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 unsigned int
 342 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 343 {
 344         unsigned int  nob = 0;
 345
 346         LASSERT(niov == 0 || kiov != NULL);
 347         while (niov-- > 0)
 348                 nob += (kiov++)->bv_len;
 349
 350         return (nob);
 351 }
 352 EXPORT_SYMBOL(lnet_kiov_nob);
 353
 354 void
 355 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
 356                     unsigned int doffset,
 357                     unsigned int nsiov, struct bio_vec *siov,
 358                     unsigned int soffset,
 359                     unsigned int nob)
 360 {
 361         /* NB diov, siov are READ-ONLY */
 362         unsigned int    this_nob;
 363         char           *daddr = NULL;
 364         char           *saddr = NULL;
 365
 366         if (nob == 0)
 367                 return;
 368
 369         LASSERT (!in_interrupt ());
 370
 371         LASSERT (ndiov > 0);
 372         while (doffset >= diov->bv_len) {
 373                 doffset -= diov->bv_len;
 374                 diov++;
 375                 ndiov--;
 376                 LASSERT(ndiov > 0);
 377         }
 378
 379         LASSERT(nsiov > 0);
 380         while (soffset >= siov->bv_len) {
 381                 soffset -= siov->bv_len;
 382                 siov++;
 383                 nsiov--;
 384                 LASSERT(nsiov > 0);
 385         }
 386
 387         do {
 388                 LASSERT(ndiov > 0);
 389                 LASSERT(nsiov > 0);
 390                 this_nob = min3(diov->bv_len - doffset,
 391                                 siov->bv_len - soffset,
 392                                 nob);
 393
 394                 if (daddr == NULL)
 395                         daddr = ((char *)kmap(diov->bv_page)) +
 396                                 diov->bv_offset + doffset;
 397                 if (saddr == NULL)
 398                         saddr = ((char *)kmap(siov->bv_page)) +
 399                                 siov->bv_offset + soffset;
 400
 401                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 402                  * However in practice at least one of the kiovs will be mapped
 403                  * kernel pages and the map/unmap will be NOOPs */
 404
 405                 memcpy (daddr, saddr, this_nob);
 406                 nob -= this_nob;
 407
 408                 if (diov->bv_len > doffset + this_nob) {
 409                         daddr += this_nob;
 410                         doffset += this_nob;
 411                 } else {
 412                         kunmap(diov->bv_page);
 413                         daddr = NULL;
 414                         diov++;
 415                         ndiov--;
 416                         doffset = 0;
 417                 }
 418
 419                 if (siov->bv_len > soffset + this_nob) {
 420                         saddr += this_nob;
 421                         soffset += this_nob;
 422                 } else {
 423                         kunmap(siov->bv_page);
 424                         saddr = NULL;
 425                         siov++;
 426                         nsiov--;
 427                         soffset = 0;
 428                 }
 429         } while (nob > 0);
 430
 431         if (daddr != NULL)
 432                 kunmap(diov->bv_page);
 433         if (saddr != NULL)
 434                 kunmap(siov->bv_page);
 435 }
 436 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 437
 438 void
 439 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 440                     unsigned int nkiov, struct bio_vec *kiov,
 441                     unsigned int kiovoffset,
 442                     unsigned int nob)
 443 {
 444         /* NB iov, kiov are READ-ONLY */
 445         unsigned int    this_nob;
 446         char           *addr = NULL;
 447
 448         if (nob == 0)
 449                 return;
 450
 451         LASSERT (!in_interrupt ());
 452
 453         LASSERT (niov > 0);
 454         while (iovoffset >= iov->iov_len) {
 455                 iovoffset -= iov->iov_len;
 456                 iov++;
 457                 niov--;
 458                 LASSERT(niov > 0);
 459         }
 460
 461         LASSERT(nkiov > 0);
 462         while (kiovoffset >= kiov->bv_len) {
 463                 kiovoffset -= kiov->bv_len;
 464                 kiov++;
