
3675                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3676                 if (status)
3677                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3678                 spin_unlock(&lpni->lpni_lock);
3679                 lnet_peer_ni_decref_locked(lpni);
3680                 lnet_net_unlock(cpt);
3681
3682                 if (status != 0)
3683                         CERROR("peer NI (%s) recovery failed with %d\n",
3684                                libcfs_nid2str(nid), status);
3685         }
3686 }
3687
3688 void
3689 lnet_mt_event_handler(struct lnet_event *event)
3690 {
3691         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3692         struct lnet_ping_buffer *pbuf;
3693
3694         /* TODO: remove assert */
3695         LASSERT(event->type == LNET_EVENT_REPLY ||
3696                 event->type == LNET_EVENT_SEND ||
3697                 event->type == LNET_EVENT_UNLINK);
3698
3699         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3700                event->status);
3701
3702         switch (event->type) {
3703         case LNET_EVENT_UNLINK:
3704                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3705                        libcfs_nid2str(ev_info->mt_nid));
3706                 /* fallthrough */
3707         case LNET_EVENT_REPLY:
3708                 lnet_handle_recovery_reply(ev_info, event->status,
3709                                            event->type == LNET_EVENT_UNLINK);
3710                 break;
3711         case LNET_EVENT_SEND:
3712                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3713                                libcfs_nid2str(ev_info->mt_nid),
3714                                (event->status) ? "unsuccessfully" :
3715                                "successfully", event->status);
3716                 break;
3717         default:
3718                 CERROR("Unexpected event: %d\n", event->type);
3719                 break;
3720         }
3721         if (event->unlinked) {
3722                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3723                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3724                 lnet_ping_buffer_decref(pbuf);
3725         }
3726 }
3727
3728 static int
3729 lnet_rsp_tracker_create(void)
3730 {
3731         struct list_head **rstqs;
3732         rstqs = lnet_create_array_of_queues();
3733
3734         if (!rstqs)
3735                 return -ENOMEM;
3736
3737         the_lnet.ln_mt_rstq = rstqs;
3738
3739         return 0;
3740 }
3741
3742 static void
3743 lnet_rsp_tracker_clean(void)
3744 {
3745         lnet_finalize_expired_responses();
3746
3747         cfs_percpt_free(the_lnet.ln_mt_rstq);
3748         the_lnet.ln_mt_rstq = NULL;
3749 }
3750
3751 int lnet_monitor_thr_start(void)
3752 {
3753         int rc = 0;
3754         struct task_struct *task;
3755
3756         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3757                 return -EALREADY;
3758
3759         rc = lnet_resendqs_create();
3760         if (rc)
3761                 return rc;
3762
3763         rc = lnet_rsp_tracker_create();
3764         if (rc)
3765                 goto clean_queues;
3766
3767         sema_init(&the_lnet.ln_mt_signal, 0);
3768
3769         lnet_net_lock(LNET_LOCK_EX);
3770         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3771         lnet_net_unlock(LNET_LOCK_EX);
3772         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3773         if (IS_ERR(task)) {
3774                 rc = PTR_ERR(task);
3775                 CERROR("Can't start monitor thread: %d\n", rc);
3776                 goto clean_thread;
3777         }
3778
3779         return 0;
3780
3781 clean_thread:
3782         lnet_net_lock(LNET_LOCK_EX);
3783         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3784         lnet_net_unlock(LNET_LOCK_EX);
3785         /* block until event callback signals exit */
3786         down(&the_lnet.ln_mt_signal);
3787         /* clean up */
3788         lnet_net_lock(LNET_LOCK_EX);
3789         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3790         lnet_net_unlock(LNET_LOCK_EX);
3791         lnet_rsp_tracker_clean();
3792         lnet_clean_local_ni_recoveryq();
3793         lnet_clean_peer_ni_recoveryq();
3794         lnet_clean_resendqs();
3795         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
3796         return rc;
3797 clean_queues:
3798         lnet_rsp_tracker_clean();
3799         lnet_clean_local_ni_recoveryq();
3800         lnet_clean_peer_ni_recoveryq();
3801         lnet_clean_resendqs();
3802         return rc;
3803 }
3804
3805 void lnet_monitor_thr_stop(void)
3806 {
3807         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3808                 return;
3809
3810         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3811         lnet_net_lock(LNET_LOCK_EX);
3812         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3813         lnet_net_unlock(LNET_LOCK_EX);
3814
3815         /* tell the monitor thread that we're shutting down */
3816         complete(&the_lnet.ln_mt_wait_complete);
3817
3818         /* block until monitor thread signals that it's done */
3819         down(&the_lnet.ln_mt_signal);
3820         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3821
3822         /* perform cleanup tasks */
3823         lnet_rsp_tracker_clean();
3824         lnet_clean_local_ni_recoveryq();
3825         lnet_clean_peer_ni_recoveryq();
3826         lnet_clean_resendqs();
3827
3828         return;
3829 }
3830
3831 void
3832 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3833                   __u32 msg_type)
3834 {
3835         lnet_net_lock(cpt);
3836         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3837         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3838         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3839         lnet_net_unlock(cpt);
3840
3841         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3842 }
3843
3844 static void
3845 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3846 {
3847         struct lnet_hdr *hdr = &msg->msg_hdr;
3848
3849         if (msg->msg_wanted != 0)
3850                 lnet_setpayloadbuffer(msg);
3851
3852         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3853
3854         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3855          * it back into the ACK during lnet_finalize() */
3856         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3857                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3858
3859         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3860                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3861 }
3862
3863 static int
3864 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3865 {
3866         struct lnet_hdr         *hdr = &msg->msg_hdr;
3867         struct lnet_match_info  info;
3868         int                     rc;
3869         bool                    ready_delay;
3870
3871         /* Convert put fields to host byte order */
3872         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3873         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3874         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3875
3876         /* Primary peer NID. */
3877         info.mi_id.nid  = msg->msg_initiator;
3878         info.mi_id.pid  = hdr->src_pid;
