lnet/lnet/lib-move.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/lnet/lib-move.c
  33  *
  34  * Data movement routines
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38
  39 #include <linux/pagemap.h>
  40
  41 #include <lnet/lib-lnet.h>
  42 #include <linux/nsproxy.h>
  43 #include <net/net_namespace.h>
  44
  45 static int local_nid_dist_zero = 1;
  46 module_param(local_nid_dist_zero, int, 0444);
  47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  48
  49 struct lnet_send_data {
  50         struct lnet_ni *sd_best_ni;
  51         struct lnet_peer_ni *sd_best_lpni;
  52         struct lnet_peer_ni *sd_final_dst_lpni;
  53         struct lnet_peer *sd_peer;
  54         struct lnet_peer *sd_gw_peer;
  55         struct lnet_peer_ni *sd_gw_lpni;
  56         struct lnet_peer_net *sd_peer_net;
  57         struct lnet_msg *sd_msg;
  58         lnet_nid_t sd_dst_nid;
  59         lnet_nid_t sd_src_nid;
  60         lnet_nid_t sd_rtr_nid;
  61         int sd_cpt;
  62         int sd_md_cpt;
  63         __u32 sd_send_case;
  64 };
  65
  66 static inline bool
  67 lnet_msg_is_response(struct lnet_msg *msg)
  68 {
  69         return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
  70 }
  71
  72 static inline struct lnet_comm_count *
  73 get_stats_counts(struct lnet_element_stats *stats,
  74                  enum lnet_stats_type stats_type)
  75 {
  76         switch (stats_type) {
  77         case LNET_STATS_TYPE_SEND:
  78                 return &stats->el_send_stats;
  79         case LNET_STATS_TYPE_RECV:
  80                 return &stats->el_recv_stats;
  81         case LNET_STATS_TYPE_DROP:
  82                 return &stats->el_drop_stats;
  83         default:
  84                 CERROR("Unknown stats type\n");
  85         }
  86
  87         return NULL;
  88 }
  89
  90 void lnet_incr_stats(struct lnet_element_stats *stats,
  91                      enum lnet_msg_type msg_type,
  92                      enum lnet_stats_type stats_type)
  93 {
  94         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
  95         if (!counts)
  96                 return;
  97
  98         switch (msg_type) {
  99         case LNET_MSG_ACK:
 100                 atomic_inc(&counts->co_ack_count);
 101                 break;
 102         case LNET_MSG_PUT:
 103                 atomic_inc(&counts->co_put_count);
 104                 break;
 105         case LNET_MSG_GET:
 106                 atomic_inc(&counts->co_get_count);
 107                 break;
 108         case LNET_MSG_REPLY:
 109                 atomic_inc(&counts->co_reply_count);
 110                 break;
 111         case LNET_MSG_HELLO:
 112                 atomic_inc(&counts->co_hello_count);
 113                 break;
 114         default:
 115                 CERROR("There is a BUG in the code. Unknown message type\n");
 116                 break;
 117         }
 118 }
 119
 120 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 121                      enum lnet_stats_type stats_type)
 122 {
 123         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 124         if (!counts)
 125                 return 0;
 126
 127         return (atomic_read(&counts->co_ack_count) +
 128                 atomic_read(&counts->co_put_count) +
 129                 atomic_read(&counts->co_get_count) +
 130                 atomic_read(&counts->co_reply_count) +
 131                 atomic_read(&counts->co_hello_count));
 132 }
 133
 134 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 135                                 struct lnet_comm_count *counts)
 136 {
 137         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 138         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 139         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 140         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 141         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 142 }
 143
 144 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 145                               struct lnet_element_stats *stats)
 146 {
 147         struct lnet_comm_count *counts;
 148
 149         LASSERT(msg_stats);
 150         LASSERT(stats);
 151
 152         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 153         if (!counts)
 154                 return;
 155         assign_stats(&msg_stats->im_send_stats, counts);
 156
 157         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 158         if (!counts)
 159                 return;
 160         assign_stats(&msg_stats->im_recv_stats, counts);
 161
 162         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 163         if (!counts)
 164                 return;
 165         assign_stats(&msg_stats->im_drop_stats, counts);
 166 }
 167
 168 int
 169 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 170 {
 171         struct lnet_test_peer *tp;
 172         struct list_head *el;
 173         struct list_head *next;
 174         LIST_HEAD(cull);
 175
 176         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 177         if (threshold != 0) {
 178                 /* Adding a new entry */
 179                 LIBCFS_ALLOC(tp, sizeof(*tp));
 180                 if (tp == NULL)
 181                         return -ENOMEM;
 182
 183                 tp->tp_nid = nid;
 184                 tp->tp_threshold = threshold;
 185
 186                 lnet_net_lock(0);
 187                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 188                 lnet_net_unlock(0);
 189                 return 0;
 190         }
 191
 192         lnet_net_lock(0);
 193
 194         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 195                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 196
 197                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 198                     nid == LNET_NID_ANY ||      /* removing all entries */
 199                     tp->tp_nid == nid) {        /* matched this one */
 200                         list_move(&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         LIST_HEAD(cull);
 222         int fail = 0;
 223
 224         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 225         lnet_net_lock(0);
 226
 227         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 228                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 229
 230                 if (tp->tp_threshold == 0) {
 231                         /* zombie entry */
 232                         if (outgoing) {
 233                                 /* only cull zombies on outgoing tests,
 234                                  * since we may be at interrupt priority on
 235                                  * incoming messages. */
 236                                 list_move(&tp->tp_list, &cull);
 237                         }
 238                         continue;
 239                 }
 240
 241                 if (tp->tp_nid == LNET_NID_ANY ||       /* fail every peer */
 242                     nid == tp->tp_nid) {                /* fail this peer */
 243                         fail = 1;
 244
 245                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 246                                 tp->tp_threshold--;
 247                                 if (outgoing &&
 248                                     tp->tp_threshold == 0) {
 249                                         /* see above */
 250                                         list_move(&tp->tp_list, &cull);
 251                                 }
 252                         }
 253                         break;
 254                 }
 255         }
 256
 257         lnet_net_unlock(0);
 258
 259         while (!list_empty(&cull)) {
 260                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 261                 list_del(&tp->tp_list);
 262
 263                 LIBCFS_FREE(tp, sizeof(*tp));
 264         }
 265
 266         return fail;
 267 }
 268
 269 unsigned int
 270 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 271 {
 272         unsigned int nob = 0;
 273
 274         LASSERT(niov == 0 || iov != NULL);
 275         while (niov-- > 0)
 276                 nob += (iov++)->iov_len;
 277
 278         return (nob);
 279 }
 280 EXPORT_SYMBOL(lnet_iov_nob);
 281
 282 void
 283 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 284                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 285                   unsigned int nob)
 286 {
 287         /* NB diov, siov are READ-ONLY */
 288         unsigned int this_nob;
 289
 290         if (nob == 0)
 291                 return;
 292
 293         /* skip complete frags before 'doffset' */
 294         LASSERT(ndiov > 0);
 295         while (doffset >= diov->iov_len) {
 296                 doffset -= diov->iov_len;
 297                 diov++;
 298                 ndiov--;
 299                 LASSERT(ndiov > 0);
 300         }
 301
 302         /* skip complete frags before 'soffset' */
 303         LASSERT(nsiov > 0);
 304         while (soffset >= siov->iov_len) {
 305                 soffset -= siov->iov_len;
 306                 siov++;
 307                 nsiov--;
 308                 LASSERT(nsiov > 0);
 309         }
 310
 311         do {
 312                 LASSERT(ndiov > 0);
 313                 LASSERT(nsiov > 0);
 314                 this_nob = min3((unsigned int)diov->iov_len - doffset,
 315                                 (unsigned int)siov->iov_len - soffset,
 316                                 nob);
 317
 318                 memcpy((char *)diov->iov_base + doffset,
 319                        (char *)siov->iov_base + soffset, this_nob);
 320                 nob -= this_nob;
 321
 322                 if (diov->iov_len > doffset + this_nob) {
 323                         doffset += this_nob;
 324                 } else {
 325                         diov++;
 326                         ndiov--;
 327                         doffset = 0;
 328                 }
 329
 330                 if (siov->iov_len > soffset + this_nob) {
 331                         soffset += this_nob;
 332                 } else {
 333                         siov++;
 334                         nsiov--;
 335                         soffset = 0;
 336                 }
 337         } while (nob > 0);
 338 }
 339 EXPORT_SYMBOL(lnet_copy_iov2iov);
 340
 341 int
 342 lnet_extract_iov(int dst_niov, struct kvec *dst,
 343                  int src_niov, struct kvec *src,
 344                  unsigned int offset, unsigned int len)
 345 {
 346         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 347          * for exactly 'len' bytes, and return the number of entries.
 348          * NB not destructive to 'src' */
 349         unsigned int    frag_len;
 350         unsigned int    niov;
 351
 352         if (len == 0)                           /* no data => */
 353                 return (0);                     /* no frags */
 354
 355         LASSERT(src_niov > 0);
 356         while (offset >= src->iov_len) {      /* skip initial frags */
 357                 offset -= src->iov_len;
 358                 src_niov--;
 359                 src++;
 360                 LASSERT(src_niov > 0);
 361         }
 362
 363         niov = 1;
 364         for (;;) {
 365                 LASSERT(src_niov > 0);
 366                 LASSERT((int)niov <= dst_niov);
 367
 368                 frag_len = src->iov_len - offset;
 369                 dst->iov_base = ((char *)src->iov_base) + offset;
 370
 371                 if (len <= frag_len) {
 372                         dst->iov_len = len;
 373                         return (niov);
 374                 }
 375
 376                 dst->iov_len = frag_len;
 377
 378                 len -= frag_len;
 379                 dst++;
 380                 src++;
 381                 niov++;
 382                 src_niov--;
 383                 offset = 0;
 384         }
 385 }
 386 EXPORT_SYMBOL(lnet_extract_iov);
 387
 388
 389 unsigned int
 390 lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
 391 {
 392         unsigned int  nob = 0;
 393
 394         LASSERT(niov == 0 || kiov != NULL);
 395         while (niov-- > 0)
 396                 nob += (kiov++)->kiov_len;
 397
 398         return (nob);
 399 }
 400 EXPORT_SYMBOL(lnet_kiov_nob);
 401
 402 void
 403 lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 404                     unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 405                     unsigned int nob)
 406 {
 407         /* NB diov, siov are READ-ONLY */
 408         unsigned int    this_nob;
 409         char           *daddr = NULL;
 410         char           *saddr = NULL;
 411
 412         if (nob == 0)
 413                 return;
 414
 415         LASSERT (!in_interrupt ());
 416
 417         LASSERT (ndiov > 0);
 418         while (doffset >= diov->kiov_len) {
 419                 doffset -= diov->kiov_len;
 420                 diov++;
 421                 ndiov--;
 422                 LASSERT(ndiov > 0);
 423         }
 424
 425         LASSERT(nsiov > 0);
 426         while (soffset >= siov->kiov_len) {
 427                 soffset -= siov->kiov_len;
 428                 siov++;
 429                 nsiov--;
 430                 LASSERT(nsiov > 0);
 431         }
 432
 433         do {
 434                 LASSERT(ndiov > 0);
 435                 LASSERT(nsiov > 0);
 436                 this_nob = min3(diov->kiov_len - doffset,
 437                                 siov->kiov_len - soffset,
 438                                 nob);
 439
 440                 if (daddr == NULL)
 441                         daddr = ((char *)kmap(diov->kiov_page)) +
 442                                 diov->kiov_offset + doffset;
 443                 if (saddr == NULL)
 444                         saddr = ((char *)kmap(siov->kiov_page)) +
 445                                 siov->kiov_offset + soffset;
 446
 447                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 448                  * However in practice at least one of the kiovs will be mapped
 449                  * kernel pages and the map/unmap will be NOOPs */
 450
 451                 memcpy (daddr, saddr, this_nob);
 452                 nob -= this_nob;
 453
 454                 if (diov->kiov_len > doffset + this_nob) {
 455                         daddr += this_nob;
 456                         doffset += this_nob;
 457                 } else {
 458                         kunmap(diov->kiov_page);
 459                         daddr = NULL;
 460                         diov++;
 461                         ndiov--;
 462                         doffset = 0;
 463                 }
 464
 465                 if (siov->kiov_len > soffset + this_nob) {
 466                         saddr += this_nob;
 467                         soffset += this_nob;
 468                 } else {
 469                         kunmap(siov->kiov_page);
 470                         saddr = NULL;
 471                         siov++;
 472                         nsiov--;
 473                         soffset = 0;
 474                 }
 475         } while (nob > 0);
 476
 477         if (daddr != NULL)
 478                 kunmap(diov->kiov_page);
 479         if (saddr != NULL)
 480                 kunmap(siov->kiov_page);
 481 }
 482 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 483
 484 void
 485 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 486                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 487                     unsigned int nob)
 488 {
 489         /* NB iov, kiov are READ-ONLY */
 490         unsigned int    this_nob;
 491         char           *addr = NULL;
 492
 493         if (nob == 0)
 494                 return;
 495
 496         LASSERT (!in_interrupt ());
 497
 498         LASSERT (niov > 0);
 499         while (iovoffset >= iov->iov_len) {
 500                 iovoffset -= iov->iov_len;
 501                 iov++;
 502                 niov--;
 503                 LASSERT(niov > 0);
 504         }
 505
 506         LASSERT(nkiov > 0);
 507         while (kiovoffset >= kiov->kiov_len) {
 508                 kiovoffset -= kiov->kiov_len;
 509                 kiov++;
 510                 nkiov--;
 511                 LASSERT(nkiov > 0);
 512         }
 513
 514         do {
 515                 LASSERT(niov > 0);
 516                 LASSERT(nkiov > 0);
 517                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 518                                 (unsigned int)kiov->kiov_len - kiovoffset,
 519                                 nob);
 520
 521                 if (addr == NULL)
 522                         addr = ((char *)kmap(kiov->kiov_page)) +
 523                                 kiov->kiov_offset + kiovoffset;
 524
 525                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 526                 nob -= this_nob;
 527
 528                 if (iov->iov_len > iovoffset + this_nob) {
 529                         iovoffset += this_nob;
 530                 } else {
 531                         iov++;
 532                         niov--;
 533                         iovoffset = 0;
 534                 }
 535
 536                 if (kiov->kiov_len > kiovoffset + this_nob) {
 537                         addr += this_nob;
 538                         kiovoffset += this_nob;
 539                 } else {
 540                         kunmap(kiov->kiov_page);
 541                         addr = NULL;
 542                         kiov++;
 543                         nkiov--;
 544                         kiovoffset = 0;
 545                 }
 546
 547         } while (nob > 0);
 548
 549         if (addr != NULL)
 550                 kunmap(kiov->kiov_page);
 551 }
 552 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 553
 554 void
 555 lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 556                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 557                    unsigned int nob)
 558 {
 559         /* NB kiov, iov are READ-ONLY */
 560         unsigned int    this_nob;
 561         char           *addr = NULL;
 562
 563         if (nob == 0)
 564                 return;
 565
 566         LASSERT (!in_interrupt ());
 567
 568         LASSERT (nkiov > 0);
 569         while (kiovoffset >= kiov->kiov_len) {
 570                 kiovoffset -= kiov->kiov_len;
 571                 kiov++;
 572                 nkiov--;
 573                 LASSERT(nkiov > 0);
 574         }
 575
 576         LASSERT(niov > 0);
 577         while (iovoffset >= iov->iov_len) {
 578                 iovoffset -= iov->iov_len;
 579                 iov++;
 580                 niov--;
 581                 LASSERT(niov > 0);
 582         }
 583
 584         do {
 585                 LASSERT(nkiov > 0);
 586                 LASSERT(niov > 0);
 587                 this_nob = min3((unsigned int)kiov->kiov_len - kiovoffset,
 588                                 (unsigned int)iov->iov_len - iovoffset,
 589                                 nob);
 590
 591                 if (addr == NULL)
 592                         addr = ((char *)kmap(kiov->kiov_page)) +
 593                                 kiov->kiov_offset + kiovoffset;
 594
 595                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 596                 nob -= this_nob;
 597
 598                 if (kiov->kiov_len > kiovoffset + this_nob) {
 599                         addr += this_nob;
 600                         kiovoffset += this_nob;
 601                 } else {
 602                         kunmap(kiov->kiov_page);
 603                         addr = NULL;
 604                         kiov++;
 605                         nkiov--;
 606                         kiovoffset = 0;
 607                 }
 608
 609                 if (iov->iov_len > iovoffset + this_nob) {
 610                         iovoffset += this_nob;
 611                 } else {
 612                         iov++;
 613                         niov--;
 614                         iovoffset = 0;
 615                 }
 616         } while (nob > 0);
 617
 618         if (addr != NULL)
 619                 kunmap(kiov->kiov_page);
 620 }
 621 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 622
 623 int
 624 lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 625                   int src_niov, lnet_kiov_t *src,
 626                   unsigned int offset, unsigned int len)
 627 {
 628         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 629          * for exactly 'len' bytes, and return the number of entries.