 465                 nkiov--;
 466                 LASSERT(nkiov > 0);
 467         }
 468
 469         do {
 470                 LASSERT(niov > 0);
 471                 LASSERT(nkiov > 0);
 472                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 473                                 (unsigned int)kiov->bv_len - kiovoffset,
 474                                 nob);
 475
 476                 if (addr == NULL)
 477                         addr = ((char *)kmap(kiov->bv_page)) +
 478                                 kiov->bv_offset + kiovoffset;
 479
 480                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 481                 nob -= this_nob;
 482
 483                 if (iov->iov_len > iovoffset + this_nob) {
 484                         iovoffset += this_nob;
 485                 } else {
 486                         iov++;
 487                         niov--;
 488                         iovoffset = 0;
 489                 }
 490
 491                 if (kiov->bv_len > kiovoffset + this_nob) {
 492                         addr += this_nob;
 493                         kiovoffset += this_nob;
 494                 } else {
 495                         kunmap(kiov->bv_page);
 496                         addr = NULL;
 497                         kiov++;
 498                         nkiov--;
 499                         kiovoffset = 0;
 500                 }
 501
 502         } while (nob > 0);
 503
 504         if (addr != NULL)
 505                 kunmap(kiov->bv_page);
 506 }
 507 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 508
 509 void
 510 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
 511                    unsigned int kiovoffset,
 512                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 513                    unsigned int nob)
 514 {
 515         /* NB kiov, iov are READ-ONLY */
 516         unsigned int    this_nob;
 517         char           *addr = NULL;
 518
 519         if (nob == 0)
 520                 return;
 521
 522         LASSERT (!in_interrupt ());
 523
 524         LASSERT (nkiov > 0);
 525         while (kiovoffset >= kiov->bv_len) {
 526                 kiovoffset -= kiov->bv_len;
 527                 kiov++;
 528                 nkiov--;
 529                 LASSERT(nkiov > 0);
 530         }
 531
 532         LASSERT(niov > 0);
 533         while (iovoffset >= iov->iov_len) {
 534                 iovoffset -= iov->iov_len;
 535                 iov++;
 536                 niov--;
 537                 LASSERT(niov > 0);
 538         }
 539
 540         do {
 541                 LASSERT(nkiov > 0);
 542                 LASSERT(niov > 0);
 543                 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
 544                                 (unsigned int)iov->iov_len - iovoffset,
 545                                 nob);
 546
 547                 if (addr == NULL)
 548                         addr = ((char *)kmap(kiov->bv_page)) +
 549                                 kiov->bv_offset + kiovoffset;
 550
 551                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 552                 nob -= this_nob;
 553
 554                 if (kiov->bv_len > kiovoffset + this_nob) {
 555                         addr += this_nob;
 556                         kiovoffset += this_nob;
 557                 } else {
 558                         kunmap(kiov->bv_page);
 559                         addr = NULL;
 560                         kiov++;
 561                         nkiov--;
 562                         kiovoffset = 0;
 563                 }
 564
 565                 if (iov->iov_len > iovoffset + this_nob) {
 566                         iovoffset += this_nob;
 567                 } else {
 568                         iov++;
 569                         niov--;
 570                         iovoffset = 0;
 571                 }
 572         } while (nob > 0);
 573
 574         if (addr != NULL)
 575                 kunmap(kiov->bv_page);
 576 }
 577 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 578
 579 int
 580 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
 581                   int src_niov, struct bio_vec *src,
 582                   unsigned int offset, unsigned int len)
 583 {
 584         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 585          * for exactly 'len' bytes, and return the number of entries.