3879         info.mi_opc     = LNET_MD_OP_PUT;
3880         info.mi_portal  = hdr->msg.put.ptl_index;
3881         info.mi_rlength = hdr->payload_length;
3882         info.mi_roffset = hdr->msg.put.offset;
3883         info.mi_mbits   = hdr->msg.put.match_bits;
3884         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3885
3886         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3887         ready_delay = msg->msg_rx_ready_delay;
3888
3889  again:
3890         rc = lnet_ptl_match_md(&info, msg);
3891         switch (rc) {
3892         default:
3893                 LBUG();
3894
3895         case LNET_MATCHMD_OK:
3896                 lnet_recv_put(ni, msg);
3897                 return 0;
3898
3899         case LNET_MATCHMD_NONE:
3900                 if (ready_delay)
3901                         /* no eager_recv or has already called it, should
3902                          * have been attached on delayed list */
3903                         return 0;
3904
3905                 rc = lnet_ni_eager_recv(ni, msg);
3906                 if (rc == 0) {
3907                         ready_delay = true;
3908                         goto again;
3909                 }
3910                 /* fall through */
3911
3912         case LNET_MATCHMD_DROP:
3913                 CNETERR("Dropping PUT from %s portal %d match %llu"
3914                         " offset %d length %d: %d\n",
3915                         libcfs_id2str(info.mi_id), info.mi_portal,
3916                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3917
3918                 return -ENOENT; /* -ve: OK but no match */
3919         }
3920 }
3921
3922 static int
3923 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3924 {
3925         struct lnet_match_info info;
3926         struct lnet_hdr *hdr = &msg->msg_hdr;
3927         struct lnet_process_id source_id;
3928         struct lnet_handle_wire reply_wmd;
3929         int rc;
3930
3931         /* Convert get fields to host byte order */
3932         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3933         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3934         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3935         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3936
3937         source_id.nid = hdr->src_nid;
3938         source_id.pid = hdr->src_pid;
3939         /* Primary peer NID */
3940         info.mi_id.nid  = msg->msg_initiator;
3941         info.mi_id.pid  = hdr->src_pid;
3942         info.mi_opc     = LNET_MD_OP_GET;
3943         info.mi_portal  = hdr->msg.get.ptl_index;
3944         info.mi_rlength = hdr->msg.get.sink_length;
3945         info.mi_roffset = hdr->msg.get.src_offset;
3946         info.mi_mbits   = hdr->msg.get.match_bits;
3947         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3948
3949         rc = lnet_ptl_match_md(&info, msg);
3950         if (rc == LNET_MATCHMD_DROP) {
3951                 CNETERR("Dropping GET from %s portal %d match %llu"
3952                         " offset %d length %d\n",
3953                         libcfs_id2str(info.mi_id), info.mi_portal,
3954                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3955                 return -ENOENT; /* -ve: OK but no match */
3956         }
3957
3958         LASSERT(rc == LNET_MATCHMD_OK);
3959
3960         lnet_build_msg_event(msg, LNET_EVENT_GET);
3961
3962         reply_wmd = hdr->msg.get.return_wmd;
3963
3964         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3965                        msg->msg_offset, msg->msg_wanted);
3966
3967         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3968
3969         if (rdma_get) {
3970                 /* The LND completes the REPLY from her recv procedure */
3971                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3972                              msg->msg_offset, msg->msg_len, msg->msg_len);
3973                 return 0;
3974         }
3975
3976         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3977         msg->msg_receiving = 0;
3978
3979         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3980         if (rc < 0) {
3981                 /* didn't get as far as lnet_ni_send() */
3982                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3983                        libcfs_nid2str(ni->ni_nid),
3984                        libcfs_id2str(info.mi_id), rc);
3985
3986                 lnet_finalize(msg, rc);
3987         }
3988
3989         return 0;
3990 }
3991
3992 static int
3993 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3994 {
3995         void *private = msg->msg_private;
3996         struct lnet_hdr *hdr = &msg->msg_hdr;
3997         struct lnet_process_id src = {0};
3998         struct lnet_libmd *md;
3999         int rlength;
4000         int mlength;
4001         int cpt;
4002
4003         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4004         lnet_res_lock(cpt);
4005
4006         src.nid = hdr->src_nid;
4007         src.pid = hdr->src_pid;
4008
4009         /* NB handles only looked up by creator (no flips) */
4010         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4011         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4012                 CNETERR("%s: Dropping REPLY from %s for %s "
4013                         "MD %#llx.%#llx\n",
4014                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4015                         (md == NULL) ? "invalid" : "inactive",
4016                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4017                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4018                 if (md != NULL && md->md_me != NULL)
4019                         CERROR("REPLY MD also attached to portal %d\n",
4020                                md->md_me->me_portal);
4021
4022                 lnet_res_unlock(cpt);
4023                 return -ENOENT; /* -ve: OK but no match */
4024         }
4025
4026         LASSERT(md->md_offset == 0);
4027
4028         rlength = hdr->payload_length;
4029         mlength = MIN(rlength, (int)md->md_length);
4030
4031         if (mlength < rlength &&
4032             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4033                 CNETERR("%s: Dropping REPLY from %s length %d "
4034                         "for MD %#llx would overflow (%d)\n",
4035                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4036                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4037                         mlength);
4038                 lnet_res_unlock(cpt);
4039                 return -ENOENT; /* -ve: OK but no match */
4040         }
4041
4042         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4043                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4044                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4045
4046         lnet_msg_attach_md(msg, md, 0, mlength);
4047
4048         if (mlength != 0)
4049                 lnet_setpayloadbuffer(msg);
4050
4051         lnet_res_unlock(cpt);
4052
4053         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4054
4055         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4056         return 0;
4057 }
4058
4059 static int
4060 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4061 {
4062         struct lnet_hdr *hdr = &msg->msg_hdr;
4063         struct lnet_process_id src = {0};
4064         struct lnet_libmd *md;