 630          * NB not destructive to 'src' */
 631         unsigned int    frag_len;
 632         unsigned int    niov;
 633
 634         if (len == 0)                           /* no data => */
 635                 return (0);                     /* no frags */
 636
 637         LASSERT(src_niov > 0);
 638         while (offset >= src->kiov_len) {      /* skip initial frags */
 639                 offset -= src->kiov_len;
 640                 src_niov--;
 641                 src++;
 642                 LASSERT(src_niov > 0);
 643         }
 644
 645         niov = 1;
 646         for (;;) {
 647                 LASSERT(src_niov > 0);
 648                 LASSERT((int)niov <= dst_niov);
 649
 650                 frag_len = src->kiov_len - offset;
 651                 dst->kiov_page = src->kiov_page;
 652                 dst->kiov_offset = src->kiov_offset + offset;
 653
 654                 if (len <= frag_len) {
 655                         dst->kiov_len = len;
 656                         LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 657                         return niov;
 658                 }
 659
 660                 dst->kiov_len = frag_len;
 661                 LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
 662
 663                 len -= frag_len;
 664                 dst++;
 665                 src++;
 666                 niov++;
 667                 src_niov--;
 668                 offset = 0;
 669         }
 670 }
 671 EXPORT_SYMBOL(lnet_extract_kiov);
 672
 673 void
 674 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 675              int delayed, unsigned int offset, unsigned int mlen,
 676              unsigned int rlen)
 677 {
 678         unsigned int  niov = 0;
 679         struct kvec *iov = NULL;
 680         lnet_kiov_t  *kiov = NULL;
 681         int           rc;
 682
 683         LASSERT (!in_interrupt ());
 684         LASSERT (mlen == 0 || msg != NULL);
 685
 686         if (msg != NULL) {
 687                 LASSERT(msg->msg_receiving);
 688                 LASSERT(!msg->msg_sending);
 689                 LASSERT(rlen == msg->msg_len);
 690                 LASSERT(mlen <= msg->msg_len);
 691                 LASSERT(msg->msg_offset == offset);
 692                 LASSERT(msg->msg_wanted == mlen);
 693
 694                 msg->msg_receiving = 0;
 695
 696                 if (mlen != 0) {
 697                         niov = msg->msg_niov;
 698                         iov  = msg->msg_iov;
 699                         kiov = msg->msg_kiov;
 700
 701                         LASSERT (niov > 0);
 702                         LASSERT ((iov == NULL) != (kiov == NULL));
 703                 }
 704         }
 705
 706         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 707                                              niov, iov, kiov, offset, mlen,
 708                                              rlen);
 709         if (rc < 0)
 710                 lnet_finalize(msg, rc);
 711 }
 712
 713 static void
 714 lnet_setpayloadbuffer(struct lnet_msg *msg)
 715 {
 716         struct lnet_libmd *md = msg->msg_md;
 717
 718         LASSERT(msg->msg_len > 0);
 719         LASSERT(!msg->msg_routing);
 720         LASSERT(md != NULL);
 721         LASSERT(msg->msg_niov == 0);
 722         LASSERT(msg->msg_iov == NULL);
 723         LASSERT(msg->msg_kiov == NULL);
 724
 725         msg->msg_niov = md->md_niov;
 726         if ((md->md_options & LNET_MD_KIOV) != 0)
 727                 msg->msg_kiov = md->md_iov.kiov;
 728         else
 729                 msg->msg_iov = md->md_iov.iov;
 730 }
 731
 732 void
 733 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 734                unsigned int offset, unsigned int len)
 735 {
 736         msg->msg_type = type;
 737         msg->msg_target = target;
 738         msg->msg_len = len;
 739         msg->msg_offset = offset;
 740
 741         if (len != 0)
 742                 lnet_setpayloadbuffer(msg);
 743
 744         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 745         msg->msg_hdr.type           = cpu_to_le32(type);
 746         /* dest_nid will be overwritten by lnet_select_pathway() */
 747         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 748         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 749         /* src_nid will be set later */
 750         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 751         msg->msg_hdr.payload_length = cpu_to_le32(len);
 752 }
 753
 754 static void
 755 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 756 {
 757         void   *priv = msg->msg_private;
 758         int rc;
 759
 760         LASSERT (!in_interrupt ());
 761         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 762                  (msg->msg_txcredit && msg->msg_peertxcredit));
 763
 764         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 765         if (rc < 0) {
 766                 msg->msg_no_resend = true;
 767                 lnet_finalize(msg, rc);
 768         }
 769 }
 770
 771 static int
 772 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 773 {
 774         int     rc;
 775
 776         LASSERT(!msg->msg_sending);
 777         LASSERT(msg->msg_receiving);
 778         LASSERT(!msg->msg_rx_ready_delay);
 779         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 780
 781         msg->msg_rx_ready_delay = 1;
 782         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 783                                                   &msg->msg_private);
 784         if (rc != 0) {
 785                 CERROR("recv from %s / send to %s aborted: "
 786                        "eager_recv failed %d\n",
 787                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 788                        libcfs_id2str(msg->msg_target), rc);
 789                 LASSERT(rc < 0); /* required by my callers */
 790         }
 791
 792         return rc;
 793 }
 794
 795 static bool
 796 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 797 {
 798         time64_t deadline;
 799
 800         deadline = lpni->lpni_last_alive +
 801                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 802
 803         /*
 804          * assume peer_ni is alive as long as we're within the configured
 805          * peer timeout
 806          */
 807         if (deadline > now)
 808                 return false;
 809
 810         return true;
 811 }
 812
 813 /* NB: returns 1 when alive, 0 when dead, negative when error;
 814  *     may drop the lnet_net_lock */
 815 static int
 816 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 817                        struct lnet_msg *msg)
 818 {
 819         time64_t now = ktime_get_seconds();
 820
 821         if (!lnet_peer_aliveness_enabled(lpni))
 822                 return -ENODEV;
 823
 824         /*
 825          * If we're resending a message, let's attempt to send it even if
 826          * the peer is down to fulfill our resend quota on the message
 827          */
 828         if (msg->msg_retry_count > 0)
 829                 return 1;
 830
 831         /* try and send recovery messages irregardless */
 832         if (msg->msg_recovery)
 833                 return 1;
 834
 835         /* always send any responses */
 836         if (lnet_msg_is_response(msg))
 837                 return 1;
 838
 839         if (!lnet_is_peer_deadline_passed(lpni, now))
 840                 return true;
 841
 842         return lnet_is_peer_ni_alive(lpni);
 843 }
 844
 845 /**
 846  * \param msg The message to be sent.
 847  * \param do_send True if lnet_ni_send() should be called in this function.
 848  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 849  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 850  *
 851  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 852  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 853  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 854  * \retval -ECANCELED If the MD of the message has been unlinked.
 855  */
 856 static int
 857 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 858 {
 859         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 860         struct lnet_ni          *ni = msg->msg_txni;
 861         int                     cpt = msg->msg_tx_cpt;
 862         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 863
 864         /* non-lnet_send() callers have checked before */
 865         LASSERT(!do_send || msg->msg_tx_delayed);
 866         LASSERT(!msg->msg_receiving);
 867         LASSERT(msg->msg_tx_committed);
 868         /* can't get here if we're sending to the loopback interface */
 869         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 870
 871         /* NB 'lp' is always the next hop */
 872         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 873             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 874                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 875                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 876                         msg->msg_len;
 877                 lnet_net_unlock(cpt);
 878                 if (msg->msg_txpeer)
 879                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 880                                         msg->msg_type,
 881                                         LNET_STATS_TYPE_DROP);
 882                 if (msg->msg_txni)
 883                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 884                                         msg->msg_type,
 885                                         LNET_STATS_TYPE_DROP);
 886
 887                 CNETERR("Dropping message for %s: peer not alive\n",
 888                         libcfs_id2str(msg->msg_target));
 889                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 890                 if (do_send)
 891                         lnet_finalize(msg, -EHOSTUNREACH);
 892
 893                 lnet_net_lock(cpt);
 894                 return -EHOSTUNREACH;
 895         }
 896
 897         if (msg->msg_md != NULL &&
 898             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 899                 lnet_net_unlock(cpt);
 900
 901                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 902                         "called on the MD/ME.\n",
 903                         libcfs_id2str(msg->msg_target));
 904                 if (do_send) {
 905                         msg->msg_no_resend = true;
 906                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 907                                msg, libcfs_id2str(msg->msg_target));
 908                         lnet_finalize(msg, -ECANCELED);
 909                 }
 910
 911                 lnet_net_lock(cpt);
 912                 return -ECANCELED;
 913         }
 914
 915         if (!msg->msg_peertxcredit) {
 916                 spin_lock(&lp->lpni_lock);
 917                 LASSERT((lp->lpni_txcredits < 0) ==
 918                         !list_empty(&lp->lpni_txq));
 919
 920                 msg->msg_peertxcredit = 1;
 921                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 922                 lp->lpni_txcredits--;
 923
 924                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 925                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 926
 927                 if (lp->lpni_txcredits < 0) {
 928                         msg->msg_tx_delayed = 1;
 929                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 930                         spin_unlock(&lp->lpni_lock);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933                 spin_unlock(&lp->lpni_lock);
 934         }
 935
 936         if (!msg->msg_txcredit) {
 937                 LASSERT((tq->tq_credits < 0) ==
 938                         !list_empty(&tq->tq_delayed));
 939
 940                 msg->msg_txcredit = 1;
 941                 tq->tq_credits--;
 942                 atomic_dec(&ni->ni_tx_credits);
 943
 944                 if (tq->tq_credits < tq->tq_credits_min)
 945                         tq->tq_credits_min = tq->tq_credits;
 946
 947                 if (tq->tq_credits < 0) {
 948                         msg->msg_tx_delayed = 1;
 949                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 950                         return LNET_CREDIT_WAIT;
 951                 }
 952         }
 953
 954         /* unset the tx_delay flag as we're going to send it now */
 955         msg->msg_tx_delayed = 0;
 956
 957         if (do_send) {
 958                 lnet_net_unlock(cpt);
 959                 lnet_ni_send(ni, msg);
 960                 lnet_net_lock(cpt);
 961         }
 962         return LNET_CREDIT_OK;
 963 }
 964
 965
 966 static struct lnet_rtrbufpool *
 967 lnet_msg2bufpool(struct lnet_msg *msg)
 968 {
 969         struct lnet_rtrbufpool  *rbp;
 970         int                     cpt;
 971
 972         LASSERT(msg->msg_rx_committed);
 973
 974         cpt = msg->msg_rx_cpt;
 975         rbp = &the_lnet.ln_rtrpools[cpt][0];
 976
 977         LASSERT(msg->msg_len <= LNET_MTU);
 978         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 979                 rbp++;
 980                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 981         }
 982
 983         return rbp;
 984 }
 985
 986 static int
 987 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 988 {
 989         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 990          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 991          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 992          * received or OK to receive */
 993         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 994         struct lnet_peer *lp;
 995         struct lnet_rtrbufpool *rbp;
 996         struct lnet_rtrbuf *rb;
 997
 998         LASSERT(msg->msg_iov == NULL);
 999         LASSERT(msg->msg_kiov == NULL);
1000         LASSERT(msg->msg_niov == 0);
1001         LASSERT(msg->msg_routing);
1002         LASSERT(msg->msg_receiving);
1003         LASSERT(!msg->msg_sending);
1004         LASSERT(lpni->lpni_peer_net);
1005         LASSERT(lpni->lpni_peer_net->lpn_peer);
1006
1007         lp = lpni->lpni_peer_net->lpn_peer;
1008
1009         /* non-lnet_parse callers only receive delayed messages */
1010         LASSERT(!do_recv || msg->msg_rx_delayed);
1011
1012         if (!msg->msg_peerrtrcredit) {
1013                 /* lpni_lock protects the credit manipulation */
1014                 spin_lock(&lpni->lpni_lock);
1015                 /* lp_lock protects the lp_rtrq */
1016                 spin_lock(&lp->lp_lock);
1017
1018                 msg->msg_peerrtrcredit = 1;
1019                 lpni->lpni_rtrcredits--;
1020                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1021                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1022
1023                 if (lpni->lpni_rtrcredits < 0) {
1024                         /* must have checked eager_recv before here */
1025                         LASSERT(msg->msg_rx_ready_delay);
1026                         msg->msg_rx_delayed = 1;
1027                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1028                         spin_unlock(&lp->lp_lock);
1029                         spin_unlock(&lpni->lpni_lock);
1030                         return LNET_CREDIT_WAIT;
1031                 }
1032                 spin_unlock(&lp->lp_lock);
1033                 spin_unlock(&lpni->lpni_lock);
1034         }
1035
1036         rbp = lnet_msg2bufpool(msg);
1037
1038         if (!msg->msg_rtrcredit) {
1039                 msg->msg_rtrcredit = 1;
1040                 rbp->rbp_credits--;
1041                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1042                         rbp->rbp_mincredits = rbp->rbp_credits;
1043
1044                 if (rbp->rbp_credits < 0) {
1045                         /* must have checked eager_recv before here */
1046                         LASSERT(msg->msg_rx_ready_delay);
1047                         msg->msg_rx_delayed = 1;
1048                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1049                         return LNET_CREDIT_WAIT;
1050                 }
1051         }
1052
1053         LASSERT(!list_empty(&rbp->rbp_bufs));
1054         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1055         list_del(&rb->rb_list);
1056
1057         msg->msg_niov = rbp->rbp_npages;
1058         msg->msg_kiov = &rb->rb_kiov[0];
1059
1060         /* unset the msg-rx_delayed flag since we're receiving the message */
1061         msg->msg_rx_delayed = 0;
1062
1063         if (do_recv) {
1064                 int cpt = msg->msg_rx_cpt;
1065
1066                 lnet_net_unlock(cpt);
1067                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1068                              0, msg->msg_len, msg->msg_len);
1069                 lnet_net_lock(cpt);
1070         }
1071         return LNET_CREDIT_OK;
1072 }
1073
1074 void
1075 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1076 {
1077         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1078         struct lnet_ni          *txni = msg->msg_txni;
1079         struct lnet_msg         *msg2;
1080
1081         if (msg->msg_txcredit) {
1082                 struct lnet_ni       *ni = msg->msg_txni;
1083                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1084
1085                 /* give back NI txcredits */
1086                 msg->msg_txcredit = 0;
1087
1088                 LASSERT((tq->tq_credits < 0) ==
1089                         !list_empty(&tq->tq_delayed));
1090
1091                 tq->tq_credits++;
1092                 atomic_inc(&ni->ni_tx_credits);
1093                 if (tq->tq_credits <= 0) {
1094                         msg2 = list_entry(tq->tq_delayed.next,
1095                                           struct lnet_msg, msg_list);
1096                         list_del(&msg2->msg_list);
1097
1098                         LASSERT(msg2->msg_txni == ni);
1099                         LASSERT(msg2->msg_tx_delayed);
1100                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1101
1102                         (void) lnet_post_send_locked(msg2, 1);
1103                 }
1104         }
1105
1106         if (msg->msg_peertxcredit) {
1107                 /* give back peer txcredits */
1108                 msg->msg_peertxcredit = 0;
1109
1110                 spin_lock(&txpeer->lpni_lock);
1111                 LASSERT((txpeer->lpni_txcredits < 0) ==
1112                         !list_empty(&txpeer->lpni_txq));
1113
1114                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1115                 LASSERT(txpeer->lpni_txqnob >= 0);
1116
1117                 txpeer->lpni_txcredits++;
1118                 if (txpeer->lpni_txcredits <= 0) {
1119                         int msg2_cpt;
1120
1121                         msg2 = list_entry(txpeer->lpni_txq.next,
1122                                               struct lnet_msg, msg_list);
1123                         list_del(&msg2->msg_list);
1124                         spin_unlock(&txpeer->lpni_lock);
1125
1126                         LASSERT(msg2->msg_txpeer == txpeer);
1127                         LASSERT(msg2->msg_tx_delayed);
1128
1129                         msg2_cpt = msg2->msg_tx_cpt;
1130
1131                         /*
1132                          * The msg_cpt can be different from the msg2_cpt
1133                          * so we need to make sure we lock the correct cpt
1134                          * for msg2.
1135                          * Once we call lnet_post_send_locked() it is no
1136                          * longer safe to access msg2, since it could've
1137                          * been freed by lnet_finalize(), but we still
1138                          * need to relock the correct cpt, so we cache the
1139                          * msg2_cpt for the purpose of the check that
1140                          * follows the call to lnet_pose_send_locked().
1141                          */
1142                         if (msg2_cpt != msg->msg_tx_cpt) {
1143                                 lnet_net_unlock(msg->msg_tx_cpt);
1144                                 lnet_net_lock(msg2_cpt);
1145                         }
1146                         (void) lnet_post_send_locked(msg2, 1);
1147                         if (msg2_cpt != msg->msg_tx_cpt) {
1148                                 lnet_net_unlock(msg2_cpt);
1149                                 lnet_net_lock(msg->msg_tx_cpt);
1150                         }
1151                 } else {
1152                         spin_unlock(&txpeer->lpni_lock);
1153                 }
1154         }
1155
1156         if (txni != NULL) {
1157                 msg->msg_txni = NULL;
1158                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1159         }
1160
1161         if (txpeer != NULL) {
1162                 msg->msg_txpeer = NULL;
1163                 lnet_peer_ni_decref_locked(txpeer);
1164         }
1165 }
1166
1167 void
1168 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1169 {
1170         struct lnet_msg *msg;
1171
1172         if (list_empty(&rbp->rbp_msgs))
1173                 return;
1174         msg = list_entry(rbp->rbp_msgs.next,
1175                          struct lnet_msg, msg_list);
1176         list_del(&msg->msg_list);
1177
1178         (void)lnet_post_routed_recv_locked(msg, 1);
1179 }
1180
1181 void
1182 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1183 {
1184         struct lnet_msg *msg;
1185         struct lnet_msg *tmp;
1186
1187         lnet_net_unlock(cpt);
1188
1189         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1190                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1191                              0, 0, 0, msg->msg_hdr.payload_length);
1192                 list_del_init(&msg->msg_list);
1193                 msg->msg_no_resend = true;
1194                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1195                 lnet_finalize(msg, -ECANCELED);
1196         }
1197
1198         lnet_net_lock(cpt);
1199 }
1200
1201 void
1202 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1203 {
1204         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1205         struct lnet_peer *lp;
1206         struct lnet_ni *rxni = msg->msg_rxni;
1207         struct lnet_msg *msg2;
1208
1209         if (msg->msg_rtrcredit) {
1210                 /* give back global router credits */
1211                 struct lnet_rtrbuf *rb;
1212                 struct lnet_rtrbufpool *rbp;
1213
1214                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1215                  * there until it gets one allocated, or aborts the wait
1216                  * itself */
1217                 LASSERT(msg->msg_kiov != NULL);
1218
1219                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1220                 rbp = rb->rb_pool;
1221
1222                 msg->msg_kiov = NULL;
1223                 msg->msg_rtrcredit = 0;
1224
1225                 LASSERT(rbp == lnet_msg2bufpool(msg));
1226
1227                 LASSERT((rbp->rbp_credits > 0) ==
1228                         !list_empty(&rbp->rbp_bufs));
1229
1230                 /* If routing is now turned off, we just drop this buffer and
1231                  * don't bother trying to return credits.  */
1232                 if (!the_lnet.ln_routing) {
1233                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1234                         goto routing_off;
1235                 }
1236
1237                 /* It is possible that a user has lowered the desired number of
1238                  * buffers in this pool.  Make sure we never put back
1239                  * more buffers than the stated number. */
1240                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1241                         /* Discard this buffer so we don't have too
1242                          * many. */
1243                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1244                         rbp->rbp_nbuffers--;
1245                 } else {
1246                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1247                         rbp->rbp_credits++;
1248                         if (rbp->rbp_credits <= 0)
1249                                 lnet_schedule_blocked_locked(rbp);
1250                 }
1251         }
1252
1253 routing_off:
1254         if (msg->msg_peerrtrcredit) {
1255                 LASSERT(rxpeerni);
1256                 LASSERT(rxpeerni->lpni_peer_net);
1257                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1258
1259                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1260
1261                 /* give back peer router credits */
1262                 msg->msg_peerrtrcredit = 0;
1263
1264                 spin_lock(&rxpeerni->lpni_lock);
1265                 spin_lock(&lp->lp_lock);
1266
1267                 rxpeerni->lpni_rtrcredits++;
1268
1269                 /* drop all messages which are queued to be routed on that
1270                  * peer. */
1271                 if (!the_lnet.ln_routing) {
1272                         LIST_HEAD(drop);
1273                         list_splice_init(&lp->lp_rtrq, &drop);
1274                         spin_unlock(&lp->lp_lock);
1275                         spin_unlock(&rxpeerni->lpni_lock);
1276                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1277                 } else if (!list_empty(&lp->lp_rtrq)) {
1278                         int msg2_cpt;
1279
1280                         msg2 = list_entry(lp->lp_rtrq.next,
1281                                           struct lnet_msg, msg_list);
1282                         list_del(&msg2->msg_list);
1283                         msg2_cpt = msg2->msg_rx_cpt;
1284                         spin_unlock(&lp->lp_lock);
1285                         spin_unlock(&rxpeerni->lpni_lock);
1286                         /*
1287                          * messages on the lp_rtrq can be from any NID in
1288                          * the peer, which means they might have different
1289                          * cpts. We need to make sure we lock the right
1290                          * one.