 586          * NB not destructive to 'src' */
 587         unsigned int    frag_len;
 588         unsigned int    niov;
 589
 590         if (len == 0)                           /* no data => */
 591                 return (0);                     /* no frags */
 592
 593         LASSERT(src_niov > 0);
 594         while (offset >= src->bv_len) {      /* skip initial frags */
 595                 offset -= src->bv_len;
 596                 src_niov--;
 597                 src++;
 598                 LASSERT(src_niov > 0);
 599         }
 600
 601         niov = 1;
 602         for (;;) {
 603                 LASSERT(src_niov > 0);
 604                 LASSERT((int)niov <= dst_niov);
 605
 606                 frag_len = src->bv_len - offset;
 607                 dst->bv_page = src->bv_page;
 608                 dst->bv_offset = src->bv_offset + offset;
 609
 610                 if (len <= frag_len) {
 611                         dst->bv_len = len;
 612                         LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 613                         return niov;
 614                 }
 615
 616                 dst->bv_len = frag_len;
 617                 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 618
 619                 len -= frag_len;
 620                 dst++;
 621                 src++;
 622                 niov++;
 623                 src_niov--;
 624                 offset = 0;
 625         }
 626 }
 627 EXPORT_SYMBOL(lnet_extract_kiov);
 628
 629 void
 630 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 631              int delayed, unsigned int offset, unsigned int mlen,
 632              unsigned int rlen)
 633 {
 634         unsigned int niov = 0;
 635         struct kvec *iov = NULL;
 636         struct bio_vec  *kiov = NULL;
 637         int rc;
 638
 639         LASSERT (!in_interrupt ());
 640         LASSERT (mlen == 0 || msg != NULL);
 641
 642         if (msg != NULL) {
 643                 LASSERT(msg->msg_receiving);
 644                 LASSERT(!msg->msg_sending);
 645                 LASSERT(rlen == msg->msg_len);
 646                 LASSERT(mlen <= msg->msg_len);
 647                 LASSERT(msg->msg_offset == offset);
 648                 LASSERT(msg->msg_wanted == mlen);
 649
 650                 msg->msg_receiving = 0;
 651
 652                 if (mlen != 0) {
 653                         niov = msg->msg_niov;
 654                         kiov = msg->msg_kiov;
 655
 656                         LASSERT (niov > 0);
 657                         LASSERT ((iov == NULL) != (kiov == NULL));
 658                 }
 659         }
 660
 661         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 662                                              niov, kiov, offset, mlen,
 663                                              rlen);
 664         if (rc < 0)
 665                 lnet_finalize(msg, rc);
 666 }
 667
 668 static void
 669 lnet_setpayloadbuffer(struct lnet_msg *msg)
 670 {
 671         struct lnet_libmd *md = msg->msg_md;
 672
 673         LASSERT(msg->msg_len > 0);
 674         LASSERT(!msg->msg_routing);
 675         LASSERT(md != NULL);
 676         LASSERT(msg->msg_niov == 0);
 677         LASSERT(msg->msg_kiov == NULL);
 678
 679         msg->msg_niov = md->md_niov;
 680         msg->msg_kiov = md->md_kiov;
 681 }
 682
 683 void
 684 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 685                unsigned int offset, unsigned int len)
 686 {
 687         msg->msg_type = type;
 688         msg->msg_target = target;
 689         msg->msg_len = len;
 690         msg->msg_offset = offset;
 691
 692         if (len != 0)
 693                 lnet_setpayloadbuffer(msg);
 694
 695         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 696         msg->msg_hdr.type           = cpu_to_le32(type);
 697         /* dest_nid will be overwritten by lnet_select_pathway() */
 698         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 699         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 700         /* src_nid will be set later */
 701         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 702         msg->msg_hdr.payload_length = cpu_to_le32(len);
 703 }
 704
 705 static void
 706 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 707 {
 708         void *priv = msg->msg_private;
 709         int rc;
 710
 711         LASSERT(!in_interrupt());
 712         LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
 713                 (msg->msg_txcredit && msg->msg_peertxcredit));
 714
 715         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 716         if (rc < 0) {
 717                 msg->msg_no_resend = true;
 718                 lnet_finalize(msg, rc);
 719         }
 720 }
 721
 722 static int
 723 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 724 {
 725         int     rc;
 726
 727         LASSERT(!msg->msg_sending);
 728         LASSERT(msg->msg_receiving);