4065         int cpt;
4066
4067         src.nid = hdr->src_nid;
4068         src.pid = hdr->src_pid;
4069
4070         /* Convert ack fields to host byte order */
4071         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4072         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4073
4074         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4075         lnet_res_lock(cpt);
4076
4077         /* NB handles only looked up by creator (no flips) */
4078         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4079         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4080                 /* Don't moan; this is expected */
4081                 CDEBUG(D_NET,
4082                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4083                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4084                        (md == NULL) ? "invalid" : "inactive",
4085                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4086                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4087                 if (md != NULL && md->md_me != NULL)
4088                         CERROR("Source MD also attached to portal %d\n",
4089                                md->md_me->me_portal);
4090
4091                 lnet_res_unlock(cpt);
4092                 return -ENOENT;                  /* -ve! */
4093         }
4094
4095         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4096                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4097                hdr->msg.ack.dst_wmd.wh_object_cookie);
4098
4099         lnet_msg_attach_md(msg, md, 0, 0);
4100
4101         lnet_res_unlock(cpt);
4102
4103         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4104
4105         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4106         return 0;
4107 }
4108
4109 /**
4110  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4111  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4112  * \retval -ve                  error code
4113  */
4114 int
4115 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4116 {
4117         int     rc = 0;
4118
4119         if (!the_lnet.ln_routing)
4120                 return -ECANCELED;
4121
4122         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4123             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4124                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4125                         msg->msg_rx_ready_delay = 1;
4126                 } else {
4127                         lnet_net_unlock(msg->msg_rx_cpt);
4128                         rc = lnet_ni_eager_recv(ni, msg);
4129                         lnet_net_lock(msg->msg_rx_cpt);
4130                 }
4131         }
4132
4133         if (rc == 0)
4134                 rc = lnet_post_routed_recv_locked(msg, 0);
4135         return rc;
4136 }
4137
4138 int
4139 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4140 {
4141         int     rc;
4142
4143         switch (msg->msg_type) {
4144         case LNET_MSG_ACK:
4145                 rc = lnet_parse_ack(ni, msg);
4146                 break;
4147         case LNET_MSG_PUT:
4148                 rc = lnet_parse_put(ni, msg);
4149                 break;
4150         case LNET_MSG_GET:
4151                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4152                 break;
4153         case LNET_MSG_REPLY:
4154                 rc = lnet_parse_reply(ni, msg);
4155                 break;
4156         default: /* prevent an unused label if !kernel */
4157                 LASSERT(0);
4158                 return -EPROTO;
4159         }
4160
4161         LASSERT(rc == 0 || rc == -ENOENT);
4162         return rc;
4163 }
4164
4165 char *
4166 lnet_msgtyp2str (int type)
4167 {
4168         switch (type) {
4169         case LNET_MSG_ACK:
4170                 return ("ACK");
4171         case LNET_MSG_PUT:
4172                 return ("PUT");
4173         case LNET_MSG_GET:
4174                 return ("GET");
4175         case LNET_MSG_REPLY:
4176                 return ("REPLY");
4177         case LNET_MSG_HELLO:
4178                 return ("HELLO");
4179         default:
4180                 return ("<UNKNOWN>");
4181         }
4182 }
4183
4184 void
4185 lnet_print_hdr(struct lnet_hdr *hdr)
4186 {
4187         struct lnet_process_id src = {
4188                 .nid = hdr->src_nid,
4189                 .pid = hdr->src_pid,
4190         };
4191         struct lnet_process_id dst = {
4192                 .nid = hdr->dest_nid,
4193                 .pid = hdr->dest_pid,
4194         };
4195         char *type_str = lnet_msgtyp2str(hdr->type);
4196
4197         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4198         CWARN("    From %s\n", libcfs_id2str(src));
4199         CWARN("    To   %s\n", libcfs_id2str(dst));
4200
4201         switch (hdr->type) {
4202         default:
4203                 break;
4204
4205         case LNET_MSG_PUT:
4206                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4207                       "match bits %llu\n",
4208                       hdr->msg.put.ptl_index,
4209                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4210                       hdr->msg.put.ack_wmd.wh_object_cookie,
4211                       hdr->msg.put.match_bits);
4212                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4213                       hdr->payload_length, hdr->msg.put.offset,
4214                       hdr->msg.put.hdr_data);
4215                 break;
4216
4217         case LNET_MSG_GET:
4218                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4219                       "match bits %llu\n", hdr->msg.get.ptl_index,
4220                       hdr->msg.get.return_wmd.wh_interface_cookie,
4221                       hdr->msg.get.return_wmd.wh_object_cookie,
4222                       hdr->msg.get.match_bits);
4223                 CWARN("    Length %d, src offset %d\n",
4224                       hdr->msg.get.sink_length,
4225                       hdr->msg.get.src_offset);
4226                 break;
4227
4228         case LNET_MSG_ACK:
4229                 CWARN("    dst md %#llx.%#llx, "
4230                       "manipulated length %d\n",
4231                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4232                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4233                       hdr->msg.ack.mlength);
4234                 break;
4235
4236         case LNET_MSG_REPLY:
4237                 CWARN("    dst md %#llx.%#llx, "
4238                       "length %d\n",
4239                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4240                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4241                       hdr->payload_length);
4242         }
4243
4244 }
4245
4246 int
4247 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4248            void *private, int rdma_req)
4249 {
4250         struct lnet_peer_ni *lpni;
4251         struct lnet_msg *msg;
4252         __u32 payload_length;
4253         lnet_pid_t dest_pid;
4254         lnet_nid_t dest_nid;
4255         lnet_nid_t src_nid;
4256         bool push = false;
4257         int for_me;
4258         __u32 type;
4259         int rc = 0;
4260         int cpt;
4261
4262         LASSERT (!in_interrupt ());
4263
4264         type = le32_to_cpu(hdr->type);