1291                          */
1292                         if (msg2_cpt != msg->msg_rx_cpt) {
1293                                 lnet_net_unlock(msg->msg_rx_cpt);
1294                                 lnet_net_lock(msg2_cpt);
1295                         }
1296                         (void) lnet_post_routed_recv_locked(msg2, 1);
1297                         if (msg2_cpt != msg->msg_rx_cpt) {
1298                                 lnet_net_unlock(msg2_cpt);
1299                                 lnet_net_lock(msg->msg_rx_cpt);
1300                         }
1301                 } else {
1302                         spin_unlock(&lp->lp_lock);
1303                         spin_unlock(&rxpeerni->lpni_lock);
1304                 }
1305         }
1306         if (rxni != NULL) {
1307                 msg->msg_rxni = NULL;
1308                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1309         }
1310         if (rxpeerni != NULL) {
1311                 msg->msg_rxpeer = NULL;
1312                 lnet_peer_ni_decref_locked(rxpeerni);
1313         }
1314 }
1315
1316 static int
1317 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1318 {
1319         if (p1->lpni_txqnob < p2->lpni_txqnob)
1320                 return 1;
1321
1322         if (p1->lpni_txqnob > p2->lpni_txqnob)
1323                 return -1;
1324
1325         if (p1->lpni_txcredits > p2->lpni_txcredits)
1326                 return 1;
1327
1328         if (p1->lpni_txcredits < p2->lpni_txcredits)
1329                 return -1;
1330
1331         return 0;
1332 }
1333
1334 static struct lnet_peer_ni *
1335 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1336                     struct lnet_peer *peer,
1337                     struct lnet_peer_net *peer_net)
1338 {
1339         /*
1340          * Look at the peer NIs for the destination peer that connect
1341          * to the chosen net. If a peer_ni is preferred when using the
1342          * best_ni to communicate, we use that one. If there is no
1343          * preferred peer_ni, or there are multiple preferred peer_ni,
1344          * the available transmit credits are used. If the transmit
1345          * credits are equal, we round-robin over the peer_ni.
1346          */
1347         struct lnet_peer_ni *lpni = NULL;
1348         struct lnet_peer_ni *best_lpni = NULL;
1349         int best_lpni_credits = INT_MIN;
1350         bool preferred = false;
1351         bool ni_is_pref;
1352         int best_lpni_healthv = 0;
1353         int lpni_healthv;
1354
1355         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1356                 /*
1357                  * if the best_ni we've chosen aleady has this lpni
1358                  * preferred, then let's use it
1359                  */
1360                 if (best_ni) {
1361                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1362                                                                 best_ni->ni_nid);
1363                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1364                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1365                 } else {
1366                         ni_is_pref = false;
1367                 }
1368
1369                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1370
1371                 if (best_lpni)
1372                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1373                                 libcfs_nid2str(lpni->lpni_nid),
1374                                 lpni->lpni_txcredits, best_lpni_credits,
1375                                 lpni->lpni_seq, best_lpni->lpni_seq);
1376
1377                 /* pick the healthiest peer ni */
1378                 if (lpni_healthv < best_lpni_healthv) {
1379                         continue;
1380                 } else if (lpni_healthv > best_lpni_healthv) {
1381                         best_lpni_healthv = lpni_healthv;
1382                 /* if this is a preferred peer use it */
1383                 } else if (!preferred && ni_is_pref) {
1384                         preferred = true;
1385                 } else if (preferred && !ni_is_pref) {
1386                         /*
1387                          * this is not the preferred peer so let's ignore
1388                          * it.
1389                          */
1390                         continue;
1391                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1392                         /*
1393                          * We already have a peer that has more credits
1394                          * available than this one. No need to consider
1395                          * this peer further.
1396                          */
1397                         continue;
1398                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1399                         /*
1400                          * The best peer found so far and the current peer
1401                          * have the same number of available credits let's
1402                          * make sure to select between them using Round
1403                          * Robin
1404                          */
1405                         if (best_lpni) {
1406                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1407                                         continue;
1408                         }
1409                 }
1410
1411                 best_lpni = lpni;
1412                 best_lpni_credits = lpni->lpni_txcredits;
1413         }
1414
1415         /* if we still can't find a peer ni then we can't reach it */
1416         if (!best_lpni) {
1417                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1418                         LNET_NIDNET(dst_nid);
1419                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1420                                 libcfs_net2str(net_id));
1421                 return NULL;
1422         }
1423
1424         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1425                libcfs_nid2str(best_lpni->lpni_nid));
1426
1427         return best_lpni;
1428 }
1429
1430 /*
1431  * Prerequisite: the best_ni should already be set in the sd
1432  */
1433 static inline struct lnet_peer_ni *
1434 lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
1435                            struct lnet_peer *peer, __u32 net_id)
1436 {
1437         struct lnet_peer_net *peer_net;
1438
1439         /*
1440          * The gateway is Multi-Rail capable so now we must select the
1441          * proper peer_ni
1442          */
1443         peer_net = lnet_peer_get_net_locked(peer, net_id);
1444
1445         return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
1446 }
1447
1448 /* Compare route priorities and hop counts */
1449 static int
1450 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1451 {
1452         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1453         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1454
1455         if (r1->lr_priority < r2->lr_priority)
1456                 return 1;
1457
1458         if (r1->lr_priority > r2->lr_priority)
1459                 return -1;
1460
1461         if (r1_hops < r2_hops)
1462                 return 1;
1463
1464         if (r1_hops > r2_hops)
1465                 return -1;
1466
1467         return 0;
1468 }
1469
1470 static struct lnet_route *
1471 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1472                        struct lnet_route **prev_route,
1473                        struct lnet_peer_ni **gwni)
1474 {
1475         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1476         struct lnet_route *best_route;
1477         struct lnet_route *last_route;
1478         struct lnet_route *route;
1479         int rc;
1480
1481         best_route = last_route = NULL;
1482         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1483                 if (!lnet_is_route_alive(route))
1484                         continue;
1485
1486                 /*
1487                  * Restrict the selection of the router NI on the src_net
1488                  * provided. If the src_net is LNET_NID_ANY, then select
1489                  * the best interface available.
1490                  */
1491                 if (!best_route) {
1492                         lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1493                                                           route->lr_gateway,
1494                                                           src_net);
1495                         if (lpni) {
1496                                 best_route = last_route = route;
1497                                 best_gw_ni = lpni;
1498                         } else
1499                                 CERROR("Gateway %s does not have a peer NI on net %s\n",
1500                                        libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1501                                        libcfs_net2str(src_net));
1502
1503                         continue;
1504                 }
1505
1506                 /* no protection on below fields, but it's harmless */
1507                 if (last_route->lr_seq - route->lr_seq < 0)
1508                         last_route = route;
1509
1510                 rc = lnet_compare_routes(route, best_route);
1511                 if (rc == -1)
1512                         continue;
1513
1514                 lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
1515                                                   route->lr_gateway,
1516                                                   src_net);
1517                 if (!lpni) {
1518                         CERROR("Gateway %s does not have a peer NI on net %s\n",
1519                                libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1520                                libcfs_net2str(src_net));
1521                         continue;
1522                 }
1523
1524                 if (rc == 1) {
1525                         best_route = route;
1526                         best_gw_ni = lpni;
1527                         continue;
1528                 }
1529
1530                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1531                 if (rc == -1)
1532                         continue;
1533
1534                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1535                         best_route = route;
1536                         best_gw_ni = lpni;
1537                         continue;
1538                 }
1539         }
1540
1541         *prev_route = last_route;
1542         *gwni = best_gw_ni;
1543
1544         return best_route;
1545 }
1546
1547 static struct lnet_ni *
1548 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1549                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1550                  int md_cpt)
1551 {
1552         struct lnet_ni *ni = NULL;
1553         unsigned int shortest_distance;
1554         int best_credits;
1555         int best_healthv;
1556
1557         /*
1558          * If there is no peer_ni that we can send to on this network,
1559          * then there is no point in looking for a new best_ni here.
1560         */
1561         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1562                 return best_ni;
1563
1564         if (best_ni == NULL) {
1565                 shortest_distance = UINT_MAX;
1566                 best_credits = INT_MIN;
1567                 best_healthv = 0;
1568         } else {
1569                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1570                                                      best_ni->ni_dev_cpt);
1571                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1572                 best_healthv = atomic_read(&best_ni->ni_healthv);
1573         }
1574
1575         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1576                 unsigned int distance;
1577                 int ni_credits;
1578                 int ni_healthv;
1579                 int ni_fatal;
1580
1581                 ni_credits = atomic_read(&ni->ni_tx_credits);
1582                 ni_healthv = atomic_read(&ni->ni_healthv);
1583                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1584
1585                 /*
1586                  * calculate the distance from the CPT on which
1587                  * the message memory is allocated to the CPT of
1588                  * the NI's physical device
1589                  */
1590                 distance = cfs_cpt_distance(lnet_cpt_table(),
1591                                             md_cpt,
1592                                             ni->ni_dev_cpt);
1593
1594                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1595                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1596                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1597                         : "not seleced", best_credits, shortest_distance,
1598                         (best_ni) ? best_ni->ni_seq : 0);
1599
1600                 /*
1601                  * All distances smaller than the NUMA range
1602                  * are treated equally.
1603                  */
1604                 if (distance < lnet_numa_range)
1605                         distance = lnet_numa_range;
1606
1607                 /*
1608                  * Select on health, shorter distance, available
1609                  * credits, then round-robin.
1610                  */
1611                 if (ni_fatal) {
1612                         continue;
1613                 } else if (ni_healthv < best_healthv) {
1614                         continue;
1615                 } else if (ni_healthv > best_healthv) {
1616                         best_healthv = ni_healthv;
1617                         /*
1618                          * If we're going to prefer this ni because it's
1619                          * the healthiest, then we should set the
1620                          * shortest_distance in the algorithm in case
1621                          * there are multiple NIs with the same health but
1622                          * different distances.
1623                          */
1624                         if (distance < shortest_distance)
1625                                 shortest_distance = distance;
1626                 } else if (distance > shortest_distance) {
1627                         continue;
1628                 } else if (distance < shortest_distance) {
1629                         shortest_distance = distance;
1630                 } else if (ni_credits < best_credits) {
1631                         continue;
1632                 } else if (ni_credits == best_credits) {
1633                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1634                                 continue;
1635                 }
1636                 best_ni = ni;
1637                 best_credits = ni_credits;
1638         }
1639
1640         CDEBUG(D_NET, "selected best_ni %s\n",
1641                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1642
1643         return best_ni;
1644 }
1645
1646 /*
1647  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1648  * because such traffic is required to perform discovery. We therefore
1649  * exclude all GET and PUT on that portal. We also exclude all ACK and
1650  * REPLY traffic, but that is because the portal is not tracked in the
1651  * message structure for these message types. We could restrict this
1652  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1653  */
1654 static bool
1655 lnet_msg_discovery(struct lnet_msg *msg)
1656 {
1657         if (msg->msg_type == LNET_MSG_PUT) {
1658                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1659                         return true;
1660         } else if (msg->msg_type == LNET_MSG_GET) {
1661                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1662                         return true;
1663         }
1664         return false;
1665 }
1666
1667 #define SRC_SPEC        0x0001
1668 #define SRC_ANY         0x0002
1669 #define LOCAL_DST       0x0004
1670 #define REMOTE_DST      0x0008
1671 #define MR_DST          0x0010
1672 #define NMR_DST         0x0020
1673 #define SND_RESP        0x0040
1674
1675 /* The following to defines are used for return codes */
1676 #define REPEAT_SEND     0x1000
1677 #define PASS_THROUGH    0x2000
1678
1679 /* The different cases lnet_select pathway needs to handle */
1680 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1681 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1682 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1683 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1684 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1685 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1686 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1687 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1688
1689 static int
1690 lnet_handle_lo_send(struct lnet_send_data *sd)
1691 {
1692         struct lnet_msg *msg = sd->sd_msg;
1693         int cpt = sd->sd_cpt;
1694
1695         /* No send credit hassles with LOLND */
1696         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1697         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1698         if (!msg->msg_routing)
1699                 msg->msg_hdr.src_nid =
1700                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1701         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1702         lnet_msg_commit(msg, cpt);
1703         msg->msg_txni = the_lnet.ln_loni;
1704
1705         return LNET_CREDIT_OK;
1706 }
1707
1708 static int
1709 lnet_handle_send(struct lnet_send_data *sd)
1710 {
1711         struct lnet_ni *best_ni = sd->sd_best_ni;
1712         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1713         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1714         struct lnet_msg *msg = sd->sd_msg;
1715         int cpt2;
1716         __u32 send_case = sd->sd_send_case;
1717         int rc;
1718         __u32 routing = send_case & REMOTE_DST;
1719          struct lnet_rsp_tracker *rspt;
1720
1721         /*
1722          * Increment sequence number of the selected peer so that we
1723          * pick the next one in Round Robin.
1724          */
1725         best_lpni->lpni_seq++;
1726
1727         /*
1728          * grab a reference on the peer_ni so it sticks around even if
1729          * we need to drop and relock the lnet_net_lock below.
1730          */
1731         lnet_peer_ni_addref_locked(best_lpni);
1732
1733         /*
1734          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1735          * message. This ensures that we get a CPT that is correct for
1736          * the NI when the NI has been restricted to a subset of all CPTs.
1737          * If the selected CPT differs from the one currently locked, we
1738          * must unlock and relock the lnet_net_lock(), and then check whether
1739          * the configuration has changed. We don't have a hold on the best_ni
1740          * yet, and it may have vanished.
1741          */
1742         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1743         if (sd->sd_cpt != cpt2) {
1744                 __u32 seq = lnet_get_dlc_seq_locked();
1745                 lnet_net_unlock(sd->sd_cpt);
1746                 sd->sd_cpt = cpt2;
1747                 lnet_net_lock(sd->sd_cpt);
1748                 if (seq != lnet_get_dlc_seq_locked()) {
1749                         lnet_peer_ni_decref_locked(best_lpni);
1750                         return REPEAT_SEND;
1751                 }
1752         }
1753
1754         /*
1755          * store the best_lpni in the message right away to avoid having
1756          * to do the same operation under different conditions
1757          */
1758         msg->msg_txpeer = best_lpni;
1759         msg->msg_txni = best_ni;
1760
1761         /*
1762          * grab a reference for the best_ni since now it's in use in this
1763          * send. The reference will be dropped in lnet_finalize()
1764          */
1765         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1766
1767         /*
1768          * Always set the target.nid to the best peer picked. Either the
1769          * NID will be one of the peer NIDs selected, or the same NID as
1770          * what was originally set in the target or it will be the NID of
1771          * a router if this message should be routed
1772          */
1773         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1774
1775         /*
1776          * lnet_msg_commit assigns the correct cpt to the message, which
1777          * is used to decrement the correct refcount on the ni when it's
1778          * time to return the credits
1779          */
1780         lnet_msg_commit(msg, sd->sd_cpt);
1781
1782         /*
1783          * If we are routing the message then we keep the src_nid that was
1784          * set by the originator. If we are not routing then we are the
1785          * originator and set it here.
1786          */
1787         if (!msg->msg_routing)
1788                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1789
1790         if (routing) {
1791                 msg->msg_target_is_router = 1;
1792                 msg->msg_target.pid = LNET_PID_LUSTRE;
1793                 /*
1794                  * since we're routing we want to ensure that the
1795                  * msg_hdr.dest_nid is set to the final destination. When
1796                  * the router receives this message it knows how to route
1797                  * it.
1798                  *
1799                  * final_dst_lpni is set at the beginning of the
1800                  * lnet_select_pathway() function and is never changed.
1801                  * It's safe to use it here.
1802                  */
1803                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1804         } else {
1805                 /*
1806                  * if we're not routing set the dest_nid to the best peer
1807                  * ni NID that we picked earlier in the algorithm.
1808                  */
1809                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1810         }
1811
1812         /*
1813          * if we have response tracker block update it with the next hop
1814          * nid
1815          */
1816         if (msg->msg_md) {
1817                 rspt = msg->msg_md->md_rspt_ptr;
1818                 if (rspt) {
1819                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1820                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1821                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1822                 }
1823         }
1824
1825         rc = lnet_post_send_locked(msg, 0);
1826
1827         if (!rc)
1828                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1829                        libcfs_nid2str(msg->msg_hdr.src_nid),
1830                        libcfs_nid2str(msg->msg_txni->ni_nid),
1831                        libcfs_nid2str(sd->sd_src_nid),
1832                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1833                        libcfs_nid2str(sd->sd_dst_nid),
1834                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1835                        libcfs_nid2str(sd->sd_rtr_nid),
1836                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1837
1838         return rc;
1839 }
1840
1841 static inline void
1842 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1843                          struct lnet_msg *msg)
1844 {
1845         if (!lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1846                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1847                        libcfs_nid2str(lni->ni_nid),
1848                        libcfs_nid2str(lpni->lpni_nid));
1849                 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1850         }
1851 }
1852
1853 /*
1854  * Source Specified
1855  * Local Destination
1856  * non-mr peer
1857  *
1858  * use the source and destination NIDs as the pathway
1859  */
1860 static int
1861 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1862 {
1863         /* the destination lpni is set before we get here. */
1864
1865         /* find local NI */
1866         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1867         if (!sd->sd_best_ni) {
1868                 CERROR("Can't send to %s: src %s is not a "
1869                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1870                                 libcfs_nid2str(sd->sd_src_nid));
1871                 return -EINVAL;
1872         }
1873
1874         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1875
1876         return lnet_handle_send(sd);
1877 }
1878
1879 /*
1880  * Source Specified
1881  * Local Destination
1882  * MR Peer
1883  *
1884  * Don't run the selection algorithm on the peer NIs. By specifying the
1885  * local NID, we're also saying that we should always use the destination NID
1886  * provided. This handles the case where we should be using the same
1887  * destination NID for the all the messages which belong to the same RPC
1888  * request.
1889  */
1890 static int
1891 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1892 {
1893         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1894         if (!sd->sd_best_ni) {
1895                 CERROR("Can't send to %s: src %s is not a "
1896                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1897                                 libcfs_nid2str(sd->sd_src_nid));
1898                 return -EINVAL;
1899         }
1900
1901         if (sd->sd_best_lpni &&
1902             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1903                 return lnet_handle_lo_send(sd);
1904         else if (sd->sd_best_lpni)
1905                 return lnet_handle_send(sd);
1906
1907         CERROR("can't send to %s. no NI on %s\n",
1908                libcfs_nid2str(sd->sd_dst_nid),
1909                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1910
1911         return -EHOSTUNREACH;
1912 }
1913
1914 struct lnet_ni *
1915 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1916                               struct lnet_peer *peer,
1917                               struct lnet_peer_net *peer_net,
1918                               int cpt,
1919                               bool incr_seq)
1920 {
1921         struct lnet_net *local_net;
1922         struct lnet_ni *best_ni;
1923
1924         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1925         if (!local_net)
1926                 return NULL;
1927
1928         /*
1929          * Iterate through the NIs in this local Net and select
1930          * the NI to send from. The selection is determined by
1931          * these 3 criterion in the following priority:
1932          *      1. NUMA
1933          *      2. NI available credits
1934          *      3. Round Robin
1935          */
1936         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1937                                    peer, peer_net, cpt);
1938
1939         if (incr_seq && best_ni)
1940                 best_ni->ni_seq++;
1941
1942         return best_ni;
1943 }
1944
1945 static int
1946 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
1947                              int cpt)
1948 {
1949         struct lnet_peer *peer;
1950         int rc;
1951
1952         lnet_peer_ni_addref_locked(lpni);
1953
1954         peer = lpni->lpni_peer_net->lpn_peer;
1955
1956         if (lnet_peer_gw_discovery(peer)) {
1957                 lnet_peer_ni_decref_locked(lpni);
1958                 return 0;
1959         }
1960
1961         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1962                 lnet_peer_ni_decref_locked(lpni);
1963                 return 0;
1964         }
1965
1966         rc = lnet_discover_peer_locked(lpni, cpt, false);
1967         if (rc) {
1968                 lnet_peer_ni_decref_locked(lpni);
1969                 return rc;
1970         }
1971         /* The peer may have changed. */
1972         peer = lpni->lpni_peer_net->lpn_peer;
1973         spin_lock(&peer->lp_lock);
1974         if (lnet_peer_is_uptodate_locked(peer)) {
1975                 spin_unlock(&peer->lp_lock);
1976                 lnet_peer_ni_decref_locked(lpni);
1977                 return 0;
1978         }
1979         /* queue message and return */
1980         msg->msg_sending = 0;
1981         msg->msg_txpeer = NULL;
1982         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1983         spin_unlock(&peer->lp_lock);
1984
1985         lnet_peer_ni_decref_locked(lpni);
1986
1987         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1988                msg, libcfs_nid2str(peer->lp_primary_nid));
1989
1990         return LNET_DC_WAIT;
1991 }
1992
1993 static int
1994 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1995                              lnet_nid_t dst_nid,
1996                              struct lnet_peer_ni **gw_lpni,
1997                              struct lnet_peer **gw_peer)
1998 {
1999         int rc;
2000         __u32 local_lnet;
2001         struct lnet_peer *gw;
2002         struct lnet_peer *lp;
2003         struct lnet_peer_net *lpn;
2004         struct lnet_peer_net *best_lpn = NULL;
2005         struct lnet_remotenet *rnet, *best_rnet = NULL;
2006         struct lnet_route *best_route = NULL;
2007         struct lnet_route *last_route = NULL;
2008         struct lnet_peer_ni *lpni = NULL;
2009         struct lnet_peer_ni *gwni = NULL;
2010         lnet_nid_t src_nid = sd->sd_src_nid;
2011
2012         /* If a router nid was specified then we are replying to a GET or
2013          * sending an ACK. In this case we use the gateway associated with the
2014          * specified router nid.