 729         LASSERT(!msg->msg_rx_ready_delay);
 730         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 731
 732         msg->msg_rx_ready_delay = 1;
 733         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 734                                                   &msg->msg_private);
 735         if (rc != 0) {
 736                 CERROR("recv from %s / send to %s aborted: "
 737                        "eager_recv failed %d\n",
 738                        libcfs_nid2str(msg->msg_rxpeer->lpni_nid),
 739                        libcfs_id2str(msg->msg_target), rc);
 740                 LASSERT(rc < 0); /* required by my callers */
 741         }
 742
 743         return rc;
 744 }
 745
 746 static bool
 747 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 748 {
 749         time64_t deadline;
 750
 751         deadline = lpni->lpni_last_alive +
 752                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 753
 754         /*
 755          * assume peer_ni is alive as long as we're within the configured
 756          * peer timeout
 757          */
 758         if (deadline > now)
 759                 return false;
 760
 761         return true;
 762 }
 763
 764 /* NB: returns 1 when alive, 0 when dead, negative when error;
 765  *     may drop the lnet_net_lock */
 766 static int
 767 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 768                        struct lnet_msg *msg)
 769 {
 770         time64_t now = ktime_get_seconds();
 771
 772         if (!lnet_peer_aliveness_enabled(lpni))
 773                 return -ENODEV;
 774
 775         /*
 776          * If we're resending a message, let's attempt to send it even if
 777          * the peer is down to fulfill our resend quota on the message
 778          */
 779         if (msg->msg_retry_count > 0)
 780                 return 1;
 781
 782         /* try and send recovery messages irregardless */
 783         if (msg->msg_recovery)
 784                 return 1;
 785
 786         /* always send any responses */
 787         if (lnet_msg_is_response(msg))
 788                 return 1;
 789
 790         if (!lnet_is_peer_deadline_passed(lpni, now))
 791                 return true;
 792
 793         return lnet_is_peer_ni_alive(lpni);
 794 }
 795
 796 /**
 797  * \param msg The message to be sent.
 798  * \param do_send True if lnet_ni_send() should be called in this function.
 799  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 800  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 801  *
 802  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 803  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 804  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 805  * \retval -ECANCELED If the MD of the message has been unlinked.
 806  */
 807 static int
 808 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 809 {
 810         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 811         struct lnet_ni          *ni = msg->msg_txni;
 812         int                     cpt = msg->msg_tx_cpt;
 813         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 814
 815         /* non-lnet_send() callers have checked before */
 816         LASSERT(!do_send || msg->msg_tx_delayed);
 817         LASSERT(!msg->msg_receiving);
 818         LASSERT(msg->msg_tx_committed);
 819         /* can't get here if we're sending to the loopback interface */
 820         LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
 821
 822         /* NB 'lp' is always the next hop */
 823         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 824             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 825                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 826                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 827                         msg->msg_len;
 828                 lnet_net_unlock(cpt);
 829                 if (msg->msg_txpeer)
 830                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 831                                         msg->msg_type,
 832                                         LNET_STATS_TYPE_DROP);
 833                 if (msg->msg_txni)
 834                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 835                                         msg->msg_type,
 836                                         LNET_STATS_TYPE_DROP);
 837
 838                 CNETERR("Dropping message for %s: peer not alive\n",
 839                         libcfs_id2str(msg->msg_target));
 840                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 841                 if (do_send)
 842                         lnet_finalize(msg, -EHOSTUNREACH);
 843
 844                 lnet_net_lock(cpt);
 845                 return -EHOSTUNREACH;
 846         }
 847
 848         if (msg->msg_md != NULL &&
 849             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 850                 lnet_net_unlock(cpt);