4265         src_nid = le64_to_cpu(hdr->src_nid);
4266         dest_nid = le64_to_cpu(hdr->dest_nid);
4267         dest_pid = le32_to_cpu(hdr->dest_pid);
4268         payload_length = le32_to_cpu(hdr->payload_length);
4269
4270         for_me = (ni->ni_nid == dest_nid);
4271         cpt = lnet_cpt_of_nid(from_nid, ni);
4272
4273         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4274                 libcfs_nid2str(dest_nid),
4275                 libcfs_nid2str(ni->ni_nid),
4276                 libcfs_nid2str(src_nid),
4277                 lnet_msgtyp2str(type),
4278                 (for_me) ? "for me" : "routed");
4279
4280         switch (type) {
4281         case LNET_MSG_ACK:
4282         case LNET_MSG_GET:
4283                 if (payload_length > 0) {
4284                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4285                                libcfs_nid2str(from_nid),
4286                                libcfs_nid2str(src_nid),
4287                                lnet_msgtyp2str(type), payload_length);
4288                         return -EPROTO;
4289                 }
4290                 break;
4291
4292         case LNET_MSG_PUT:
4293         case LNET_MSG_REPLY:
4294                 if (payload_length >
4295                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4296                         CERROR("%s, src %s: bad %s payload %d "
4297                                "(%d max expected)\n",
4298                                libcfs_nid2str(from_nid),
4299                                libcfs_nid2str(src_nid),
4300                                lnet_msgtyp2str(type),
4301                                payload_length,
4302                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4303                         return -EPROTO;
4304                 }
4305                 break;
4306
4307         default:
4308                 CERROR("%s, src %s: Bad message type 0x%x\n",
4309                        libcfs_nid2str(from_nid),
4310                        libcfs_nid2str(src_nid), type);
4311                 return -EPROTO;
4312         }
4313
4314         if (the_lnet.ln_routing &&
4315             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4316                 lnet_ni_lock(ni);
4317                 spin_lock(&ni->ni_net->net_lock);
4318                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4319                 spin_unlock(&ni->ni_net->net_lock);
4320                 if (ni->ni_status != NULL &&
4321                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4322                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4323                         push = true;
4324                 }
4325                 lnet_ni_unlock(ni);
4326         }
4327
4328         if (push)
4329                 lnet_push_update_to_peers(1);
4330
4331         /* Regard a bad destination NID as a protocol error.  Senders should
4332          * know what they're doing; if they don't they're misconfigured, buggy
4333          * or malicious so we chop them off at the knees :) */
4334
4335         if (!for_me) {
4336                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4337                         /* should have gone direct */
4338                         CERROR("%s, src %s: Bad dest nid %s "
4339                                "(should have been sent direct)\n",
4340                                 libcfs_nid2str(from_nid),
4341                                 libcfs_nid2str(src_nid),
4342                                 libcfs_nid2str(dest_nid));
4343                         return -EPROTO;
4344                 }
4345
4346                 if (lnet_islocalnid(dest_nid)) {
4347                         /* dest is another local NI; sender should have used
4348                          * this node's NID on its own network */
4349                         CERROR("%s, src %s: Bad dest nid %s "
4350                                "(it's my nid but on a different network)\n",
4351                                 libcfs_nid2str(from_nid),
4352                                 libcfs_nid2str(src_nid),
4353                                 libcfs_nid2str(dest_nid));
4354                         return -EPROTO;
4355                 }
4356
4357                 if (rdma_req && type == LNET_MSG_GET) {
4358                         CERROR("%s, src %s: Bad optimized GET for %s "
4359                                "(final destination must be me)\n",
4360                                 libcfs_nid2str(from_nid),
4361                                 libcfs_nid2str(src_nid),
4362                                 libcfs_nid2str(dest_nid));
4363                         return -EPROTO;
4364                 }
4365
4366                 if (!the_lnet.ln_routing) {
4367                         CERROR("%s, src %s: Dropping message for %s "
4368                                "(routing not enabled)\n",
4369                                 libcfs_nid2str(from_nid),
4370                                 libcfs_nid2str(src_nid),
4371                                 libcfs_nid2str(dest_nid));
4372                         goto drop;
4373                 }
4374         }
4375
4376         /* Message looks OK; we're not going to return an error, so we MUST
4377          * call back lnd_recv() come what may... */
4378
4379         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4380             fail_peer(src_nid, 0)) {                    /* shall we now? */
4381                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4382                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4383                        lnet_msgtyp2str(type));
4384                 goto drop;
4385         }
4386
4387         if (!list_empty(&the_lnet.ln_drop_rules) &&
4388             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4389                 CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
4390                               "silent message loss\n",
4391                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4392                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4393                 goto drop;
4394         }
4395
4396         if (lnet_drop_asym_route && for_me &&
4397             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4398                 struct lnet_net *net;
4399                 struct lnet_remotenet *rnet;
4400                 bool found = true;
4401
4402                 /* we are dealing with a routed message,
4403                  * so see if route to reach src_nid goes through from_nid
4404                  */
4405                 lnet_net_lock(cpt);
4406                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4407                 if (!net) {
4408                         lnet_net_unlock(cpt);
4409                         CERROR("net %s not found\n",
4410                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4411                         return -EPROTO;
4412                 }
4413
4414                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4415                 if (rnet) {
4416                         struct lnet_peer *gw = NULL;
4417                         struct lnet_peer_ni *lpni = NULL;
4418                         struct lnet_route *route;
4419
4420                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4421                                 found = false;