2015          */
2016         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2017                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2018                 if (!gwni) {
2019                         CERROR("No peer NI for gateway %s\n",
2020                                libcfs_nid2str(sd->sd_rtr_nid));
2021                         return -EHOSTUNREACH;
2022                 }
2023                 gw = gwni->lpni_peer_net->lpn_peer;
2024                 lnet_peer_ni_decref_locked(gwni);
2025                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2026         } else {
2027                 /* we've already looked up the initial lpni using dst_nid */
2028                 lpni = sd->sd_best_lpni;
2029                 /* the peer tree must be in existence */
2030                 LASSERT(lpni && lpni->lpni_peer_net &&
2031                         lpni->lpni_peer_net->lpn_peer);
2032                 lp = lpni->lpni_peer_net->lpn_peer;
2033
2034                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2035                         /* is this remote network reachable?  */
2036                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2037                         if (!rnet)
2038                                 continue;
2039
2040                         if (!best_lpn) {
2041                                 best_lpn = lpn;
2042                                 best_rnet = rnet;
2043                         }
2044
2045                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2046                                 continue;
2047
2048                         best_lpn = lpn;
2049                         best_rnet = rnet;
2050                 }
2051
2052                 if (!best_lpn) {
2053                         CERROR("peer %s has no available nets\n",
2054                                libcfs_nid2str(sd->sd_dst_nid));
2055                         return -EHOSTUNREACH;
2056                 }
2057
2058                 sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
2059                                                               sd->sd_dst_nid,
2060                                                               lp,
2061                                                               best_lpn->lpn_net_id);
2062                 if (!sd->sd_best_lpni) {
2063                         CERROR("peer %s down\n",
2064                                libcfs_nid2str(sd->sd_dst_nid));
2065                         return -EHOSTUNREACH;
2066                 }
2067
2068                 best_route = lnet_find_route_locked(best_rnet,
2069                                                     LNET_NIDNET(src_nid),
2070                                                     &last_route, &gwni);
2071                 if (!best_route) {
2072                         CERROR("no route to %s from %s\n",
2073                                libcfs_nid2str(dst_nid),
2074                                libcfs_nid2str(src_nid));
2075                         return -EHOSTUNREACH;
2076                 }
2077
2078                 if (!gwni) {
2079                         CERROR("Internal Error. Route expected to %s from %s\n",
2080                                libcfs_nid2str(dst_nid),
2081                                libcfs_nid2str(src_nid));
2082                         return -EFAULT;
2083                 }
2084
2085                 gw = best_route->lr_gateway;
2086                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2087                 local_lnet = best_route->lr_lnet;
2088
2089         }
2090
2091         /*
2092          * Discover this gateway if it hasn't already been discovered.
2093          * This means we might delay the message until discovery has
2094          * completed
2095          */
2096         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2097         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2098         if (rc)
2099                 return rc;
2100
2101         if (!sd->sd_best_ni)
2102                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2103                                         lnet_peer_get_net_locked(gw,
2104                                                                  local_lnet),
2105                                         sd->sd_md_cpt,
2106                                         true);
2107
2108         if (!sd->sd_best_ni) {
2109                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2110                        libcfs_net2str(local_lnet),
2111                        libcfs_nid2str(sd->sd_src_nid));
2112                 return -EFAULT;
2113         }
2114
2115         *gw_lpni = gwni;
2116         *gw_peer = gw;
2117
2118         /*
2119          * increment the sequence numbers since now we're sure we're
2120          * going to use this path
2121          */
2122         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2123                 LASSERT(best_route && last_route);
2124                 best_route->lr_seq = last_route->lr_seq + 1;
2125                 best_lpn->lpn_seq++;
2126         }
2127
2128         return 0;
2129 }
2130
2131 /*
2132  * Handle two cases:
2133  *
2134  * Case 1:
2135  *  Source specified
2136  *  Remote destination
2137  *  Non-MR destination
2138  *
2139  * Case 2:
2140  *  Source specified
2141  *  Remote destination
2142  *  MR destination
2143  *
2144  * The handling of these two cases is similar. Even though the destination
2145  * can be MR or non-MR, we'll deal directly with the router.
2146  */
2147 static int
2148 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2149 {
2150         int rc;
2151         struct lnet_peer_ni *gw_lpni = NULL;
2152         struct lnet_peer *gw_peer = NULL;
2153
2154         /* find local NI */
2155         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2156         if (!sd->sd_best_ni) {
2157                 CERROR("Can't send to %s: src %s is not a "
2158                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2159                                 libcfs_nid2str(sd->sd_src_nid));
2160                 return -EINVAL;
2161         }
2162
2163         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2164                                      &gw_peer);
2165         if (rc)
2166                 return rc;
2167
2168         if (sd->sd_send_case & NMR_DST)
2169                 /*
2170                  * since the final destination is non-MR let's set its preferred
2171                  * NID before we send
2172                  */
2173                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2174                                          sd->sd_msg);
2175
2176         /*
2177          * We're going to send to the gw found so let's set its
2178          * info
2179          */
2180         sd->sd_peer = gw_peer;
2181         sd->sd_best_lpni = gw_lpni;
2182
2183         return lnet_handle_send(sd);
2184 }
2185
2186 struct lnet_ni *
2187 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2188                                bool discovery)
2189 {
2190         struct lnet_peer_net *peer_net = NULL;
2191         struct lnet_ni *best_ni = NULL;
2192         int lpn_healthv = 0;
2193
2194         /*
2195          * The peer can have multiple interfaces, some of them can be on
2196          * the local network and others on a routed network. We should
2197          * prefer the local network. However if the local network is not
2198          * available then we need to try the routed network
2199          */
2200
2201         /* go through all the peer nets and find the best_ni */
2202         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2203                 /*
2204                  * The peer's list of nets can contain non-local nets. We
2205                  * want to only examine the local ones.
2206                  */
2207                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2208                         continue;
2209
2210                 /* always select the lpn with the best health */
2211                 if (lpn_healthv <= peer_net->lpn_healthv)
2212                         lpn_healthv = peer_net->lpn_healthv;
2213                 else
2214                         continue;
2215
2216                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2217                                                         md_cpt, false);
2218
2219                 /*
2220                  * if this is a discovery message and lp_disc_net_id is
2221                  * specified then use that net to send the discovery on.
2222                  */
2223                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2224                     discovery)
2225                         break;
2226         }
2227
2228         if (best_ni)
2229                 /* increment sequence number so we can round robin */
2230                 best_ni->ni_seq++;
2231
2232         return best_ni;
2233 }
2234
2235 static struct lnet_ni *
2236 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2237 {
2238         struct lnet_ni *best_ni = NULL;
2239         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2240         struct lnet_peer_ni *lpni_entry;
2241
2242         /*
2243          * We must use a consistent source address when sending to a
2244          * non-MR peer. However, a non-MR peer can have multiple NIDs
2245          * on multiple networks, and we may even need to talk to this
2246          * peer on multiple networks -- certain types of
2247          * load-balancing configuration do this.
2248          *
2249          * So we need to pick the NI the peer prefers for this
2250          * particular network.
2251          */
2252         LASSERT(peer_net);
2253         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2254                             lpni_peer_nis) {
2255                 if (lpni_entry->lpni_pref_nnids == 0)
2256                         continue;
2257                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2258                 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2259                 break;
2260         }
2261
2262         return best_ni;
2263 }
2264
2265 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2266 static int
2267 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2268 {
2269         struct lnet_ni *best_ni = NULL;
2270         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2271
2272         /*
2273          * We must use a consistent source address when sending to a
2274          * non-MR peer. However, a non-MR peer can have multiple NIDs
2275          * on multiple networks, and we may even need to talk to this
2276          * peer on multiple networks -- certain types of
2277          * load-balancing configuration do this.
2278          *
2279          * So we need to pick the NI the peer prefers for this
2280          * particular network.
2281          */
2282
2283         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2284                                                        sd->sd_cpt);
2285
2286         /* if best_ni is still not set just pick one */
2287         if (!best_ni) {
2288                 best_ni =
2289                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2290                                                 sd->sd_best_lpni->lpni_peer_net,
2291                                                 sd->sd_md_cpt, true);
2292                 /* If there is no best_ni we don't have a route */
2293                 if (!best_ni) {
2294                         CERROR("no path to %s from net %s\n",
2295                                 libcfs_nid2str(best_lpni->lpni_nid),
2296                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2297                         return -EHOSTUNREACH;
2298                 }
2299         }
2300
2301         sd->sd_best_ni = best_ni;
2302
2303         /* Set preferred NI if necessary. */
2304         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2305
2306         return 0;
2307 }
2308
2309
2310 /*
2311  * Source not specified
2312  * Local destination
2313  * Non-MR Peer
2314  *
2315  * always use the same source NID for NMR peers
2316  * If we've talked to that peer before then we already have a preferred
2317  * source NI associated with it. Otherwise, we select a preferred local NI
2318  * and store it in the peer
2319  */
2320 static int
2321 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2322 {
2323         int rc;
2324
2325         /* sd->sd_best_lpni is already set to the final destination */
2326
2327         /*
2328          * At this point we should've created the peer ni and peer. If we
2329          * can't find it, then something went wrong. Instead of assert
2330          * output a relevant message and fail the send
2331          */
2332         if (!sd->sd_best_lpni) {
2333                 CERROR("Internal fault. Unable to send msg %s to %s. "
2334                        "NID not known\n",
2335                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2336                        libcfs_nid2str(sd->sd_dst_nid));
2337                 return -EFAULT;
2338         }
2339
2340         rc = lnet_select_preferred_best_ni(sd);
2341         if (!rc)
2342                 rc = lnet_handle_send(sd);
2343
2344         return rc;
2345 }
2346
2347 static int
2348 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2349 {
2350         /*
2351          * NOTE we've already handled the remote peer case. So we only
2352          * need to worry about the local case here.
2353          *
2354          * if we're sending a response, ACK or reply, we need to send it
2355          * to the destination NID given to us. At this point we already
2356          * have the peer_ni we're suppose to send to, so just find the
2357          * best_ni on the peer net and use that. Since we're sending to an
2358          * MR peer then we can just run the selection algorithm on our
2359          * local NIs and pick the best one.
2360          */
2361         if (sd->sd_send_case & SND_RESP) {
2362                 sd->sd_best_ni =
2363                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2364                                                 sd->sd_best_lpni->lpni_peer_net,
2365                                                 sd->sd_md_cpt, true);
2366
2367                 if (!sd->sd_best_ni) {
2368                         /*
2369                          * We're not going to deal with not able to send
2370                          * a response to the provided final destination
2371                          */
2372                         CERROR("Can't send response to %s. "
2373                                "No local NI available\n",
2374                                 libcfs_nid2str(sd->sd_dst_nid));
2375                         return -EHOSTUNREACH;
2376                 }
2377
2378                 return lnet_handle_send(sd);
2379         }
2380
2381         /*
2382          * If we get here that means we're sending a fresh request, PUT or
2383          * GET, so we need to run our standard selection algorithm.
2384          * First find the best local interface that's on any of the peer's
2385          * networks.
2386          */
2387         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2388                                         sd->sd_md_cpt,
2389                                         lnet_msg_discovery(sd->sd_msg));
2390         if (sd->sd_best_ni) {
2391                 sd->sd_best_lpni =
2392                   lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
2393                                              sd->sd_peer,
2394                                              sd->sd_best_ni->ni_net->net_id);
2395
2396                 /*
2397                  * if we're successful in selecting a peer_ni on the local
2398                  * network, then send to it. Otherwise fall through and
2399                  * try and see if we can reach it over another routed
2400                  * network
2401                  */
2402                 if (sd->sd_best_lpni &&
2403                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2404                         /*
2405                          * in case we initially started with a routed
2406                          * destination, let's reset to local
2407                          */
2408                         sd->sd_send_case &= ~REMOTE_DST;
2409                         sd->sd_send_case |= LOCAL_DST;
2410                         return lnet_handle_lo_send(sd);
2411                 } else if (sd->sd_best_lpni) {
2412                         /*
2413                          * in case we initially started with a routed
2414                          * destination, let's reset to local
2415                          */
2416                         sd->sd_send_case &= ~REMOTE_DST;
2417                         sd->sd_send_case |= LOCAL_DST;
2418                         return lnet_handle_send(sd);
2419                 }
2420
2421                 CERROR("Internal Error. Expected to have a best_lpni: "
2422                        "%s -> %s\n",
2423                        libcfs_nid2str(sd->sd_src_nid),
2424                        libcfs_nid2str(sd->sd_dst_nid));
2425
2426                 return -EFAULT;
2427         }
2428
2429         /*
2430          * Peer doesn't have a local network. Let's see if there is
2431          * a remote network we can reach it on.
2432          */
2433         return PASS_THROUGH;
2434 }
2435
2436 /*
2437  * Case 1:
2438  *      Source NID not specified
2439  *      Local destination
2440  *      MR peer
2441  *
2442  * Case 2:
2443  *      Source NID not speified
2444  *      Remote destination
2445  *      MR peer
2446  *
2447  * In both of these cases if we're sending a response, ACK or REPLY, then
2448  * we need to send to the destination NID provided.
2449  *
2450  * In the remote case let's deal with MR routers.
2451  *
2452  */
2453
2454 static int
2455 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2456 {
2457         int rc = 0;
2458         struct lnet_peer *gw_peer = NULL;
2459         struct lnet_peer_ni *gw_lpni = NULL;
2460
2461         /*
2462          * handle sending a response to a remote peer here so we don't
2463          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2464          */
2465         if (sd->sd_send_case & REMOTE_DST &&
2466             sd->sd_send_case & SND_RESP) {
2467                 struct lnet_peer_ni *gw;
2468                 struct lnet_peer *gw_peer;
2469
2470                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2471                                                   &gw_peer);
2472                 if (rc < 0) {
2473                         CERROR("Can't send response to %s. "
2474                                "No route available\n",
2475                                 libcfs_nid2str(sd->sd_dst_nid));
2476                         return -EHOSTUNREACH;
2477                 } else if (rc > 0) {
2478                         return rc;
2479                 }
2480
2481                 sd->sd_best_lpni = gw;
2482                 sd->sd_peer = gw_peer;
2483
2484                 return lnet_handle_send(sd);
2485         }
2486
2487         /*
2488          * Even though the NID for the peer might not be on a local network,
2489          * since the peer is MR there could be other interfaces on the
2490          * local network. In that case we'd still like to prefer the local
2491          * network over the routed network. If we're unable to do that
2492          * then we select the best router among the different routed networks,
2493          * and if the router is MR then we can deal with it as such.
2494          */
2495         rc = lnet_handle_any_mr_dsta(sd);
2496         if (rc != PASS_THROUGH)
2497                 return rc;
2498
2499         /*
2500          * Now that we must route to the destination, we must consider the
2501          * MR case, where the destination has multiple interfaces, some of
2502          * which we can route to and others we do not. For this reason we
2503          * need to select the destination which we can route to and if
2504          * there are multiple, we need to round robin.
2505          */
2506         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2507                                           &gw_peer);
2508         if (rc)
2509                 return rc;
2510
2511         sd->sd_send_case &= ~LOCAL_DST;
2512         sd->sd_send_case |= REMOTE_DST;
2513
2514         sd->sd_peer = gw_peer;
2515         sd->sd_best_lpni = gw_lpni;
2516
2517         return lnet_handle_send(sd);
2518 }
2519
2520 /*
2521  * Source not specified
2522  * Remote destination
2523  * Non-MR peer
2524  *
2525  * Must send to the specified peer NID using the same source NID that
2526  * we've used before. If it's the first time to talk to that peer then
2527  * find the source NI and assign it as preferred to that peer
2528  */
2529 static int
2530 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2531 {
2532         int rc;
2533         struct lnet_peer_ni *gw_lpni = NULL;
2534         struct lnet_peer *gw_peer = NULL;
2535
2536         /*
2537          * Let's see if we have a preferred NI to talk to this NMR peer
2538          */
2539         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2540                                                               sd->sd_cpt);
2541
2542         /*
2543          * find the router and that'll find the best NI if we didn't find
2544          * it already.
2545          */
2546         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2547                                           &gw_peer);
2548         if (rc)
2549                 return rc;
2550
2551         /*
2552          * set the best_ni we've chosen as the preferred one for
2553          * this peer
2554          */
2555         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2556
2557         /* we'll be sending to the gw */
2558         sd->sd_best_lpni = gw_lpni;
2559         sd->sd_peer = gw_peer;
2560
2561         return lnet_handle_send(sd);
2562 }
2563
2564 static int
2565 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2566 {
2567         /*
2568          * turn off the SND_RESP bit.
2569          * It will be checked in the case handling
2570          */
2571         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2572
2573         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2574                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2575                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2576                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2577                 libcfs_nid2str(sd->sd_dst_nid),
2578                 (send_case & LOCAL_DST) ? "local" : "routed");
2579
2580         switch (send_case) {
2581         /*
2582          * For all cases where the source is specified, we should always
2583          * use the destination NID, whether it's an MR destination or not,
2584          * since we're continuing a series of related messages for the
2585          * same RPC
2586          */
2587         case SRC_SPEC_LOCAL_NMR_DST:
2588                 return lnet_handle_spec_local_nmr_dst(sd);
2589         case SRC_SPEC_LOCAL_MR_DST:
2590                 return lnet_handle_spec_local_mr_dst(sd);
2591         case SRC_SPEC_ROUTER_NMR_DST:
2592         case SRC_SPEC_ROUTER_MR_DST:
2593                 return lnet_handle_spec_router_dst(sd);
2594         case SRC_ANY_LOCAL_NMR_DST:
2595                 return lnet_handle_any_local_nmr_dst(sd);
2596         case SRC_ANY_LOCAL_MR_DST:
2597         case SRC_ANY_ROUTER_MR_DST:
2598                 return lnet_handle_any_mr_dst(sd);
2599         case SRC_ANY_ROUTER_NMR_DST:
2600                 return lnet_handle_any_router_nmr_dst(sd);
2601         default:
2602                 CERROR("Unknown send case\n");
2603                 return -1;
2604         }
2605 }
2606
2607 static int
2608 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2609                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2610 {
2611         struct lnet_peer_ni     *lpni;
2612         struct lnet_peer        *peer;
2613         struct lnet_send_data   send_data;
2614         int                     cpt, rc;
2615         int                     md_cpt;
2616         __u32                   send_case = 0;
2617
2618         memset(&send_data, 0, sizeof(send_data));
2619
2620         /*
2621          * get an initial CPT to use for locking. The idea here is not to
2622          * serialize the calls to select_pathway, so that as many
2623          * operations can run concurrently as possible. To do that we use
2624          * the CPT where this call is being executed. Later on when we
2625          * determine the CPT to use in lnet_message_commit, we switch the
2626          * lock and check if there was any configuration change.  If none,
2627          * then we proceed, if there is, then we restart the operation.