 851
 852                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 853                         "called on the MD/ME.\n",
 854                         libcfs_id2str(msg->msg_target));
 855                 if (do_send) {
 856                         msg->msg_no_resend = true;
 857                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 858                                msg, libcfs_id2str(msg->msg_target));
 859                         lnet_finalize(msg, -ECANCELED);
 860                 }
 861
 862                 lnet_net_lock(cpt);
 863                 return -ECANCELED;
 864         }
 865
 866         if (!msg->msg_peertxcredit) {
 867                 spin_lock(&lp->lpni_lock);
 868                 LASSERT((lp->lpni_txcredits < 0) ==
 869                         !list_empty(&lp->lpni_txq));
 870
 871                 msg->msg_peertxcredit = 1;
 872                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 873                 lp->lpni_txcredits--;
 874
 875                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 876                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 877
 878                 if (lp->lpni_txcredits < 0) {
 879                         msg->msg_tx_delayed = 1;
 880                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 881                         spin_unlock(&lp->lpni_lock);
 882                         return LNET_CREDIT_WAIT;
 883                 }
 884                 spin_unlock(&lp->lpni_lock);
 885         }
 886
 887         if (!msg->msg_txcredit) {
 888                 LASSERT((tq->tq_credits < 0) ==
 889                         !list_empty(&tq->tq_delayed));
 890
 891                 msg->msg_txcredit = 1;
 892                 tq->tq_credits--;
 893                 atomic_dec(&ni->ni_tx_credits);
 894
 895                 if (tq->tq_credits < tq->tq_credits_min)
 896                         tq->tq_credits_min = tq->tq_credits;
 897
 898                 if (tq->tq_credits < 0) {
 899                         msg->msg_tx_delayed = 1;
 900                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 901                         return LNET_CREDIT_WAIT;
 902                 }
 903         }
 904
 905         /* unset the tx_delay flag as we're going to send it now */
 906         msg->msg_tx_delayed = 0;
 907
 908         if (do_send) {
 909                 lnet_net_unlock(cpt);
 910                 lnet_ni_send(ni, msg);
 911                 lnet_net_lock(cpt);
 912         }
 913         return LNET_CREDIT_OK;
 914 }
 915
 916
 917 static struct lnet_rtrbufpool *
 918 lnet_msg2bufpool(struct lnet_msg *msg)
 919 {
 920         struct lnet_rtrbufpool  *rbp;
 921         int                     cpt;
 922
 923         LASSERT(msg->msg_rx_committed);
 924
 925         cpt = msg->msg_rx_cpt;
 926         rbp = &the_lnet.ln_rtrpools[cpt][0];
 927
 928         LASSERT(msg->msg_len <= LNET_MTU);
 929         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 930                 rbp++;
 931                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 932         }
 933
 934         return rbp;
 935 }
 936
 937 static int
 938 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 939 {
 940         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 941          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 942          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 943          * received or OK to receive */
 944         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 945         struct lnet_peer *lp;
 946         struct lnet_rtrbufpool *rbp;
 947         struct lnet_rtrbuf *rb;
 948
 949         LASSERT(msg->msg_kiov == NULL);
 950         LASSERT(msg->msg_niov == 0);
 951         LASSERT(msg->msg_routing);
 952         LASSERT(msg->msg_receiving);
 953         LASSERT(!msg->msg_sending);
 954         LASSERT(lpni->lpni_peer_net);
 955         LASSERT(lpni->lpni_peer_net->lpn_peer);
 956
 957         lp = lpni->lpni_peer_net->lpn_peer;
 958
 959         /* non-lnet_parse callers only receive delayed messages */
 960         LASSERT(!do_recv || msg->msg_rx_delayed);
 961
 962         if (!msg->msg_peerrtrcredit) {
 963                 /* lpni_lock protects the credit manipulation */
 964                 spin_lock(&lpni->lpni_lock);
 965
 966                 msg->msg_peerrtrcredit = 1;
 967                 lpni->lpni_rtrcredits--;
 968                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
 969                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
 970
 971                 if (lpni->lpni_rtrcredits < 0) {
 972                         spin_unlock(&lpni->lpni_lock);
 973                         /* must have checked eager_recv before here */
 974                         LASSERT(msg->msg_rx_ready_delay);
 975                         msg->msg_rx_delayed = 1;