4422                                 gw = route->lr_gateway;
4423                                 if (route->lr_lnet != net->net_id)
4424                                         continue;
4425                                 /*
4426                                  * if the nid is one of the gateway's NIDs
4427                                  * then this is a valid gateway
4428                                  */
4429                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4430                                                 NULL, lpni)) != NULL) {
4431                                         if (lpni->lpni_nid == from_nid) {
4432                                                 found = true;
4433                                                 break;
4434                                         }
4435                                 }
4436                         }
4437                 }
4438                 lnet_net_unlock(cpt);
4439                 if (!found) {
4440                         /* we would not use from_nid to route a message to
4441                          * src_nid
4442                          * => asymmetric routing detected but forbidden
4443                          */
4444                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4445                                libcfs_nid2str(from_nid),
4446                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4447                         goto drop;
4448                 }
4449         }
4450
4451         msg = lnet_msg_alloc();
4452         if (msg == NULL) {
4453                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4454                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4455                        lnet_msgtyp2str(type));
4456                 goto drop;
4457         }
4458
4459         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4460          * pointers NULL etc */
4461
4462         msg->msg_type = type;
4463         msg->msg_private = private;
4464         msg->msg_receiving = 1;
4465         msg->msg_rdma_get = rdma_req;
4466         msg->msg_len = msg->msg_wanted = payload_length;
4467         msg->msg_offset = 0;
4468         msg->msg_hdr = *hdr;
4469         /* for building message event */
4470         msg->msg_from = from_nid;
4471         if (!for_me) {
4472                 msg->msg_target.pid     = dest_pid;
4473                 msg->msg_target.nid     = dest_nid;
4474                 msg->msg_routing        = 1;
4475
4476         } else {
4477                 /* convert common msg->hdr fields to host byteorder */
4478                 msg->msg_hdr.type       = type;
4479                 msg->msg_hdr.src_nid    = src_nid;
4480                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4481                 msg->msg_hdr.dest_nid   = dest_nid;
4482                 msg->msg_hdr.dest_pid   = dest_pid;
4483                 msg->msg_hdr.payload_length = payload_length;
4484         }
4485
4486         lnet_net_lock(cpt);
4487         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4488         if (IS_ERR(lpni)) {
4489                 lnet_net_unlock(cpt);
4490                 CERROR("%s, src %s: Dropping %s "
4491                        "(error %ld looking up sender)\n",
4492                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4493                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4494                 lnet_msg_free(msg);
4495                 if (rc == -ESHUTDOWN)
4496                         /* We are shutting down.  Don't do anything more */
4497                         return 0;
4498                 goto drop;
4499         }
4500
4501         if (the_lnet.ln_routing)
4502                 lpni->lpni_last_alive = ktime_get_seconds();
4503
4504         msg->msg_rxpeer = lpni;
4505         msg->msg_rxni = ni;
4506         lnet_ni_addref_locked(ni, cpt);
4507         /* Multi-Rail: Primary NID of source. */
4508         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4509
4510         /*
4511          * mark the status of this lpni as UP since we received a message
4512          * from it. The ping response reports back the ns_status which is
4513          * marked on the remote as up or down and we cache it here.
4514          */
4515         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4516
4517         lnet_msg_commit(msg, cpt);
4518
4519         /* message delay simulation */
4520         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4521                      lnet_delay_rule_match_locked(hdr, msg))) {
4522                 lnet_net_unlock(cpt);
4523                 return 0;
4524         }
4525
4526         if (!for_me) {
4527                 rc = lnet_parse_forward_locked(ni, msg);
4528                 lnet_net_unlock(cpt);
4529
4530                 if (rc < 0)
4531                         goto free_drop;
4532
4533                 if (rc == LNET_CREDIT_OK) {
4534                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4535                                      0, payload_length, payload_length);
4536                 }
4537                 return 0;
4538         }
4539
4540         lnet_net_unlock(cpt);
4541
4542         rc = lnet_parse_local(ni, msg);
4543         if (rc != 0)
4544                 goto free_drop;
4545         return 0;
4546
4547  free_drop:
4548         LASSERT(msg->msg_md == NULL);
4549         lnet_finalize(msg, rc);
4550
4551  drop:
4552         lnet_drop_message(ni, cpt, private, payload_length, type);
4553         return 0;
4554 }
4555 EXPORT_SYMBOL(lnet_parse);
4556
4557 void
4558 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4559 {
4560         while (!list_empty(head)) {
4561                 struct lnet_process_id id = {0};
4562                 struct lnet_msg *msg;
4563
4564                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4565                 list_del(&msg->msg_list);
4566
4567                 id.nid = msg->msg_hdr.src_nid;
4568                 id.pid = msg->msg_hdr.src_pid;
4569
4570                 LASSERT(msg->msg_md == NULL);
4571                 LASSERT(msg->msg_rx_delayed);
4572                 LASSERT(msg->msg_rxpeer != NULL);
4573                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4574
4575                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4576                       " offset %d length %d: %s\n",
4577                       libcfs_id2str(id),
4578                       msg->msg_hdr.msg.put.ptl_index,
4579                       msg->msg_hdr.msg.put.match_bits,
4580                       msg->msg_hdr.msg.put.offset,
4581                       msg->msg_hdr.payload_length, reason);
4582
4583                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4584                  * called lnet_drop_message(), so I just hang onto msg as well
4585                  * until that's done */
4586
4587                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4588                                   msg->msg_private, msg->msg_len,
4589                                   msg->msg_type);
4590
4591                 msg->msg_no_resend = true;
4592                 /*
4593                  * NB: message will not generate event because w/o attached MD,
4594                  * but we still should give error code so lnet_msg_decommit()
4595                  * can skip counters operations and other checks.