2628          */
2629         cpt = lnet_net_lock_current();
2630
2631         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2632         if (md_cpt == CFS_CPT_ANY)
2633                 md_cpt = cpt;
2634
2635 again:
2636
2637         /*
2638          * If we're being asked to send to the loopback interface, there
2639          * is no need to go through any selection. We can just shortcut
2640          * the entire process and send over lolnd
2641          */
2642         send_data.sd_msg = msg;
2643         send_data.sd_cpt = cpt;
2644         if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
2645                 rc = lnet_handle_lo_send(&send_data);
2646                 lnet_net_unlock(cpt);
2647                 return rc;
2648         }
2649
2650         /*
2651          * find an existing peer_ni, or create one and mark it as having been
2652          * created due to network traffic. This call will create the
2653          * peer->peer_net->peer_ni tree.
2654          */
2655         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2656         if (IS_ERR(lpni)) {
2657                 lnet_net_unlock(cpt);
2658                 return PTR_ERR(lpni);
2659         }
2660
2661         /*
2662          * Cache the original src_nid and rtr_nid. If we need to resend the
2663          * message then we'll need to know whether the src_nid was originally
2664          * specified for this message. If it was originally specified,
2665          * then we need to keep using the same src_nid since it's
2666          * continuing the same sequence of messages. Similarly, rtr_nid will
2667          * affect our choice of next hop.
2668          */
2669         msg->msg_src_nid_param = src_nid;
2670         msg->msg_rtr_nid_param = rtr_nid;
2671
2672         /*
2673          * If necessary, perform discovery on the peer that owns this peer_ni.
2674          * Note, this can result in the ownership of this peer_ni changing
2675          * to another peer object.
2676          */
2677         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2678         if (rc) {
2679                 lnet_peer_ni_decref_locked(lpni);
2680                 lnet_net_unlock(cpt);
2681                 return rc;
2682         }
2683         lnet_peer_ni_decref_locked(lpni);
2684
2685         peer = lpni->lpni_peer_net->lpn_peer;
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 (lnet_msg_is_response(msg))
2711                 send_case |= SND_RESP;
2712
2713         /* assign parameters to the send_data */
2714         send_data.sd_rtr_nid = rtr_nid;
2715         send_data.sd_src_nid = src_nid;
2716         send_data.sd_dst_nid = dst_nid;
2717         send_data.sd_best_lpni = lpni;
2718         /*
2719          * keep a pointer to the final destination in case we're going to
2720          * route, so we'll need to access it later
2721          */
2722         send_data.sd_final_dst_lpni = lpni;
2723         send_data.sd_peer = peer;
2724         send_data.sd_md_cpt = md_cpt;
2725         send_data.sd_send_case = send_case;
2726
2727         rc = lnet_handle_send_case_locked(&send_data);
2728
2729         /*
2730          * Update the local cpt since send_data.sd_cpt might've been
2731          * updated as a result of calling lnet_handle_send_case_locked().
2732          */
2733         cpt = send_data.sd_cpt;
2734
2735         if (rc == REPEAT_SEND)
2736                 goto again;
2737
2738         lnet_net_unlock(cpt);
2739
2740         return rc;
2741 }
2742
2743 int
2744 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2745 {
2746         lnet_nid_t              dst_nid = msg->msg_target.nid;
2747         int                     rc;
2748
2749         /*
2750          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2751          * but we might want to use pre-determined router for ACK/REPLY
2752          * in the future
2753          */
2754         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2755         LASSERT(msg->msg_txpeer == NULL);
2756         LASSERT(msg->msg_txni == NULL);
2757         LASSERT(!msg->msg_sending);
2758         LASSERT(!msg->msg_target_is_router);
2759         LASSERT(!msg->msg_receiving);
2760
2761         msg->msg_sending = 1;
2762
2763         LASSERT(!msg->msg_tx_committed);
2764
2765         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2766         if (rc < 0) {
2767                 if (rc == -EHOSTUNREACH)
2768                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2769                 else
2770                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2771                 return rc;
2772         }
2773
2774         if (rc == LNET_CREDIT_OK)
2775                 lnet_ni_send(msg->msg_txni, msg);
2776
2777         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2778         return 0;
2779 }
2780
2781 enum lnet_mt_event_type {
2782         MT_TYPE_LOCAL_NI = 0,
2783         MT_TYPE_PEER_NI
2784 };
2785
2786 struct lnet_mt_event_info {
2787         enum lnet_mt_event_type mt_type;
2788         lnet_nid_t mt_nid;
2789 };
2790
2791 /* called with res_lock held */
2792 void
2793 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2794 {
2795         struct lnet_rsp_tracker *rspt;
2796
2797         /*
2798          * msg has a refcount on the MD so the MD is not going away.
2799          * The rspt queue for the cpt is protected by
2800          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2801          */
2802         if (!md->md_rspt_ptr)
2803                 return;
2804
2805         rspt = md->md_rspt_ptr;
2806
2807         /* debug code */
2808         LASSERT(rspt->rspt_cpt == cpt);
2809
2810         md->md_rspt_ptr = NULL;
2811
2812         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2813                 /*
2814                  * The monitor thread has invalidated this handle because the
2815                  * response timed out, but it failed to lookup the MD. That
2816                  * means this response tracker is on the zombie list. We can
2817                  * safely remove it under the resource lock (held by caller) and
2818                  * free the response tracker block.
2819                  */
2820                 list_del(&rspt->rspt_on_list);
2821                 lnet_rspt_free(rspt, cpt);
2822         } else {
2823                 /*
2824                  * invalidate the handle to indicate that a response has been
2825                  * received, which will then lead the monitor thread to clean up
2826                  * the rspt block.
2827                  */
2828                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2829         }
2830 }
2831
2832 void
2833 lnet_clean_zombie_rstqs(void)
2834 {
2835         struct lnet_rsp_tracker *rspt, *tmp;
2836         int i;
2837
2838         cfs_cpt_for_each(i, lnet_cpt_table()) {
2839                 list_for_each_entry_safe(rspt, tmp,
2840                                          the_lnet.ln_mt_zombie_rstqs[i],
2841                                          rspt_on_list) {
2842                         list_del(&rspt->rspt_on_list);
2843                         lnet_rspt_free(rspt, i);
2844                 }
2845         }
2846
2847         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2848 }
2849
2850 static void
2851 lnet_finalize_expired_responses(void)
2852 {
2853         struct lnet_libmd *md;
2854         struct lnet_rsp_tracker *rspt, *tmp;
2855         ktime_t now;
2856         int i;
2857
2858         if (the_lnet.ln_mt_rstq == NULL)
2859                 return;
2860
2861         cfs_cpt_for_each(i, lnet_cpt_table()) {
2862                 LIST_HEAD(local_queue);
2863
2864                 lnet_net_lock(i);
2865                 if (!the_lnet.ln_mt_rstq[i]) {
2866                         lnet_net_unlock(i);
2867                         continue;
2868                 }
2869                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2870                 lnet_net_unlock(i);
2871
2872                 now = ktime_get();
2873
2874                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2875                         /*
2876                          * The rspt mdh will be invalidated when a response
2877                          * is received or whenever we want to discard the
2878                          * block the monitor thread will walk the queue
2879                          * and clean up any rsts with an invalid mdh.
2880                          * The monitor thread will walk the queue until
2881                          * the first unexpired rspt block. This means that
2882                          * some rspt blocks which received their
2883                          * corresponding responses will linger in the
2884                          * queue until they are cleaned up eventually.
2885                          */
2886                         lnet_res_lock(i);
2887                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2888                                 lnet_res_unlock(i);
2889                                 list_del(&rspt->rspt_on_list);
2890                                 lnet_rspt_free(rspt, i);
2891                                 continue;
2892                         }
2893
2894                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2895                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2896                                 struct lnet_peer_ni *lpni;
2897                                 lnet_nid_t nid;
2898
2899                                 md = lnet_handle2md(&rspt->rspt_mdh);
2900                                 if (!md) {
2901                                         /* MD has been queued for unlink, but
2902                                          * rspt hasn't been detached (Note we've
2903                                          * checked above that the rspt_mdh is
2904                                          * valid). Since we cannot lookup the MD
2905                                          * we're unable to detach the rspt
2906                                          * ourselves. Thus, move the rspt to the
2907                                          * zombie list where we'll wait for
2908                                          * either:
2909                                          *   1. The remaining operations on the
2910                                          *   MD to complete. In this case the
2911                                          *   final operation will result in
2912                                          *   lnet_msg_detach_md()->
2913                                          *   lnet_detach_rsp_tracker() where
2914                                          *   we will clean up this response
2915                                          *   tracker.
2916                                          *   2. LNet to shutdown. In this case
2917                                          *   we'll wait until after all LND Nets
2918                                          *   have shutdown and then we can
2919                                          *   safely free any remaining response
2920                                          *   tracker blocks on the zombie list.
2921                                          * Note: We need to hold the resource
2922                                          * lock when adding to the zombie list
2923                                          * because we may have concurrent access
2924                                          * with lnet_detach_rsp_tracker().
2925                                          */
2926                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2927                                         list_move(&rspt->rspt_on_list,
2928                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2929                                         lnet_res_unlock(i);
2930                                         continue;
2931                                 }
2932                                 LASSERT(md->md_rspt_ptr == rspt);
2933                                 md->md_rspt_ptr = NULL;
2934                                 lnet_res_unlock(i);
2935
2936                                 LNetMDUnlink(rspt->rspt_mdh);
2937
2938                                 nid = rspt->rspt_next_hop_nid;
2939
2940                                 list_del(&rspt->rspt_on_list);
2941                                 lnet_rspt_free(rspt, i);
2942
2943                                 /* If we're shutting down we just want to clean
2944                                  * up the rspt blocks
2945                                  */
2946                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2947                                         continue;
2948
2949                                 lnet_net_lock(i);
2950                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2951                                 lnet_net_unlock(i);
2952
2953                                 CDEBUG(D_NET,
2954                                        "Response timeout: md = %p: nid = %s\n",
2955                                        md, libcfs_nid2str(nid));
2956
2957                                 /*
2958                                  * If there is a timeout on the response
2959                                  * from the next hop decrement its health
2960                                  * value so that we don't use it
2961                                  */
2962                                 lnet_net_lock(0);
2963                                 lpni = lnet_find_peer_ni_locked(nid);
2964                                 if (lpni) {
2965                                         lnet_handle_remote_failure_locked(lpni);
2966                                         lnet_peer_ni_decref_locked(lpni);
2967                                 }
2968                                 lnet_net_unlock(0);
2969                         } else {
2970                                 lnet_res_unlock(i);
2971                                 break;
2972                         }
2973                 }
2974
2975                 if (!list_empty(&local_queue)) {
2976                         lnet_net_lock(i);
2977                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
2978                         lnet_net_unlock(i);
2979                 }
2980         }
2981 }
2982
2983 static void
2984 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
2985 {
2986         struct lnet_msg *msg;
2987
2988         while (!list_empty(resendq)) {
2989                 struct lnet_peer_ni *lpni;
2990
2991                 msg = list_entry(resendq->next, struct lnet_msg,
2992                                  msg_list);
2993
2994                 list_del_init(&msg->msg_list);
2995
2996                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
2997                 if (!lpni) {
2998                         lnet_net_unlock(cpt);
2999                         CERROR("Expected that a peer is already created for %s\n",
3000                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3001                         msg->msg_no_resend = true;
3002                         lnet_finalize(msg, -EFAULT);
3003                         lnet_net_lock(cpt);
3004                 } else {
3005                         int rc;
3006
3007                         lnet_peer_ni_decref_locked(lpni);
3008
3009                         lnet_net_unlock(cpt);
3010                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3011                                libcfs_nid2str(msg->msg_src_nid_param),
3012                                libcfs_id2str(msg->msg_target),
3013                                lnet_msgtyp2str(msg->msg_type),
3014                                msg->msg_recovery,
3015                                msg->msg_retry_count);
3016                         rc = lnet_send(msg->msg_src_nid_param, msg,
3017                                        msg->msg_rtr_nid_param);
3018                         if (rc) {
3019                                 CERROR("Error sending %s to %s: %d\n",
3020                                        lnet_msgtyp2str(msg->msg_type),
3021                                        libcfs_id2str(msg->msg_target), rc);
3022                                 msg->msg_no_resend = true;
3023                                 lnet_finalize(msg, rc);
3024                         }
3025                         lnet_net_lock(cpt);
3026                         if (!rc)
3027                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3028                 }
3029         }
3030 }
3031
3032 static void
3033 lnet_resend_pending_msgs(void)
3034 {
3035         int i;
3036
3037         cfs_cpt_for_each(i, lnet_cpt_table()) {
3038                 lnet_net_lock(i);
3039                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3040                 lnet_net_unlock(i);
3041         }
3042 }
3043
3044 /* called with cpt and ni_lock held */
3045 static void
3046 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3047 {
3048         struct lnet_handle_md recovery_mdh;
3049
3050         LNetInvalidateMDHandle(&recovery_mdh);
3051
3052         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3053             force) {
3054                 recovery_mdh = ni->ni_ping_mdh;
3055                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3056         }
3057         lnet_ni_unlock(ni);
3058         lnet_net_unlock(cpt);
3059         if (!LNetMDHandleIsInvalid(recovery_mdh))
3060                 LNetMDUnlink(recovery_mdh);
3061         lnet_net_lock(cpt);
3062         lnet_ni_lock(ni);
3063 }
3064
3065 static void
3066 lnet_recover_local_nis(void)
3067 {
3068         struct lnet_mt_event_info *ev_info;
3069         LIST_HEAD(processed_list);
3070         LIST_HEAD(local_queue);
3071         struct lnet_handle_md mdh;
3072         struct lnet_ni *tmp;
3073         struct lnet_ni *ni;
3074         lnet_nid_t nid;
3075         int healthv;
3076         int rc;
3077
3078         /*
3079          * splice the recovery queue on a local queue. We will iterate
3080          * through the local queue and update it as needed. Once we're
3081          * done with the traversal, we'll splice the local queue back on
3082          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3083          * will be traversed in the next iteration.
3084          */
3085         lnet_net_lock(0);
3086         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3087                          &local_queue);
3088         lnet_net_unlock(0);
3089
3090         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3091                 /*
3092                  * if an NI is being deleted or it is now healthy, there
3093                  * is no need to keep it around in the recovery queue.
3094                  * The monitor thread is the only thread responsible for
3095                  * removing the NI from the recovery queue.
3096                  * Multiple threads can be adding NIs to the recovery
3097                  * queue.
3098                  */
3099                 healthv = atomic_read(&ni->ni_healthv);
3100
3101                 lnet_net_lock(0);
3102                 lnet_ni_lock(ni);
3103                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3104                     healthv == LNET_MAX_HEALTH_VALUE) {
3105                         list_del_init(&ni->ni_recovery);
3106                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3107                         lnet_ni_unlock(ni);
3108                         lnet_ni_decref_locked(ni, 0);
3109                         lnet_net_unlock(0);
3110                         continue;
3111                 }
3112
3113                 /*
3114                  * if the local NI failed recovery we must unlink the md.
3115                  * But we want to keep the local_ni on the recovery queue
3116                  * so we can continue the attempts to recover it.
3117                  */
3118                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3119                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3120                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3121                 }
3122
3123                 lnet_ni_unlock(ni);
3124                 lnet_net_unlock(0);
3125
3126
3127                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3128                        libcfs_nid2str(ni->ni_nid));
3129
3130                 lnet_ni_lock(ni);
3131                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3132                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3133                         lnet_ni_unlock(ni);
3134
3135                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3136                         if (!ev_info) {
3137                                 CERROR("out of memory. Can't recover %s\n",
3138                                        libcfs_nid2str(ni->ni_nid));
3139                                 lnet_ni_lock(ni);
3140                                 ni->ni_recovery_state &=
3141                                   ~LNET_NI_RECOVERY_PENDING;
3142                                 lnet_ni_unlock(ni);
3143                                 continue;
3144                         }
3145
3146                         mdh = ni->ni_ping_mdh;
3147                         /*
3148                          * Invalidate the ni mdh in case it's deleted.
3149                          * We'll unlink the mdh in this case below.
3150                          */
3151                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3152                         nid = ni->ni_nid;
3153
3154                         /*
3155                          * remove the NI from the local queue and drop the
3156                          * reference count to it while we're recovering
3157                          * it. The reason for that, is that the NI could
3158                          * be deleted, and the way the code is structured
3159                          * is if we don't drop the NI, then the deletion
3160                          * code will enter a loop waiting for the
3161                          * reference count to be removed while holding the
3162                          * ln_mutex_lock(). When we look up the peer to
3163                          * send to in lnet_select_pathway() we will try to
3164                          * lock the ln_mutex_lock() as well, leading to
3165                          * a deadlock. By dropping the refcount and
3166                          * removing it from the list, we allow for the NI
3167                          * to be removed, then we use the cached NID to
3168                          * look it up again. If it's gone, then we just
3169                          * continue examining the rest of the queue.
3170                          */
3171                         lnet_net_lock(0);
3172                         list_del_init(&ni->ni_recovery);
3173                         lnet_ni_decref_locked(ni, 0);
3174                         lnet_net_unlock(0);
3175
3176                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3177                         ev_info->mt_nid = nid;
3178                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3179                                             ev_info, the_lnet.ln_mt_eq, true);
3180                         /* lookup the nid again */
3181                         lnet_net_lock(0);
3182                         ni = lnet_nid2ni_locked(nid, 0);
3183                         if (!ni) {
3184                                 /*
3185                                  * the NI has been deleted when we dropped
3186                                  * the ref count
3187                                  */
3188                                 lnet_net_unlock(0);
3189                                 LNetMDUnlink(mdh);
3190                                 continue;
3191                         }
3192                         /*
3193                          * Same note as in lnet_recover_peer_nis(). When
3194                          * we're sending the ping, the NI is free to be
3195                          * deleted or manipulated. By this point it
3196                          * could've been added back on the recovery queue,
3197                          * and a refcount taken on it.
3198                          * So we can't just add it blindly again or we'll
3199                          * corrupt the queue. We must check under lock if
3200                          * it's not on any list and if not then add it
3201                          * to the processed list, which will eventually be
3202                          * spliced back on to the recovery queue.
3203                          */
3204                         ni->ni_ping_mdh = mdh;
3205                         if (list_empty(&ni->ni_recovery)) {
3206                                 list_add_tail(&ni->ni_recovery, &processed_list);
3207                                 lnet_ni_addref_locked(ni, 0);
3208                         }
3209                         lnet_net_unlock(0);
3210
3211                         lnet_ni_lock(ni);
3212                         if (rc)
3213                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3214                 }
3215                 lnet_ni_unlock(ni);
3216         }
3217
3218         /*
3219          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3220          * reexamined in the next iteration.