 976                         /* lp_lock protects the lp_rtrq */
 977                         spin_lock(&lp->lp_lock);
 978                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
 979                         spin_unlock(&lp->lp_lock);
 980                         return LNET_CREDIT_WAIT;
 981                 }
 982                 spin_unlock(&lpni->lpni_lock);
 983         }
 984
 985         rbp = lnet_msg2bufpool(msg);
 986
 987         if (!msg->msg_rtrcredit) {
 988                 msg->msg_rtrcredit = 1;
 989                 rbp->rbp_credits--;
 990                 if (rbp->rbp_credits < rbp->rbp_mincredits)
 991                         rbp->rbp_mincredits = rbp->rbp_credits;
 992
 993                 if (rbp->rbp_credits < 0) {
 994                         /* must have checked eager_recv before here */
 995                         LASSERT(msg->msg_rx_ready_delay);
 996                         msg->msg_rx_delayed = 1;
 997                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
 998                         return LNET_CREDIT_WAIT;
 999                 }
1000         }
1001
1002         LASSERT(!list_empty(&rbp->rbp_bufs));
1003         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1004         list_del(&rb->rb_list);
1005
1006         msg->msg_niov = rbp->rbp_npages;
1007         msg->msg_kiov = &rb->rb_kiov[0];
1008
1009         /* unset the msg-rx_delayed flag since we're receiving the message */
1010         msg->msg_rx_delayed = 0;
1011
1012         if (do_recv) {
1013                 int cpt = msg->msg_rx_cpt;
1014
1015                 lnet_net_unlock(cpt);
1016                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1017                              0, msg->msg_len, msg->msg_len);
1018                 lnet_net_lock(cpt);
1019         }
1020         return LNET_CREDIT_OK;
1021 }
1022
1023 void
1024 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1025 {
1026         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1027         struct lnet_ni          *txni = msg->msg_txni;
1028         struct lnet_msg         *msg2;
1029
1030         if (msg->msg_txcredit) {
1031                 struct lnet_ni       *ni = msg->msg_txni;
1032                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1033
1034                 /* give back NI txcredits */
1035                 msg->msg_txcredit = 0;
1036
1037                 LASSERT((tq->tq_credits < 0) ==
1038                         !list_empty(&tq->tq_delayed));
1039
1040                 tq->tq_credits++;
1041                 atomic_inc(&ni->ni_tx_credits);
1042                 if (tq->tq_credits <= 0) {
1043                         msg2 = list_entry(tq->tq_delayed.next,
1044                                           struct lnet_msg, msg_list);
1045                         list_del(&msg2->msg_list);
1046
1047                         LASSERT(msg2->msg_txni == ni);
1048                         LASSERT(msg2->msg_tx_delayed);
1049                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1050
1051                         (void) lnet_post_send_locked(msg2, 1);
1052                 }
1053         }
1054
1055         if (msg->msg_peertxcredit) {
1056                 /* give back peer txcredits */
1057                 msg->msg_peertxcredit = 0;
1058
1059                 spin_lock(&txpeer->lpni_lock);
1060                 LASSERT((txpeer->lpni_txcredits < 0) ==
1061                         !list_empty(&txpeer->lpni_txq));
1062
1063                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1064                 LASSERT(txpeer->lpni_txqnob >= 0);
1065
1066                 txpeer->lpni_txcredits++;
1067                 if (txpeer->lpni_txcredits <= 0) {
1068                         int msg2_cpt;
1069
1070                         msg2 = list_entry(txpeer->lpni_txq.next,
1071                                               struct lnet_msg, msg_list);
1072                         list_del(&msg2->msg_list);
1073                         spin_unlock(&txpeer->lpni_lock);
1074
1075                         LASSERT(msg2->msg_txpeer == txpeer);
1076                         LASSERT(msg2->msg_tx_delayed);
1077
1078                         msg2_cpt = msg2->msg_tx_cpt;
1079
1080                         /*
1081                          * The msg_cpt can be different from the msg2_cpt
1082                          * so we need to make sure we lock the correct cpt
1083                          * for msg2.
1084                          * Once we call lnet_post_send_locked() it is no
1085                          * longer safe to access msg2, since it could've
1086                          * been freed by lnet_finalize(), but we still
1087                          * need to relock the correct cpt, so we cache the
1088                          * msg2_cpt for the purpose of the check that
1089                          * follows the call to lnet_pose_send_locked().