4596                  */
4597                 lnet_finalize(msg, -ENOENT);
4598         }
4599 }
4600
4601 void
4602 lnet_recv_delayed_msg_list(struct list_head *head)
4603 {
4604         while (!list_empty(head)) {
4605                 struct lnet_msg *msg;
4606                 struct lnet_process_id id;
4607
4608                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4609                 list_del(&msg->msg_list);
4610
4611                 /* md won't disappear under me, since each msg
4612                  * holds a ref on it */
4613
4614                 id.nid = msg->msg_hdr.src_nid;
4615                 id.pid = msg->msg_hdr.src_pid;
4616
4617                 LASSERT(msg->msg_rx_delayed);
4618                 LASSERT(msg->msg_md != NULL);
4619                 LASSERT(msg->msg_rxpeer != NULL);
4620                 LASSERT(msg->msg_rxni != NULL);
4621                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4622
4623                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4624                        "match %llu offset %d length %d.\n",
4625                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4626                         msg->msg_hdr.msg.put.match_bits,
4627                         msg->msg_hdr.msg.put.offset,
4628                         msg->msg_hdr.payload_length);
4629
4630                 lnet_recv_put(msg->msg_rxni, msg);
4631         }
4632 }
4633
4634 static void
4635 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4636                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4637 {
4638         s64 timeout_ns;
4639         bool new_entry = true;
4640         struct lnet_rsp_tracker *local_rspt;
4641
4642         /*
4643          * MD has a refcount taken by message so it's not going away.
4644          * The MD however can be looked up. We need to secure the access
4645          * to the md_rspt_ptr by taking the res_lock.
4646          * The rspt can be accessed without protection up to when it gets
4647          * added to the list.
4648          */
4649
4650         lnet_res_lock(cpt);
4651         local_rspt = md->md_rspt_ptr;
4652         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4653         if (local_rspt != NULL) {
4654                 /*
4655                  * we already have an rspt attached to the md, so we'll
4656                  * update the deadline on that one.
4657                  */
4658                 LIBCFS_FREE(rspt, sizeof(*rspt));
4659                 new_entry = false;
4660         } else {
4661                 /* new md */
4662                 rspt->rspt_mdh = mdh;
4663                 rspt->rspt_cpt = cpt;
4664                 /* store the rspt so we can access it when we get the REPLY */
4665                 md->md_rspt_ptr = rspt;
4666                 local_rspt = rspt;
4667         }
4668         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4669
4670         /*
4671          * add to the list of tracked responses. It's added to tail of the
4672          * list in order to expire all the older entries first.
4673          */
4674         lnet_net_lock(cpt);
4675         if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
4676                 list_del_init(&local_rspt->rspt_on_list);
4677         list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4678         lnet_net_unlock(cpt);
4679         lnet_res_unlock(cpt);
4680 }
4681
4682 /**
4683  * Initiate an asynchronous PUT operation.
4684  *
4685  * There are several events associated with a PUT: completion of the send on
4686  * the initiator node (LNET_EVENT_SEND), and when the send completes
4687  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4688  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4689  * used at the target node to indicate the completion of incoming data
4690  * delivery.
4691  *
4692  * The local events will be logged in the EQ associated with the MD pointed to
4693  * by \a mdh handle. Using a MD without an associated EQ results in these
4694  * events being discarded. In this case, the caller must have another
4695  * mechanism (e.g., a higher level protocol) for determining when it is safe
4696  * to modify the memory region associated with the MD.
4697  *
4698  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4699  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4700  *
4701  * \param self Indicates the NID of a local interface through which to send
4702  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4703  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4704  * must be "free floating" (See LNetMDBind()).
4705  * \param ack Controls whether an acknowledgment is requested.
4706  * Acknowledgments are only sent when they are requested by the initiating
4707  * process and the target MD enables them.
4708  * \param target A process identifier for the target process.
4709  * \param portal The index in the \a target's portal table.
4710  * \param match_bits The match bits to use for MD selection at the target
4711  * process.
4712  * \param offset The offset into the target MD (only used when the target
4713  * MD has the LNET_MD_MANAGE_REMOTE option set).
4714  * \param hdr_data 64 bits of user data that can be included in the message
4715  * header. This data is written to an event queue entry at the target if an
4716  * EQ is present on the matching MD.
4717  *
4718  * \retval  0      Success, and only in this case events will be generated
4719  * and logged to EQ (if it exists).
4720  * \retval -EIO    Simulated failure.
4721  * \retval -ENOMEM Memory allocation failure.
4722  * \retval -ENOENT Invalid MD object.