3221          */
3222         list_splice_init(&processed_list, &local_queue);
3223         lnet_net_lock(0);
3224         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3225         lnet_net_unlock(0);
3226 }
3227
3228 static int
3229 lnet_resendqs_create(void)
3230 {
3231         struct list_head **resendqs;
3232         resendqs = lnet_create_array_of_queues();
3233
3234         if (!resendqs)
3235                 return -ENOMEM;
3236
3237         lnet_net_lock(LNET_LOCK_EX);
3238         the_lnet.ln_mt_resendqs = resendqs;
3239         lnet_net_unlock(LNET_LOCK_EX);
3240
3241         return 0;
3242 }
3243
3244 static void
3245 lnet_clean_local_ni_recoveryq(void)
3246 {
3247         struct lnet_ni *ni;
3248
3249         /* This is only called when the monitor thread has stopped */
3250         lnet_net_lock(0);
3251
3252         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3253                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3254                                 struct lnet_ni, ni_recovery);
3255                 list_del_init(&ni->ni_recovery);
3256                 lnet_ni_lock(ni);
3257                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3258                 lnet_ni_unlock(ni);
3259                 lnet_ni_decref_locked(ni, 0);
3260         }
3261
3262         lnet_net_unlock(0);
3263 }
3264
3265 static void
3266 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3267                                      bool force)
3268 {
3269         struct lnet_handle_md recovery_mdh;
3270
3271         LNetInvalidateMDHandle(&recovery_mdh);
3272
3273         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3274                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3275                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3276         }
3277         spin_unlock(&lpni->lpni_lock);
3278         lnet_net_unlock(cpt);
3279         if (!LNetMDHandleIsInvalid(recovery_mdh))
3280                 LNetMDUnlink(recovery_mdh);
3281         lnet_net_lock(cpt);
3282         spin_lock(&lpni->lpni_lock);
3283 }
3284
3285 static void
3286 lnet_clean_peer_ni_recoveryq(void)
3287 {
3288         struct lnet_peer_ni *lpni, *tmp;
3289
3290         lnet_net_lock(LNET_LOCK_EX);
3291
3292         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3293                                  lpni_recovery) {
3294                 list_del_init(&lpni->lpni_recovery);
3295                 spin_lock(&lpni->lpni_lock);
3296                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3297                 spin_unlock(&lpni->lpni_lock);
3298                 lnet_peer_ni_decref_locked(lpni);
3299         }
3300
3301         lnet_net_unlock(LNET_LOCK_EX);
3302 }
3303
3304 static void
3305 lnet_clean_resendqs(void)
3306 {
3307         struct lnet_msg *msg, *tmp;
3308         LIST_HEAD(msgs);
3309         int i;
3310
3311         cfs_cpt_for_each(i, lnet_cpt_table()) {
3312                 lnet_net_lock(i);
3313                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3314                 lnet_net_unlock(i);
3315                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3316                         list_del_init(&msg->msg_list);
3317                         msg->msg_no_resend = true;
3318                         lnet_finalize(msg, -ESHUTDOWN);
3319                 }
3320         }
3321
3322         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3323 }
3324
3325 static void
3326 lnet_recover_peer_nis(void)
3327 {
3328         struct lnet_mt_event_info *ev_info;
3329         LIST_HEAD(processed_list);
3330         LIST_HEAD(local_queue);
3331         struct lnet_handle_md mdh;
3332         struct lnet_peer_ni *lpni;
3333         struct lnet_peer_ni *tmp;
3334         lnet_nid_t nid;
3335         int healthv;
3336         int rc;
3337
3338         /*
3339          * Always use cpt 0 for locking across all interactions with
3340          * ln_mt_peerNIRecovq
3341          */
3342         lnet_net_lock(0);
3343         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3344                          &local_queue);
3345         lnet_net_unlock(0);
3346
3347         list_for_each_entry_safe(lpni, tmp, &local_queue,
3348                                  lpni_recovery) {
3349                 /*
3350                  * The same protection strategy is used here as is in the
3351                  * local recovery case.
3352                  */
3353                 lnet_net_lock(0);
3354                 healthv = atomic_read(&lpni->lpni_healthv);
3355                 spin_lock(&lpni->lpni_lock);
3356                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3357                     healthv == LNET_MAX_HEALTH_VALUE) {
3358                         list_del_init(&lpni->lpni_recovery);
3359                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3360                         spin_unlock(&lpni->lpni_lock);
3361                         lnet_peer_ni_decref_locked(lpni);
3362                         lnet_net_unlock(0);
3363                         continue;
3364                 }
3365
3366                 /*
3367                  * If the peer NI has failed recovery we must unlink the
3368                  * md. But we want to keep the peer ni on the recovery
3369                  * queue so we can try to continue recovering it
3370                  */
3371                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3372                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3373                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3374                 }
3375
3376                 spin_unlock(&lpni->lpni_lock);
3377                 lnet_net_unlock(0);
3378
3379                 /*
3380                  * NOTE: we're racing with peer deletion from user space.
3381                  * It's possible that a peer is deleted after we check its
3382                  * state. In this case the recovery can create a new peer
3383                  */
3384                 spin_lock(&lpni->lpni_lock);
3385                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3386                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3387                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3388                         spin_unlock(&lpni->lpni_lock);
3389
3390                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3391                         if (!ev_info) {
3392                                 CERROR("out of memory. Can't recover %s\n",
3393                                        libcfs_nid2str(lpni->lpni_nid));
3394                                 spin_lock(&lpni->lpni_lock);
3395                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3396                                 spin_unlock(&lpni->lpni_lock);
3397                                 continue;
3398                         }
3399
3400                         /* look at the comments in lnet_recover_local_nis() */
3401                         mdh = lpni->lpni_recovery_ping_mdh;
3402                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3403                         nid = lpni->lpni_nid;
3404                         lnet_net_lock(0);
3405                         list_del_init(&lpni->lpni_recovery);
3406                         lnet_peer_ni_decref_locked(lpni);
3407                         lnet_net_unlock(0);
3408
3409                         ev_info->mt_type = MT_TYPE_PEER_NI;
3410                         ev_info->mt_nid = nid;
3411                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3412                                             ev_info, the_lnet.ln_mt_eq, true);
3413                         lnet_net_lock(0);
3414                         /*
3415                          * lnet_find_peer_ni_locked() grabs a refcount for
3416                          * us. No need to take it explicitly.
3417                          */
3418                         lpni = lnet_find_peer_ni_locked(nid);
3419                         if (!lpni) {
3420                                 lnet_net_unlock(0);
3421                                 LNetMDUnlink(mdh);
3422                                 continue;
3423                         }
3424
3425                         lpni->lpni_recovery_ping_mdh = mdh;
3426                         /*
3427                          * While we're unlocked the lpni could've been
3428                          * readded on the recovery queue. In this case we
3429                          * don't need to add it to the local queue, since
3430                          * it's already on there and the thread that added
3431                          * it would've incremented the refcount on the
3432                          * peer, which means we need to decref the refcount
3433                          * that was implicitly grabbed by find_peer_ni_locked.
3434                          * Otherwise, if the lpni is still not on
3435                          * the recovery queue, then we'll add it to the
3436                          * processed list.
3437                          */
3438                         if (list_empty(&lpni->lpni_recovery))
3439                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3440                         else
3441                                 lnet_peer_ni_decref_locked(lpni);
3442                         lnet_net_unlock(0);
3443
3444                         spin_lock(&lpni->lpni_lock);
3445                         if (rc)
3446                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3447                 }
3448                 spin_unlock(&lpni->lpni_lock);
3449         }
3450
3451         list_splice_init(&processed_list, &local_queue);
3452         lnet_net_lock(0);
3453         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3454         lnet_net_unlock(0);
3455 }
3456
3457 static int
3458 lnet_monitor_thread(void *arg)
3459 {
3460         time64_t recovery_timeout = 0;
3461         time64_t rsp_timeout = 0;
3462         int interval;
3463         time64_t now;
3464
3465         wait_for_completion(&the_lnet.ln_started);
3466         /*
3467          * The monitor thread takes care of the following:
3468          *  1. Checks the aliveness of routers
3469          *  2. Checks if there are messages on the resend queue to resend
3470          *     them.
3471          *  3. Check if there are any NIs on the local recovery queue and
3472          *     pings them
3473          *  4. Checks if there are any NIs on the remote recovery queue
3474          *     and pings them.
3475          */
3476         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3477                 now = ktime_get_real_seconds();
3478
3479                 if (lnet_router_checker_active())
3480                         lnet_check_routers();
3481
3482                 lnet_resend_pending_msgs();
3483
3484                 if (now >= rsp_timeout) {
3485                         lnet_finalize_expired_responses();
3486                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3487                 }
3488
3489                 if (now >= recovery_timeout) {
3490                         lnet_recover_local_nis();
3491                         lnet_recover_peer_nis();
3492                         recovery_timeout = now + lnet_recovery_interval;
3493                 }
3494
3495                 /*
3496                  * TODO do we need to check if we should sleep without
3497                  * timeout?  Technically, an active system will always
3498                  * have messages in flight so this check will always
3499                  * evaluate to false. And on an idle system do we care
3500                  * if we wake up every 1 second? Although, we've seen
3501                  * cases where we get a complaint that an idle thread
3502                  * is waking up unnecessarily.
3503                  *
3504                  * Take into account the current net_count when you wake
3505                  * up for alive router checking, since we need to check
3506                  * possibly as many networks as we have configured.
3507                  */
3508                 interval = min(lnet_recovery_interval,
3509                                min((unsigned int) alive_router_check_interval /
3510                                         lnet_current_net_count,
3511                                    lnet_transaction_timeout / 2));
3512                 wait_for_completion_interruptible_timeout(
3513                         &the_lnet.ln_mt_wait_complete,
3514                         cfs_time_seconds(interval));
3515                 /* Must re-init the completion before testing anything,
3516                  * including ln_mt_state.
3517                  */
3518                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3519         }
3520
3521         /* Shutting down */
3522         lnet_net_lock(LNET_LOCK_EX);
3523         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3524         lnet_net_unlock(LNET_LOCK_EX);
3525
3526         /* signal that the monitor thread is exiting */
3527         up(&the_lnet.ln_mt_signal);
3528
3529         return 0;
3530 }
3531
3532 /*
3533  * lnet_send_ping
3534  * Sends a ping.
3535  * Returns == 0 if success
3536  * Returns > 0 if LNetMDBind or prior fails
3537  * Returns < 0 if LNetGet fails
3538  */
3539 int
3540 lnet_send_ping(lnet_nid_t dest_nid,
3541                struct lnet_handle_md *mdh, int nnis,
3542                void *user_data, struct lnet_eq *eq, bool recovery)
3543 {
3544         struct lnet_md md = { NULL };
3545         struct lnet_process_id id;
3546         struct lnet_ping_buffer *pbuf;
3547         int rc;
3548
3549         if (dest_nid == LNET_NID_ANY) {
3550                 rc = -EHOSTUNREACH;
3551                 goto fail_error;
3552         }
3553
3554         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3555         if (!pbuf) {
3556                 rc = ENOMEM;
3557                 goto fail_error;
3558         }
3559
3560         /* initialize md content */
3561         md.start     = &pbuf->pb_info;
3562         md.length    = LNET_PING_INFO_SIZE(nnis);
3563         md.threshold = 2; /* GET/REPLY */
3564         md.max_size  = 0;
3565         md.options   = LNET_MD_TRUNCATE;
3566         md.user_ptr  = user_data;
3567         md.eq_handle = eq;
3568
3569         rc = LNetMDBind(md, LNET_UNLINK, mdh);
3570         if (rc) {
3571                 lnet_ping_buffer_decref(pbuf);
3572                 CERROR("Can't bind MD: %d\n", rc);
3573                 rc = -rc; /* change the rc to positive */
3574                 goto fail_error;
3575         }
3576         id.pid = LNET_PID_LUSTRE;
3577         id.nid = dest_nid;
3578
3579         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3580                      LNET_RESERVED_PORTAL,
3581                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3582
3583         if (rc)
3584                 goto fail_unlink_md;
3585
3586         return 0;
3587
3588 fail_unlink_md:
3589         LNetMDUnlink(*mdh);
3590         LNetInvalidateMDHandle(mdh);
3591 fail_error:
3592         return rc;
3593 }
3594
3595 static void
3596 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3597                            int status, bool unlink_event)
3598 {
3599         lnet_nid_t nid = ev_info->mt_nid;
3600
3601         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3602                 struct lnet_ni *ni;
3603
3604                 lnet_net_lock(0);
3605                 ni = lnet_nid2ni_locked(nid, 0);
3606                 if (!ni) {
3607                         lnet_net_unlock(0);
3608                         return;
3609                 }
3610                 lnet_ni_lock(ni);
3611                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3612                 if (status)
3613                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3614                 lnet_ni_unlock(ni);
3615                 lnet_net_unlock(0);
3616
3617                 if (status != 0) {
3618                         CERROR("local NI (%s) recovery failed with %d\n",
3619                                libcfs_nid2str(nid), status);
3620                         return;
3621                 }
3622                 /*
3623                  * need to increment healthv for the ni here, because in
3624                  * the lnet_finalize() path we don't have access to this
3625                  * NI. And in order to get access to it, we'll need to
3626                  * carry forward too much information.
3627                  * In the peer case, it'll naturally be incremented
3628                  */
3629                 if (!unlink_event)
3630                         lnet_inc_healthv(&ni->ni_healthv);
3631         } else {
3632                 struct lnet_peer_ni *lpni;
3633                 int cpt;
3634
3635                 cpt = lnet_net_lock_current();
3636                 lpni = lnet_find_peer_ni_locked(nid);
3637                 if (!lpni) {
3638                         lnet_net_unlock(cpt);
3639                         return;
3640                 }
3641                 spin_lock(&lpni->lpni_lock);
3642                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3643                 if (status)
3644                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3645                 spin_unlock(&lpni->lpni_lock);
3646                 lnet_peer_ni_decref_locked(lpni);
3647                 lnet_net_unlock(cpt);
3648
3649                 if (status != 0)
3650                         CERROR("peer NI (%s) recovery failed with %d\n",
3651                                libcfs_nid2str(nid), status);
3652         }
3653 }
3654
3655 void
3656 lnet_mt_event_handler(struct lnet_event *event)
3657 {
3658         struct lnet_mt_event_info *ev_info = event->md.user_ptr;
3659         struct lnet_ping_buffer *pbuf;
3660
3661         /* TODO: remove assert */
3662         LASSERT(event->type == LNET_EVENT_REPLY ||
3663                 event->type == LNET_EVENT_SEND ||
3664                 event->type == LNET_EVENT_UNLINK);
3665
3666         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3667                event->status);
3668
3669         switch (event->type) {
3670         case LNET_EVENT_UNLINK:
3671                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3672                        libcfs_nid2str(ev_info->mt_nid));
3673                 /* fallthrough */
3674         case LNET_EVENT_REPLY:
3675                 lnet_handle_recovery_reply(ev_info, event->status,
3676                                            event->type == LNET_EVENT_UNLINK);
3677                 break;
3678         case LNET_EVENT_SEND:
3679                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3680                                libcfs_nid2str(ev_info->mt_nid),
3681                                (event->status) ? "unsuccessfully" :
3682                                "successfully", event->status);
3683                 break;
3684         default:
3685                 CERROR("Unexpected event: %d\n", event->type);
3686                 break;
3687         }
3688         if (event->unlinked) {
3689                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3690                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
3691                 lnet_ping_buffer_decref(pbuf);
3692         }
3693 }
3694
3695 static int
3696 lnet_rsp_tracker_create(void)
3697 {
3698         struct list_head **rstqs;
3699         rstqs = lnet_create_array_of_queues();
3700
3701         if (!rstqs)
3702                 return -ENOMEM;
3703
3704         the_lnet.ln_mt_rstq = rstqs;
3705
3706         return 0;
3707 }
3708
3709 static void
3710 lnet_rsp_tracker_clean(void)
3711 {
3712         lnet_finalize_expired_responses();
3713
3714         cfs_percpt_free(the_lnet.ln_mt_rstq);
3715         the_lnet.ln_mt_rstq = NULL;
3716 }
3717
3718 int lnet_monitor_thr_start(void)
3719 {
3720         int rc = 0;
3721         struct task_struct *task;
3722
3723         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3724                 return -EALREADY;
3725
3726         rc = lnet_resendqs_create();
3727         if (rc)
3728                 return rc;
3729
3730         rc = lnet_rsp_tracker_create();
3731         if (rc)
3732                 goto clean_queues;
3733
3734         sema_init(&the_lnet.ln_mt_signal, 0);
3735
3736         lnet_net_lock(LNET_LOCK_EX);
3737         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3738         lnet_net_unlock(LNET_LOCK_EX);
3739         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3740         if (IS_ERR(task)) {
3741                 rc = PTR_ERR(task);
3742                 CERROR("Can't start monitor thread: %d\n", rc);
3743                 goto clean_thread;
3744         }
3745
3746         return 0;
3747
3748 clean_thread:
3749         lnet_net_lock(LNET_LOCK_EX);
3750         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3751         lnet_net_unlock(LNET_LOCK_EX);
3752         /* block until event callback signals exit */
3753         down(&the_lnet.ln_mt_signal);
3754         /* clean up */
3755         lnet_net_lock(LNET_LOCK_EX);
3756         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3757         lnet_net_unlock(LNET_LOCK_EX);
3758         lnet_rsp_tracker_clean();
3759         lnet_clean_local_ni_recoveryq();
3760         lnet_clean_peer_ni_recoveryq();
3761         lnet_clean_resendqs();
3762         the_lnet.ln_mt_eq = NULL;
3763         return rc;
3764 clean_queues:
3765         lnet_rsp_tracker_clean();
3766         lnet_clean_local_ni_recoveryq();
3767         lnet_clean_peer_ni_recoveryq();
3768         lnet_clean_resendqs();
3769         return rc;
3770 }
3771
3772 void lnet_monitor_thr_stop(void)
3773 {
3774         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3775                 return;
3776
3777         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3778         lnet_net_lock(LNET_LOCK_EX);
3779         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3780         lnet_net_unlock(LNET_LOCK_EX);
3781
3782         /* tell the monitor thread that we're shutting down */
3783         complete(&the_lnet.ln_mt_wait_complete);
3784
3785         /* block until monitor thread signals that it's done */
3786         down(&the_lnet.ln_mt_signal);
3787         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3788
3789         /* perform cleanup tasks */
3790         lnet_rsp_tracker_clean();
3791         lnet_clean_local_ni_recoveryq();
3792         lnet_clean_peer_ni_recoveryq();
3793         lnet_clean_resendqs();
3794 }
3795
3796 void
3797 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3798                   __u32 msg_type)
3799 {
3800         lnet_net_lock(cpt);
3801         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3802         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3803         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3804         lnet_net_unlock(cpt);
3805
3806         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3807 }
3808
3809 static void
3810 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3811 {
3812         struct lnet_hdr *hdr = &msg->msg_hdr;
3813
3814         if (msg->msg_wanted != 0)
3815                 lnet_setpayloadbuffer(msg);
3816
3817         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3818
3819         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3820          * it back into the ACK during lnet_finalize() */
3821         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3822                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3823
3824         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3825                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3826 }
3827
3828 static int
3829 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3830 {
3831         struct lnet_hdr         *hdr = &msg->msg_hdr;
3832         struct lnet_match_info  info;
3833         int                     rc;
3834         bool                    ready_delay;
3835
3836         /* Convert put fields to host byte order */
3837         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3838         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3839         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3840
3841         /* Primary peer NID. */
3842         info.mi_id.nid  = msg->msg_initiator;
3843         info.mi_id.pid  = hdr->src_pid;
3844         info.mi_opc     = LNET_MD_OP_PUT;
3845         info.mi_portal  = hdr->msg.put.ptl_index;
3846         info.mi_rlength = hdr->payload_length;
3847         info.mi_roffset = hdr->msg.put.offset;
3848         info.mi_mbits   = hdr->msg.put.match_bits;
3849         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3850
3851         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3852         ready_delay = msg->msg_rx_ready_delay;
3853
3854  again:
3855         rc = lnet_ptl_match_md(&info, msg);
3856         switch (rc) {
3857         default:
3858                 LBUG();
3859
3860         case LNET_MATCHMD_OK:
3861                 lnet_recv_put(ni, msg);
3862                 return 0;
3863
3864         case LNET_MATCHMD_NONE:
3865                 if (ready_delay)
3866                         /* no eager_recv or has already called it, should
3867                          * have been attached on delayed list */
3868                         return 0;
3869
3870                 rc = lnet_ni_eager_recv(ni, msg);
3871                 if (rc == 0) {
3872                         ready_delay = true;
3873                         goto again;
3874                 }
3875                 /* fall through */
3876
3877         case LNET_MATCHMD_DROP:
3878                 CNETERR("Dropping PUT from %s portal %d match %llu"
3879                         " offset %d length %d: %d\n",
3880                         libcfs_id2str(info.mi_id), info.mi_portal,
3881                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3882
3883                 return -ENOENT; /* -ve: OK but no match */
3884         }
3885 }
3886
3887 static int
3888 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3889 {
3890         struct lnet_match_info info;
3891         struct lnet_hdr *hdr = &msg->msg_hdr;
3892         struct lnet_process_id source_id;
3893         struct lnet_handle_wire reply_wmd;
3894         int rc;
3895
3896         /* Convert get fields to host byte order */
3897         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3898         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3899         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3900         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3901
3902         source_id.nid = hdr->src_nid;
3903         source_id.pid = hdr->src_pid;
3904         /* Primary peer NID */
3905         info.mi_id.nid  = msg->msg_initiator;
3906         info.mi_id.pid  = hdr->src_pid;
3907         info.mi_opc     = LNET_MD_OP_GET;
3908         info.mi_portal  = hdr->msg.get.ptl_index;
3909         info.mi_rlength = hdr->msg.get.sink_length;
3910         info.mi_roffset = hdr->msg.get.src_offset;
3911         info.mi_mbits   = hdr->msg.get.match_bits;
3912         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3913
3914         rc = lnet_ptl_match_md(&info, msg);
3915         if (rc == LNET_MATCHMD_DROP) {
3916                 CNETERR("Dropping GET from %s portal %d match %llu"
3917                         " offset %d length %d\n",
3918                         libcfs_id2str(info.mi_id), info.mi_portal,
3919                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3920                 return -ENOENT; /* -ve: OK but no match */
3921         }
3922
3923         LASSERT(rc == LNET_MATCHMD_OK);
3924
3925         lnet_build_msg_event(msg, LNET_EVENT_GET);
3926
3927         reply_wmd = hdr->msg.get.return_wmd;
3928
3929         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3930                        msg->msg_offset, msg->msg_wanted);
3931
3932         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3933
3934         if (rdma_get) {
3935                 /* The LND completes the REPLY from her recv procedure */
3936                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3937                              msg->msg_offset, msg->msg_len, msg->msg_len);
3938                 return 0;
3939         }
3940
3941         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3942         msg->msg_receiving = 0;
3943
3944         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3945         if (rc < 0) {
3946                 /* didn't get as far as lnet_ni_send() */
3947                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3948                        libcfs_nid2str(ni->ni_nid),
3949                        libcfs_id2str(info.mi_id), rc);
3950
3951                 lnet_finalize(msg, rc);
3952         }
3953
3954         return 0;
3955 }
3956
3957 static int
3958 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
3959 {
3960         void *private = msg->msg_private;
3961         struct lnet_hdr *hdr = &msg->msg_hdr;
3962         struct lnet_process_id src = {0};
3963         struct lnet_libmd *md;
3964         unsigned int rlength;
3965         unsigned int mlength;
3966         int cpt;
3967
3968         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
3969         lnet_res_lock(cpt);
3970
3971         src.nid = hdr->src_nid;
3972         src.pid = hdr->src_pid;
3973
3974         /* NB handles only looked up by creator (no flips) */
3975         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
3976         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
3977                 CNETERR("%s: Dropping REPLY from %s for %s "
3978                         "MD %#llx.%#llx\n",
3979                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
3980                         (md == NULL) ? "invalid" : "inactive",
3981                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
3982                         hdr->msg.reply.dst_wmd.wh_object_cookie);
3983                 if (md != NULL && md->md_me != NULL)
3984                         CERROR("REPLY MD also attached to portal %d\n",
3985                                md->md_me->me_portal);
3986
3987                 lnet_res_unlock(cpt);
3988                 return -ENOENT; /* -ve: OK but no match */
3989         }
3990
3991         LASSERT(md->md_offset == 0);
3992
3993         rlength = hdr->payload_length;
3994         mlength = min(rlength, md->md_length);
3995
3996         if (mlength < rlength &&
3997             (md->md_options & LNET_MD_TRUNCATE) == 0) {
3998                 CNETERR("%s: Dropping REPLY from %s length %d "
3999                         "for MD %#llx would overflow (%d)\n",
4000                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4001                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4002                         mlength);
4003                 lnet_res_unlock(cpt);
4004                 return -ENOENT; /* -ve: OK but no match */
4005         }
4006
4007         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4008                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4009                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4010
4011         lnet_msg_attach_md(msg, md, 0, mlength);
4012
4013         if (mlength != 0)
4014                 lnet_setpayloadbuffer(msg);
4015
4016         lnet_res_unlock(cpt);
4017
4018         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4019
4020         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4021         return 0;
4022 }
4023
4024 static int
4025 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4026 {
4027         struct lnet_hdr *hdr = &msg->msg_hdr;
4028         struct lnet_process_id src = {0};
4029         struct lnet_libmd *md;
4030         int cpt;
4031
4032         src.nid = hdr->src_nid;
4033         src.pid = hdr->src_pid;
4034
4035         /* Convert ack fields to host byte order */
4036         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4037         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4038
4039         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4040         lnet_res_lock(cpt);
4041
4042         /* NB handles only looked up by creator (no flips) */
4043         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4044         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4045                 /* Don't moan; this is expected */
4046                 CDEBUG(D_NET,
4047                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4048                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4049                        (md == NULL) ? "invalid" : "inactive",
4050                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4051                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4052                 if (md != NULL && md->md_me != NULL)
4053                         CERROR("Source MD also attached to portal %d\n",
4054                                md->md_me->me_portal);
4055
4056                 lnet_res_unlock(cpt);
4057                 return -ENOENT;                  /* -ve! */
4058         }
4059
4060         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4061                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4062                hdr->msg.ack.dst_wmd.wh_object_cookie);
4063
4064         lnet_msg_attach_md(msg, md, 0, 0);
4065
4066         lnet_res_unlock(cpt);
4067
4068         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4069
4070         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4071         return 0;
4072 }
4073
4074 /**
4075  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4076  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4077  * \retval -ve                  error code
4078  */
4079 int
4080 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4081 {
4082         int     rc = 0;
4083
4084         if (!the_lnet.ln_routing)
4085                 return -ECANCELED;
4086
4087         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4088             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4089                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4090                         msg->msg_rx_ready_delay = 1;
4091                 } else {
4092                         lnet_net_unlock(msg->msg_rx_cpt);
4093                         rc = lnet_ni_eager_recv(ni, msg);
4094                         lnet_net_lock(msg->msg_rx_cpt);
4095                 }
4096         }
4097
4098         if (rc == 0)
4099                 rc = lnet_post_routed_recv_locked(msg, 0);
4100         return rc;
4101 }
4102
4103 int
4104 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4105 {
4106         int     rc;
4107
4108         switch (msg->msg_type) {
4109         case LNET_MSG_ACK:
4110                 rc = lnet_parse_ack(ni, msg);
4111                 break;
4112         case LNET_MSG_PUT:
4113                 rc = lnet_parse_put(ni, msg);
4114                 break;
4115         case LNET_MSG_GET:
4116                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4117                 break;
4118         case LNET_MSG_REPLY:
4119                 rc = lnet_parse_reply(ni, msg);
4120                 break;
4121         default: /* prevent an unused label if !kernel */
4122                 LASSERT(0);
4123                 return -EPROTO;
4124         }
4125
4126         LASSERT(rc == 0 || rc == -ENOENT);
4127         return rc;
4128 }
4129
4130 char *
4131 lnet_msgtyp2str (int type)
4132 {
4133         switch (type) {
4134         case LNET_MSG_ACK:
4135                 return ("ACK");
4136         case LNET_MSG_PUT:
4137                 return ("PUT");
4138         case LNET_MSG_GET:
4139                 return ("GET");
4140         case LNET_MSG_REPLY:
4141                 return ("REPLY");
4142         case LNET_MSG_HELLO:
4143                 return ("HELLO");
4144         default:
4145                 return ("<UNKNOWN>");
4146         }
4147 }
4148
4149 void
4150 lnet_print_hdr(struct lnet_hdr *hdr)
4151 {
4152         struct lnet_process_id src = {
4153                 .nid = hdr->src_nid,
4154                 .pid = hdr->src_pid,
4155         };
4156         struct lnet_process_id dst = {
4157                 .nid = hdr->dest_nid,
4158                 .pid = hdr->dest_pid,
4159         };
4160         char *type_str = lnet_msgtyp2str(hdr->type);
4161
4162         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4163         CWARN("    From %s\n", libcfs_id2str(src));
4164         CWARN("    To   %s\n", libcfs_id2str(dst));
4165
4166         switch (hdr->type) {
4167         default:
4168                 break;
4169
4170         case LNET_MSG_PUT:
4171                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4172                       "match bits %llu\n",
4173                       hdr->msg.put.ptl_index,
4174                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4175                       hdr->msg.put.ack_wmd.wh_object_cookie,
4176                       hdr->msg.put.match_bits);
4177                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4178                       hdr->payload_length, hdr->msg.put.offset,
4179                       hdr->msg.put.hdr_data);
4180                 break;
4181
4182         case LNET_MSG_GET:
4183                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4184                       "match bits %llu\n", hdr->msg.get.ptl_index,
4185                       hdr->msg.get.return_wmd.wh_interface_cookie,
4186                       hdr->msg.get.return_wmd.wh_object_cookie,
4187                       hdr->msg.get.match_bits);
4188                 CWARN("    Length %d, src offset %d\n",
4189                       hdr->msg.get.sink_length,
4190                       hdr->msg.get.src_offset);
4191                 break;
4192
4193         case LNET_MSG_ACK:
4194                 CWARN("    dst md %#llx.%#llx, "
4195                       "manipulated length %d\n",
4196                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4197                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4198                       hdr->msg.ack.mlength);
4199                 break;
4200
4201         case LNET_MSG_REPLY:
4202                 CWARN("    dst md %#llx.%#llx, "
4203                       "length %d\n",
4204                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4205                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4206                       hdr->payload_length);
4207         }
4208
4209 }
4210
4211 int
4212 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4213            void *private, int rdma_req)
4214 {
4215         struct lnet_peer_ni *lpni;
4216         struct lnet_msg *msg;
4217         __u32 payload_length;
4218         lnet_pid_t dest_pid;
4219         lnet_nid_t dest_nid;
4220         lnet_nid_t src_nid;
4221         bool push = false;
4222         int for_me;
4223         __u32 type;
4224         int rc = 0;
4225         int cpt;
4226
4227         LASSERT (!in_interrupt ());
4228
4229         type = le32_to_cpu(hdr->type);
4230         src_nid = le64_to_cpu(hdr->src_nid);
4231         dest_nid = le64_to_cpu(hdr->dest_nid);
4232         dest_pid = le32_to_cpu(hdr->dest_pid);
4233         payload_length = le32_to_cpu(hdr->payload_length);
4234
4235         for_me = (ni->ni_nid == dest_nid);
4236         cpt = lnet_cpt_of_nid(from_nid, ni);
4237
4238         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4239                 libcfs_nid2str(dest_nid),
4240                 libcfs_nid2str(ni->ni_nid),
4241                 libcfs_nid2str(src_nid),
4242                 lnet_msgtyp2str(type),
4243                 (for_me) ? "for me" : "routed");
4244
4245         switch (type) {
4246         case LNET_MSG_ACK:
4247         case LNET_MSG_GET:
4248                 if (payload_length > 0) {
4249                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4250                                libcfs_nid2str(from_nid),
4251                                libcfs_nid2str(src_nid),
4252                                lnet_msgtyp2str(type), payload_length);
4253                         return -EPROTO;
4254                 }
4255                 break;
4256
4257         case LNET_MSG_PUT:
4258         case LNET_MSG_REPLY:
4259                 if (payload_length >
4260                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4261                         CERROR("%s, src %s: bad %s payload %d "
4262                                "(%d max expected)\n",
4263                                libcfs_nid2str(from_nid),
4264                                libcfs_nid2str(src_nid),
4265                                lnet_msgtyp2str(type),
4266                                payload_length,
4267                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4268                         return -EPROTO;
4269                 }
4270                 break;
4271
4272         default:
4273                 CERROR("%s, src %s: Bad message type 0x%x\n",
4274                        libcfs_nid2str(from_nid),
4275                        libcfs_nid2str(src_nid), type);
4276                 return -EPROTO;
4277         }
4278
4279         if (the_lnet.ln_routing &&
4280             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4281                 lnet_ni_lock(ni);
4282                 spin_lock(&ni->ni_net->net_lock);
4283                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4284                 spin_unlock(&ni->ni_net->net_lock);
4285                 if (ni->ni_status != NULL &&
4286                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4287                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4288                         push = true;
4289                 }
4290                 lnet_ni_unlock(ni);
4291         }
4292
4293         if (push)
4294                 lnet_push_update_to_peers(1);
4295
4296         /* Regard a bad destination NID as a protocol error.  Senders should
4297          * know what they're doing; if they don't they're misconfigured, buggy
4298          * or malicious so we chop them off at the knees :) */
4299
4300         if (!for_me) {
4301                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4302                         /* should have gone direct */
4303                         CERROR("%s, src %s: Bad dest nid %s "
4304                                "(should have been sent direct)\n",
4305                                 libcfs_nid2str(from_nid),
4306                                 libcfs_nid2str(src_nid),
4307                                 libcfs_nid2str(dest_nid));
4308                         return -EPROTO;
4309                 }
4310
4311                 if (lnet_islocalnid(dest_nid)) {
4312                         /* dest is another local NI; sender should have used
4313                          * this node's NID on its own network */
4314                         CERROR("%s, src %s: Bad dest nid %s "
4315                                "(it's my nid but on a different network)\n",
4316                                 libcfs_nid2str(from_nid),
4317                                 libcfs_nid2str(src_nid),
4318                                 libcfs_nid2str(dest_nid));
4319                         return -EPROTO;
4320                 }
4321
4322                 if (rdma_req && type == LNET_MSG_GET) {
4323                         CERROR("%s, src %s: Bad optimized GET for %s "
4324                                "(final destination must be me)\n",
4325                                 libcfs_nid2str(from_nid),
4326                                 libcfs_nid2str(src_nid),
4327                                 libcfs_nid2str(dest_nid));
4328                         return -EPROTO;
4329                 }
4330
4331                 if (!the_lnet.ln_routing) {
4332                         CERROR("%s, src %s: Dropping message for %s "
4333                                "(routing not enabled)\n",
4334                                 libcfs_nid2str(from_nid),
4335                                 libcfs_nid2str(src_nid),
4336                                 libcfs_nid2str(dest_nid));
4337                         goto drop;
4338                 }
4339         }
4340
4341         /* Message looks OK; we're not going to return an error, so we MUST
4342          * call back lnd_recv() come what may... */
4343
4344         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4345             fail_peer(src_nid, 0)) {                    /* shall we now? */
4346                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4347                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4348                        lnet_msgtyp2str(type));
4349                 goto drop;
4350         }
4351
4352         if (!list_empty(&the_lnet.ln_drop_rules) &&
4353             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4354                 CDEBUG(D_NET,
4355                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4356                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4357                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4358                 goto drop;
4359         }
4360
4361         if (lnet_drop_asym_route && for_me &&
4362             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4363                 struct lnet_net *net;
4364                 struct lnet_remotenet *rnet;
4365                 bool found = true;
4366
4367                 /* we are dealing with a routed message,
4368                  * so see if route to reach src_nid goes through from_nid
4369                  */
4370                 lnet_net_lock(cpt);
4371                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4372                 if (!net) {
4373                         lnet_net_unlock(cpt);
4374                         CERROR("net %s not found\n",
4375                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4376                         return -EPROTO;
4377                 }
4378
4379                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4380                 if (rnet) {
4381                         struct lnet_peer *gw = NULL;
4382                         struct lnet_peer_ni *lpni = NULL;
4383                         struct lnet_route *route;
4384
4385                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4386                                 found = false;
4387                                 gw = route->lr_gateway;
4388                                 if (route->lr_lnet != net->net_id)
4389                                         continue;
4390                                 /*
4391                                  * if the nid is one of the gateway's NIDs
4392                                  * then this is a valid gateway
4393                                  */
4394                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4395                                                 NULL, lpni)) != NULL) {
4396                                         if (lpni->lpni_nid == from_nid) {
4397                                                 found = true;
4398                                                 break;
4399                                         }
4400                                 }
4401                         }
4402                 }
4403                 lnet_net_unlock(cpt);
4404                 if (!found) {
4405                         /* we would not use from_nid to route a message to
4406                          * src_nid
4407                          * => asymmetric routing detected but forbidden
4408                          */
4409                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4410                                libcfs_nid2str(from_nid),
4411                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4412                         goto drop;
4413                 }
4414         }
4415
4416         msg = lnet_msg_alloc();
4417         if (msg == NULL) {
4418                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4419                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4420                        lnet_msgtyp2str(type));
4421                 goto drop;
4422         }
4423
4424         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4425          * pointers NULL etc */
4426
4427         msg->msg_type = type;
4428         msg->msg_private = private;
4429         msg->msg_receiving = 1;
4430         msg->msg_rdma_get = rdma_req;
4431         msg->msg_len = msg->msg_wanted = payload_length;
4432         msg->msg_offset = 0;
4433         msg->msg_hdr = *hdr;
4434         /* for building message event */
4435         msg->msg_from = from_nid;
4436         if (!for_me) {
4437                 msg->msg_target.pid     = dest_pid;
4438                 msg->msg_target.nid     = dest_nid;
4439                 msg->msg_routing        = 1;
4440
4441         } else {
4442                 /* convert common msg->hdr fields to host byteorder */
4443                 msg->msg_hdr.type       = type;
4444                 msg->msg_hdr.src_nid    = src_nid;
4445                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4446                 msg->msg_hdr.dest_nid   = dest_nid;
4447                 msg->msg_hdr.dest_pid   = dest_pid;
4448                 msg->msg_hdr.payload_length = payload_length;
4449         }
4450
4451         lnet_net_lock(cpt);
4452         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4453         if (IS_ERR(lpni)) {
4454                 lnet_net_unlock(cpt);
4455                 CERROR("%s, src %s: Dropping %s "
4456                        "(error %ld looking up sender)\n",
4457                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4458                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4459                 lnet_msg_free(msg);
4460                 if (rc == -ESHUTDOWN)
4461                         /* We are shutting down.  Don't do anything more */
4462                         return 0;
4463                 goto drop;
4464         }
4465
4466         if (the_lnet.ln_routing)
4467                 lpni->lpni_last_alive = ktime_get_seconds();
4468
4469         msg->msg_rxpeer = lpni;
4470         msg->msg_rxni = ni;
4471         lnet_ni_addref_locked(ni, cpt);
4472         /* Multi-Rail: Primary NID of source. */
4473         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4474
4475         /*
4476          * mark the status of this lpni as UP since we received a message
4477          * from it. The ping response reports back the ns_status which is
4478          * marked on the remote as up or down and we cache it here.