4723  *
4724  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4725  */
4726 int
4727 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4728         struct lnet_process_id target, unsigned int portal,
4729         __u64 match_bits, unsigned int offset,
4730         __u64 hdr_data)
4731 {
4732         struct lnet_msg *msg;
4733         struct lnet_libmd *md;
4734         int cpt;
4735         int rc;
4736         struct lnet_rsp_tracker *rspt = NULL;
4737
4738         LASSERT(the_lnet.ln_refcount > 0);
4739
4740         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4741             fail_peer(target.nid, 1)) {                 /* shall we now? */
4742                 CERROR("Dropping PUT to %s: simulated failure\n",
4743                        libcfs_id2str(target));
4744                 return -EIO;
4745         }
4746
4747         msg = lnet_msg_alloc();
4748         if (msg == NULL) {
4749                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4750                        libcfs_id2str(target));
4751                 return -ENOMEM;
4752         }
4753         msg->msg_vmflush = !!memory_pressure_get();
4754
4755         cpt = lnet_cpt_of_cookie(mdh.cookie);
4756
4757         if (ack == LNET_ACK_REQ) {
4758                 rspt = lnet_rspt_alloc(cpt);
4759                 if (!rspt) {
4760                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4761                                 libcfs_id2str(target));
4762                         return -ENOMEM;
4763                 }
4764                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4765         }
4766
4767         lnet_res_lock(cpt);
4768
4769         md = lnet_handle2md(&mdh);
4770         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4771                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4772                        match_bits, portal, libcfs_id2str(target),
4773                        md == NULL ? -1 : md->md_threshold);
4774                 if (md != NULL && md->md_me != NULL)
4775                         CERROR("Source MD also attached to portal %d\n",
4776                                md->md_me->me_portal);
4777                 lnet_res_unlock(cpt);
4778
4779                 LIBCFS_FREE(rspt, sizeof(*rspt));
4780                 lnet_msg_free(msg);
4781                 return -ENOENT;
4782         }
4783
4784         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4785
4786         lnet_msg_attach_md(msg, md, 0, 0);
4787
4788         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4789
4790         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4791         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4792         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4793         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4794
4795         /* NB handles only looked up by creator (no flips) */
4796         if (ack == LNET_ACK_REQ) {
4797                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4798                         the_lnet.ln_interface_cookie;
4799                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4800                         md->md_lh.lh_cookie;
4801         } else {
4802                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4803                         LNET_WIRE_HANDLE_COOKIE_NONE;
4804                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4805                         LNET_WIRE_HANDLE_COOKIE_NONE;
4806         }
4807
4808         lnet_res_unlock(cpt);
4809
4810         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4811
4812         if (ack == LNET_ACK_REQ)
4813                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4814
4815         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4816                                  CFS_FAIL_ONCE))
4817                 rc = -EIO;
4818         else
4819                 rc = lnet_send(self, msg, LNET_NID_ANY);
4820
4821         if (rc != 0) {
4822                 CNETERR("Error sending PUT to %s: %d\n",
4823                         libcfs_id2str(target), rc);
4824                 msg->msg_no_resend = true;
4825                 lnet_finalize(msg, rc);
4826         }
4827
4828         /* completion will be signalled by an event */
4829         return 0;
4830 }
4831 EXPORT_SYMBOL(LNetPut);
4832
4833 /*
4834  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4835  * returns a msg for the LND to pass to lnet_finalize() when the sink
4836  * data has been received.
4837  *
4838  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4839  * lnet_finalize() is called on it, so the LND must call this first
4840  */
4841 struct lnet_msg *
4842 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4843 {
4844         struct lnet_msg *msg = lnet_msg_alloc();
4845         struct lnet_libmd *getmd = getmsg->msg_md;
4846         struct lnet_process_id peer_id = getmsg->msg_target;
4847         int cpt;
4848
4849         LASSERT(!getmsg->msg_target_is_router);
4850         LASSERT(!getmsg->msg_routing);
4851
4852         if (msg == NULL) {
4853                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4854                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4855                 goto drop;
4856         }
4857
4858         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4859         lnet_res_lock(cpt);
4860
4861         LASSERT(getmd->md_refcount > 0);
4862
4863         if (getmd->md_threshold == 0) {
4864                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4865                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4866                         getmd);
4867                 lnet_res_unlock(cpt);
4868                 goto drop;
4869         }
4870
4871         LASSERT(getmd->md_offset == 0);
4872
4873         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4874                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4875
4876         /* setup information for lnet_build_msg_event */
4877         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4878         msg->msg_from = peer_id.nid;
4879         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4880         msg->msg_hdr.src_nid = peer_id.nid;
4881         msg->msg_hdr.payload_length = getmd->md_length;
4882         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4883
4884         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4885         lnet_res_unlock(cpt);
4886
4887         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4888
4889         lnet_net_lock(cpt);
4890         lnet_msg_commit(msg, cpt);
4891         lnet_net_unlock(cpt);
4892
4893         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4894
4895         return msg;
4896
4897  drop:
4898         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4899
4900         lnet_net_lock(cpt);
4901         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4902         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4903         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4904                 getmd->md_length;
4905         lnet_net_unlock(cpt);
4906
4907         if (msg != NULL)
4908                 lnet_msg_free(msg);
4909
4910         return NULL;
4911 }
4912 EXPORT_SYMBOL(lnet_create_reply_msg);
4913
4914 void
4915 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4916                        unsigned int len)
4917 {
4918         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4919          * completed and I know it. */
4920         LASSERT(reply != NULL);
4921         LASSERT(reply->msg_type == LNET_MSG_GET);
4922         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4923
4924         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4925          * the end of my buffer, I might as well be dead. */
4926         LASSERT(len <= reply->msg_ev.mlength);
4927
4928         reply->msg_ev.mlength = len;
4929 }
4930 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4931
4932 /**
4933  * Initiate an asynchronous GET operation.
4934  *
4935  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4936  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4937  * the target node in the REPLY has been written to local MD.
4938  *
4939  * On the target node, an LNET_EVENT_GET is logged when the GET request
4940  * arrives and is accepted into a MD.
4941  *
4942  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4943  * \param mdh A handle for the MD that describes the memory into which the
4944  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4945  *
4946  * \retval  0      Success, and only in this case events will be generated
4947  * and logged to EQ (if it exists) of the MD.
4948  * \retval -EIO    Simulated failure.
4949  * \retval -ENOMEM Memory allocation failure.
4950  * \retval -ENOENT Invalid MD object.
4951  */
4952 int
4953 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4954         struct lnet_process_id target, unsigned int portal,
4955         __u64 match_bits, unsigned int offset, bool recovery)
4956 {
4957         struct lnet_msg *msg;
4958         struct lnet_libmd *md;
4959         struct lnet_rsp_tracker *rspt;
4960         int cpt;
4961         int rc;
4962
4963         LASSERT(the_lnet.ln_refcount > 0);
4964
4965         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4966             fail_peer(target.nid, 1))                   /* shall we now? */
4967         {
4968                 CERROR("Dropping GET to %s: simulated failure\n",
4969                        libcfs_id2str(target));
4970                 return -EIO;
4971         }
4972
4973         msg = lnet_msg_alloc();
4974         if (!msg) {
4975                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4976                        libcfs_id2str(target));
4977                 return -ENOMEM;
4978         }
4979
4980         cpt = lnet_cpt_of_cookie(mdh.cookie);
4981
4982         rspt = lnet_rspt_alloc(cpt);
4983         if (!rspt) {
4984                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4985                        libcfs_id2str(target));
4986                 return -ENOMEM;
4987         }
4988         INIT_LIST_HEAD(&rspt->rspt_on_list);
4989
4990         msg->msg_recovery = recovery;
4991
4992         lnet_res_lock(cpt);
4993
4994         md = lnet_handle2md(&mdh);
4995         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4996                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4997                        match_bits, portal, libcfs_id2str(target),
4998                        md == NULL ? -1 : md->md_threshold);
4999                 if (md != NULL && md->md_me != NULL)
5000                         CERROR("REPLY MD also attached to portal %d\n",
5001                                md->md_me->me_portal);
5002
5003                 lnet_res_unlock(cpt);
5004
5005                 lnet_msg_free(msg);
5006                 LIBCFS_FREE(rspt, sizeof(*rspt));
5007                 return -ENOENT;
5008         }
5009
5010         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5011
5012         lnet_msg_attach_md(msg, md, 0, 0);
5013
5014         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5015
5016         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5017         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5018         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5019         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5020
5021         /* NB handles only looked up by creator (no flips) */
5022         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5023                 the_lnet.ln_interface_cookie;
5024         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5025                 md->md_lh.lh_cookie;
5026
5027         lnet_res_unlock(cpt);
5028
5029         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5030
5031         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5032
5033         rc = lnet_send(self, msg, LNET_NID_ANY);
5034         if (rc < 0) {
5035                 CNETERR("Error sending GET to %s: %d\n",
5036                         libcfs_id2str(target), rc);
5037                 msg->msg_no_resend = true;
5038                 lnet_finalize(msg, rc);
5039         }
5040
5041         /* completion will be signalled by an event */
5042         return 0;
5043 }
5044 EXPORT_SYMBOL(LNetGet);
5045
5046 /**
5047  * Calculate distance to node at \a dstnid.
5048  *
5049  * \param dstnid Target NID.
5050  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5051  * is saved here.
5052  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5053  * here.
5054  *
5055  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5056  * local_nid_dist_zero is set, which is the default.
5057  * \retval positives Distance to target NID, i.e. number of hops plus one.
5058  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5059  */
5060 int
5061 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5062 {
5063         struct list_head        *e;
5064         struct lnet_ni *ni = NULL;
5065         struct lnet_remotenet *rnet;
5066         __u32                   dstnet = LNET_NIDNET(dstnid);
5067         int                     hops;
5068         int                     cpt;
5069         __u32                   order = 2;
5070         struct list_head        *rn_list;
5071
5072         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5073          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5074          * keep order 0 free for 0@lo and order 1 free for a local NID
5075          * match */
5076
5077         LASSERT(the_lnet.ln_refcount > 0);
5078
5079         cpt = lnet_net_lock_current();
5080
5081         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5082                 if (ni->ni_nid == dstnid) {
5083                         if (srcnidp != NULL)
5084                                 *srcnidp = dstnid;
5085                         if (orderp != NULL) {
5086                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5087                                         *orderp = 0;
5088                                 else
5089                                         *orderp = 1;
5090                         }
5091                         lnet_net_unlock(cpt);
5092
5093                         return local_nid_dist_zero ? 0 : 1;
5094                 }
5095
5096                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5097                         /* Check if ni was originally created in
5098                          * current net namespace.
5099                          * If not, assign order above 0xffff0000,
5100                          * to make this ni not a priority. */
5101                         if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5102                                 order += 0xffff0000;
5103
5104                         if (srcnidp != NULL)
5105                                 *srcnidp = ni->ni_nid;
5106                         if (orderp != NULL)
5107                                 *orderp = order;
5108                         lnet_net_unlock(cpt);
5109                         return 1;
5110                 }
5111
5112                 order++;
5113         }
5114
5115         rn_list = lnet_net2rnethash(dstnet);
5116         list_for_each(e, rn_list) {
5117                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5118
5119                 if (rnet->lrn_net == dstnet) {
5120                         struct lnet_route *route;
5121                         struct lnet_route *shortest = NULL;
5122                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5123                         __u32 route_hops;
5124
5125                         LASSERT(!list_empty(&rnet->lrn_routes));
5126
5127                         list_for_each_entry(route, &rnet->lrn_routes,
5128                                             lr_list) {
5129                                 route_hops = route->lr_hops;
5130                                 if (route_hops == LNET_UNDEFINED_HOPS)
5131                                         route_hops = 1;
5132                                 if (shortest == NULL ||
5133                                     route_hops < shortest_hops) {
5134                                         shortest = route;
5135                                         shortest_hops = route_hops;
5136                                 }
5137                         }
5138
5139                         LASSERT(shortest != NULL);
5140                         hops = shortest_hops;
5141                         if (srcnidp != NULL) {
5142                                 struct lnet_net *net;
5143                                 net = lnet_get_net_locked(shortest->lr_lnet);
5144                                 LASSERT(net);
5145                                 ni = lnet_get_next_ni_locked(net, NULL);
5146                                 *srcnidp = ni->ni_nid;
5147                         }
5148                         if (orderp != NULL)
5149                                 *orderp = order;
5150                         lnet_net_unlock(cpt);
5151                         return hops + 1;
5152                 }
5153                 order++;
5154         }
5155
5156         lnet_net_unlock(cpt);
5157         return -EHOSTUNREACH;
5158 }
5159 EXPORT_SYMBOL(LNetDist);