4479          */
4480         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4481
4482         lnet_msg_commit(msg, cpt);
4483
4484         /* message delay simulation */
4485         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4486                      lnet_delay_rule_match_locked(hdr, msg))) {
4487                 lnet_net_unlock(cpt);
4488                 return 0;
4489         }
4490
4491         if (!for_me) {
4492                 rc = lnet_parse_forward_locked(ni, msg);
4493                 lnet_net_unlock(cpt);
4494
4495                 if (rc < 0)
4496                         goto free_drop;
4497
4498                 if (rc == LNET_CREDIT_OK) {
4499                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4500                                      0, payload_length, payload_length);
4501                 }
4502                 return 0;
4503         }
4504
4505         lnet_net_unlock(cpt);
4506
4507         rc = lnet_parse_local(ni, msg);
4508         if (rc != 0)
4509                 goto free_drop;
4510         return 0;
4511
4512  free_drop:
4513         LASSERT(msg->msg_md == NULL);
4514         lnet_finalize(msg, rc);
4515
4516  drop:
4517         lnet_drop_message(ni, cpt, private, payload_length, type);
4518         return 0;
4519 }
4520 EXPORT_SYMBOL(lnet_parse);
4521
4522 void
4523 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4524 {
4525         while (!list_empty(head)) {
4526                 struct lnet_process_id id = {0};
4527                 struct lnet_msg *msg;
4528
4529                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4530                 list_del(&msg->msg_list);
4531
4532                 id.nid = msg->msg_hdr.src_nid;
4533                 id.pid = msg->msg_hdr.src_pid;
4534
4535                 LASSERT(msg->msg_md == NULL);
4536                 LASSERT(msg->msg_rx_delayed);
4537                 LASSERT(msg->msg_rxpeer != NULL);
4538                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4539
4540                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4541                       " offset %d length %d: %s\n",
4542                       libcfs_id2str(id),
4543                       msg->msg_hdr.msg.put.ptl_index,
4544                       msg->msg_hdr.msg.put.match_bits,
4545                       msg->msg_hdr.msg.put.offset,
4546                       msg->msg_hdr.payload_length, reason);
4547
4548                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4549                  * called lnet_drop_message(), so I just hang onto msg as well
4550                  * until that's done */
4551
4552                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4553                                   msg->msg_private, msg->msg_len,
4554                                   msg->msg_type);
4555
4556                 msg->msg_no_resend = true;
4557                 /*
4558                  * NB: message will not generate event because w/o attached MD,
4559                  * but we still should give error code so lnet_msg_decommit()
4560                  * can skip counters operations and other checks.
4561                  */
4562                 lnet_finalize(msg, -ENOENT);
4563         }
4564 }
4565
4566 void
4567 lnet_recv_delayed_msg_list(struct list_head *head)
4568 {
4569         while (!list_empty(head)) {
4570                 struct lnet_msg *msg;
4571                 struct lnet_process_id id;
4572
4573                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4574                 list_del(&msg->msg_list);
4575
4576                 /* md won't disappear under me, since each msg
4577                  * holds a ref on it */
4578
4579                 id.nid = msg->msg_hdr.src_nid;
4580                 id.pid = msg->msg_hdr.src_pid;
4581
4582                 LASSERT(msg->msg_rx_delayed);
4583                 LASSERT(msg->msg_md != NULL);
4584                 LASSERT(msg->msg_rxpeer != NULL);
4585                 LASSERT(msg->msg_rxni != NULL);
4586                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4587
4588                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4589                        "match %llu offset %d length %d.\n",
4590                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4591                         msg->msg_hdr.msg.put.match_bits,
4592                         msg->msg_hdr.msg.put.offset,
4593                         msg->msg_hdr.payload_length);
4594
4595                 lnet_recv_put(msg->msg_rxni, msg);
4596         }
4597 }
4598
4599 static void
4600 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4601                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4602 {
4603         s64 timeout_ns;
4604         struct lnet_rsp_tracker *local_rspt;
4605
4606         /*
4607          * MD has a refcount taken by message so it's not going away.
4608          * The MD however can be looked up. We need to secure the access
4609          * to the md_rspt_ptr by taking the res_lock.
4610          * The rspt can be accessed without protection up to when it gets
4611          * added to the list.
4612          */
4613
4614         lnet_res_lock(cpt);
4615         local_rspt = md->md_rspt_ptr;
4616         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4617         if (local_rspt != NULL) {
4618                 /*
4619                  * we already have an rspt attached to the md, so we'll
4620                  * update the deadline on that one.
4621                  */
4622                 lnet_rspt_free(rspt, cpt);
4623         } else {
4624                 /* new md */
4625                 rspt->rspt_mdh = mdh;
4626                 rspt->rspt_cpt = cpt;
4627                 /* store the rspt so we can access it when we get the REPLY */
4628                 md->md_rspt_ptr = rspt;
4629                 local_rspt = rspt;
4630         }
4631         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4632
4633         /*
4634          * add to the list of tracked responses. It's added to tail of the
4635          * list in order to expire all the older entries first.
4636          */
4637         lnet_net_lock(cpt);
4638         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4639         lnet_net_unlock(cpt);
4640         lnet_res_unlock(cpt);
4641 }
4642
4643 /**
4644  * Initiate an asynchronous PUT operation.
4645  *
4646  * There are several events associated with a PUT: completion of the send on
4647  * the initiator node (LNET_EVENT_SEND), and when the send completes
4648  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4649  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4650  * used at the target node to indicate the completion of incoming data
4651  * delivery.
4652  *
4653  * The local events will be logged in the EQ associated with the MD pointed to
4654  * by \a mdh handle. Using a MD without an associated EQ results in these
4655  * events being discarded. In this case, the caller must have another
4656  * mechanism (e.g., a higher level protocol) for determining when it is safe
4657  * to modify the memory region associated with the MD.
4658  *
4659  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4660  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4661  *
4662  * \param self Indicates the NID of a local interface through which to send
4663  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4664  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4665  * must be "free floating" (See LNetMDBind()).
4666  * \param ack Controls whether an acknowledgment is requested.
4667  * Acknowledgments are only sent when they are requested by the initiating
4668  * process and the target MD enables them.
4669  * \param target A process identifier for the target process.
4670  * \param portal The index in the \a target's portal table.
4671  * \param match_bits The match bits to use for MD selection at the target
4672  * process.
4673  * \param offset The offset into the target MD (only used when the target
4674  * MD has the LNET_MD_MANAGE_REMOTE option set).
4675  * \param hdr_data 64 bits of user data that can be included in the message
4676  * header. This data is written to an event queue entry at the target if an
4677  * EQ is present on the matching MD.
4678  *
4679  * \retval  0      Success, and only in this case events will be generated
4680  * and logged to EQ (if it exists).
4681  * \retval -EIO    Simulated failure.
4682  * \retval -ENOMEM Memory allocation failure.
4683  * \retval -ENOENT Invalid MD object.
4684  *
4685  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4686  */
4687 int
4688 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4689         struct lnet_process_id target, unsigned int portal,
4690         __u64 match_bits, unsigned int offset,
4691         __u64 hdr_data)
4692 {
4693         struct lnet_msg *msg;
4694         struct lnet_libmd *md;
4695         int cpt;
4696         int rc;
4697         struct lnet_rsp_tracker *rspt = NULL;
4698
4699         LASSERT(the_lnet.ln_refcount > 0);
4700
4701         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4702             fail_peer(target.nid, 1)) {                 /* shall we now? */
4703                 CERROR("Dropping PUT to %s: simulated failure\n",
4704                        libcfs_id2str(target));
4705                 return -EIO;
4706         }
4707
4708         msg = lnet_msg_alloc();
4709         if (msg == NULL) {
4710                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4711                        libcfs_id2str(target));
4712                 return -ENOMEM;
4713         }
4714         msg->msg_vmflush = !!memory_pressure_get();
4715
4716         cpt = lnet_cpt_of_cookie(mdh.cookie);
4717
4718         if (ack == LNET_ACK_REQ) {
4719                 rspt = lnet_rspt_alloc(cpt);
4720                 if (!rspt) {
4721                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4722                                 libcfs_id2str(target));
4723                         return -ENOMEM;
4724                 }
4725                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4726         }
4727
4728         lnet_res_lock(cpt);
4729
4730         md = lnet_handle2md(&mdh);
4731         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4732                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4733                        match_bits, portal, libcfs_id2str(target),
4734                        md == NULL ? -1 : md->md_threshold);
4735                 if (md != NULL && md->md_me != NULL)
4736                         CERROR("Source MD also attached to portal %d\n",
4737                                md->md_me->me_portal);
4738                 lnet_res_unlock(cpt);
4739
4740                 lnet_rspt_free(rspt, cpt);
4741                 lnet_msg_free(msg);
4742                 return -ENOENT;
4743         }
4744
4745         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4746
4747         lnet_msg_attach_md(msg, md, 0, 0);
4748
4749         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4750
4751         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4752         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4753         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4754         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4755
4756         /* NB handles only looked up by creator (no flips) */
4757         if (ack == LNET_ACK_REQ) {
4758                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4759                         the_lnet.ln_interface_cookie;
4760                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4761                         md->md_lh.lh_cookie;
4762         } else {
4763                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4764                         LNET_WIRE_HANDLE_COOKIE_NONE;
4765                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4766                         LNET_WIRE_HANDLE_COOKIE_NONE;
4767         }
4768
4769         lnet_res_unlock(cpt);
4770
4771         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4772
4773         if (ack == LNET_ACK_REQ)
4774                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4775
4776         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4777                                  CFS_FAIL_ONCE))
4778                 rc = -EIO;
4779         else
4780                 rc = lnet_send(self, msg, LNET_NID_ANY);
4781
4782         if (rc != 0) {
4783                 CNETERR("Error sending PUT to %s: %d\n",
4784                         libcfs_id2str(target), rc);
4785                 msg->msg_no_resend = true;
4786                 lnet_finalize(msg, rc);
4787         }
4788
4789         /* completion will be signalled by an event */
4790         return 0;
4791 }
4792 EXPORT_SYMBOL(LNetPut);
4793
4794 /*
4795  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4796  * returns a msg for the LND to pass to lnet_finalize() when the sink
4797  * data has been received.
4798  *
4799  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4800  * lnet_finalize() is called on it, so the LND must call this first
4801  */
4802 struct lnet_msg *
4803 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4804 {
4805         struct lnet_msg *msg = lnet_msg_alloc();
4806         struct lnet_libmd *getmd = getmsg->msg_md;
4807         struct lnet_process_id peer_id = getmsg->msg_target;
4808         int cpt;
4809
4810         LASSERT(!getmsg->msg_target_is_router);
4811         LASSERT(!getmsg->msg_routing);
4812
4813         if (msg == NULL) {
4814                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4815                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4816                 goto drop;
4817         }
4818
4819         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4820         lnet_res_lock(cpt);
4821
4822         LASSERT(getmd->md_refcount > 0);
4823
4824         if (getmd->md_threshold == 0) {
4825                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4826                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4827                         getmd);
4828                 lnet_res_unlock(cpt);
4829                 goto drop;
4830         }
4831
4832         LASSERT(getmd->md_offset == 0);
4833
4834         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4835                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4836
4837         /* setup information for lnet_build_msg_event */
4838         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4839         msg->msg_from = peer_id.nid;
4840         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4841         msg->msg_hdr.src_nid = peer_id.nid;
4842         msg->msg_hdr.payload_length = getmd->md_length;
4843         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4844
4845         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4846         lnet_res_unlock(cpt);
4847
4848         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4849
4850         lnet_net_lock(cpt);
4851         lnet_msg_commit(msg, cpt);
4852         lnet_net_unlock(cpt);
4853
4854         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4855
4856         return msg;
4857
4858  drop:
4859         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4860
4861         lnet_net_lock(cpt);
4862         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4863         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4864         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4865                 getmd->md_length;
4866         lnet_net_unlock(cpt);
4867
4868         if (msg != NULL)
4869                 lnet_msg_free(msg);
4870
4871         return NULL;
4872 }
4873 EXPORT_SYMBOL(lnet_create_reply_msg);
4874
4875 void
4876 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4877                        unsigned int len)
4878 {
4879         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4880          * completed and I know it. */
4881         LASSERT(reply != NULL);
4882         LASSERT(reply->msg_type == LNET_MSG_GET);
4883         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4884
4885         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4886          * the end of my buffer, I might as well be dead. */
4887         LASSERT(len <= reply->msg_ev.mlength);
4888
4889         reply->msg_ev.mlength = len;
4890 }
4891 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4892
4893 /**
4894  * Initiate an asynchronous GET operation.
4895  *
4896  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4897  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4898  * the target node in the REPLY has been written to local MD.
4899  *
4900  * On the target node, an LNET_EVENT_GET is logged when the GET request
4901  * arrives and is accepted into a MD.
4902  *
4903  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4904  * \param mdh A handle for the MD that describes the memory into which the
4905  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4906  *
4907  * \retval  0      Success, and only in this case events will be generated
4908  * and logged to EQ (if it exists) of the MD.
4909  * \retval -EIO    Simulated failure.
4910  * \retval -ENOMEM Memory allocation failure.
4911  * \retval -ENOENT Invalid MD object.
4912  */
4913 int
4914 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4915         struct lnet_process_id target, unsigned int portal,
4916         __u64 match_bits, unsigned int offset, bool recovery)
4917 {
4918         struct lnet_msg *msg;
4919         struct lnet_libmd *md;
4920         struct lnet_rsp_tracker *rspt;
4921         int cpt;
4922         int rc;
4923
4924         LASSERT(the_lnet.ln_refcount > 0);
4925
4926         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4927             fail_peer(target.nid, 1))                   /* shall we now? */
4928         {
4929                 CERROR("Dropping GET to %s: simulated failure\n",
4930                        libcfs_id2str(target));
4931                 return -EIO;
4932         }
4933
4934         msg = lnet_msg_alloc();
4935         if (!msg) {
4936                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4937                        libcfs_id2str(target));
4938                 return -ENOMEM;
4939         }
4940
4941         cpt = lnet_cpt_of_cookie(mdh.cookie);
4942
4943         rspt = lnet_rspt_alloc(cpt);
4944         if (!rspt) {
4945                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4946                        libcfs_id2str(target));
4947                 return -ENOMEM;
4948         }
4949         INIT_LIST_HEAD(&rspt->rspt_on_list);
4950
4951         msg->msg_recovery = recovery;
4952
4953         lnet_res_lock(cpt);
4954
4955         md = lnet_handle2md(&mdh);
4956         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4957                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
4958                        match_bits, portal, libcfs_id2str(target),
4959                        md == NULL ? -1 : md->md_threshold);
4960                 if (md != NULL && md->md_me != NULL)
4961                         CERROR("REPLY MD also attached to portal %d\n",
4962                                md->md_me->me_portal);
4963
4964                 lnet_res_unlock(cpt);
4965
4966                 lnet_msg_free(msg);
4967                 lnet_rspt_free(rspt, cpt);
4968                 return -ENOENT;
4969         }
4970
4971         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
4972
4973         lnet_msg_attach_md(msg, md, 0, 0);
4974
4975         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
4976
4977         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
4978         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
4979         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
4980         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
4981
4982         /* NB handles only looked up by creator (no flips) */
4983         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
4984                 the_lnet.ln_interface_cookie;
4985         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
4986                 md->md_lh.lh_cookie;
4987
4988         lnet_res_unlock(cpt);
4989
4990         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4991
4992         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4993
4994         rc = lnet_send(self, msg, LNET_NID_ANY);
4995         if (rc < 0) {
4996                 CNETERR("Error sending GET to %s: %d\n",
4997                         libcfs_id2str(target), rc);
4998                 msg->msg_no_resend = true;
4999                 lnet_finalize(msg, rc);
5000         }
5001
5002         /* completion will be signalled by an event */
5003         return 0;
5004 }
5005 EXPORT_SYMBOL(LNetGet);
5006
5007 /**
5008  * Calculate distance to node at \a dstnid.
5009  *
5010  * \param dstnid Target NID.
5011  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5012  * is saved here.
5013  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5014  * here.
5015  *
5016  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5017  * local_nid_dist_zero is set, which is the default.
5018  * \retval positives Distance to target NID, i.e. number of hops plus one.
5019  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5020  */
5021 int
5022 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5023 {
5024         struct list_head        *e;
5025         struct lnet_ni *ni = NULL;
5026         struct lnet_remotenet *rnet;
5027         __u32                   dstnet = LNET_NIDNET(dstnid);
5028         int                     hops;
5029         int                     cpt;
5030         __u32                   order = 2;
5031         struct list_head        *rn_list;
5032
5033         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5034          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5035          * keep order 0 free for 0@lo and order 1 free for a local NID
5036          * match */
5037
5038         LASSERT(the_lnet.ln_refcount > 0);
5039
5040         cpt = lnet_net_lock_current();
5041
5042         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5043                 if (ni->ni_nid == dstnid) {
5044                         if (srcnidp != NULL)
5045                                 *srcnidp = dstnid;
5046                         if (orderp != NULL) {
5047                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
5048                                         *orderp = 0;
5049                                 else
5050                                         *orderp = 1;
5051                         }
5052                         lnet_net_unlock(cpt);
5053
5054                         return local_nid_dist_zero ? 0 : 1;
5055                 }
5056
5057                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5058                         /* Check if ni was originally created in
5059                          * current net namespace.
5060                          * If not, assign order above 0xffff0000,
5061                          * to make this ni not a priority. */
5062                         if (current->nsproxy &&
5063                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5064                                         order += 0xffff0000;
5065                         if (srcnidp != NULL)
5066                                 *srcnidp = ni->ni_nid;
5067                         if (orderp != NULL)
5068                                 *orderp = order;
5069                         lnet_net_unlock(cpt);
5070                         return 1;
5071                 }
5072
5073                 order++;
5074         }
5075
5076         rn_list = lnet_net2rnethash(dstnet);
5077         list_for_each(e, rn_list) {
5078                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5079
5080                 if (rnet->lrn_net == dstnet) {
5081                         struct lnet_route *route;
5082                         struct lnet_route *shortest = NULL;
5083                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5084                         __u32 route_hops;
5085
5086                         LASSERT(!list_empty(&rnet->lrn_routes));
5087
5088                         list_for_each_entry(route, &rnet->lrn_routes,
5089                                             lr_list) {
5090                                 route_hops = route->lr_hops;
5091                                 if (route_hops == LNET_UNDEFINED_HOPS)
5092                                         route_hops = 1;
5093                                 if (shortest == NULL ||
5094                                     route_hops < shortest_hops) {
5095                                         shortest = route;
5096                                         shortest_hops = route_hops;
5097                                 }
5098                         }
5099
5100                         LASSERT(shortest != NULL);
5101                         hops = shortest_hops;
5102                         if (srcnidp != NULL) {
5103                                 struct lnet_net *net;
5104                                 net = lnet_get_net_locked(shortest->lr_lnet);
5105                                 LASSERT(net);
5106                                 ni = lnet_get_next_ni_locked(net, NULL);
5107                                 *srcnidp = ni->ni_nid;
5108                         }
5109                         if (orderp != NULL)
5110                                 *orderp = order;
5111                         lnet_net_unlock(cpt);
5112                         return hops + 1;
5113                 }
5114                 order++;
5115         }
5116
5117         lnet_net_unlock(cpt);
5118         return -EHOSTUNREACH;
5119 }
5120 EXPORT_SYMBOL(LNetDist);