1090                          */
1091                         if (msg2_cpt != msg->msg_tx_cpt) {
1092                                 lnet_net_unlock(msg->msg_tx_cpt);
1093                                 lnet_net_lock(msg2_cpt);
1094                         }
1095                         (void) lnet_post_send_locked(msg2, 1);
1096                         if (msg2_cpt != msg->msg_tx_cpt) {
1097                                 lnet_net_unlock(msg2_cpt);
1098                                 lnet_net_lock(msg->msg_tx_cpt);
1099                         }
1100                 } else {
1101                         spin_unlock(&txpeer->lpni_lock);
1102                 }
1103         }
1104
1105         if (txni != NULL) {
1106                 msg->msg_txni = NULL;
1107                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1108         }
1109
1110         if (txpeer != NULL) {
1111                 msg->msg_txpeer = NULL;
1112                 lnet_peer_ni_decref_locked(txpeer);
1113         }
1114 }
1115
1116 void
1117 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1118 {
1119         struct lnet_msg *msg;
1120
1121         if (list_empty(&rbp->rbp_msgs))
1122                 return;
1123         msg = list_entry(rbp->rbp_msgs.next,
1124                          struct lnet_msg, msg_list);
1125         list_del(&msg->msg_list);
1126
1127         (void)lnet_post_routed_recv_locked(msg, 1);
1128 }
1129
1130 void
1131 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1132 {
1133         struct lnet_msg *msg;
1134         struct lnet_msg *tmp;
1135
1136         lnet_net_unlock(cpt);
1137
1138         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1139                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1140                              0, 0, 0, msg->msg_hdr.payload_length);
1141                 list_del_init(&msg->msg_list);
1142                 msg->msg_no_resend = true;
1143                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1144                 lnet_finalize(msg, -ECANCELED);
1145         }
1146
1147         lnet_net_lock(cpt);
1148 }
1149
1150 void
1151 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1152 {
1153         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1154         struct lnet_peer *lp;
1155         struct lnet_ni *rxni = msg->msg_rxni;
1156         struct lnet_msg *msg2;
1157
1158         if (msg->msg_rtrcredit) {
1159                 /* give back global router credits */
1160                 struct lnet_rtrbuf *rb;
1161                 struct lnet_rtrbufpool *rbp;
1162
1163                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1164                  * there until it gets one allocated, or aborts the wait
1165                  * itself */
1166                 LASSERT(msg->msg_kiov != NULL);
1167
1168                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1169                 rbp = rb->rb_pool;
1170
1171                 msg->msg_kiov = NULL;
1172                 msg->msg_rtrcredit = 0;
1173
1174                 LASSERT(rbp == lnet_msg2bufpool(msg));
1175
1176                 LASSERT((rbp->rbp_credits > 0) ==
1177                         !list_empty(&rbp->rbp_bufs));
1178
1179                 /* If routing is now turned off, we just drop this buffer and
1180                  * don't bother trying to return credits.  */
1181                 if (!the_lnet.ln_routing) {
1182                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1183                         goto routing_off;
1184                 }
1185
1186                 /* It is possible that a user has lowered the desired number of
1187                  * buffers in this pool.  Make sure we never put back
1188                  * more buffers than the stated number. */
1189                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1190                         /* Discard this buffer so we don't have too
1191                          * many. */
1192                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1193                         rbp->rbp_nbuffers--;
1194                 } else {
1195                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1196                         rbp->rbp_credits++;
1197                         if (rbp->rbp_credits <= 0)
1198                                 lnet_schedule_blocked_locked(rbp);
1199                 }
1200         }
1201
1202 routing_off:
1203         if (msg->msg_peerrtrcredit) {
1204                 LASSERT(rxpeerni);
1205                 LASSERT(rxpeerni->lpni_peer_net);
1206                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1207
1208                 /* give back peer router credits */
1209                 msg->msg_peerrtrcredit = 0;
1210
1211                 spin_lock(&rxpeerni->lpni_lock);
1212                 rxpeerni->lpni_rtrcredits++;
1213                 spin_unlock(&rxpeerni->lpni_lock);
1214
1215                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1216                 spin_lock(&lp->lp_lock);
1217
1218                 /* drop all messages which are queued to be routed on that
1219                  * peer. */
1220                 if (!the_lnet.ln_routing) {
1221                         LIST_HEAD(drop);
1222                         list_splice_init(&lp->lp_rtrq, &drop);
1223                         spin_unlock(&lp->lp_lock);
1224                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1225                 } else if (!list_empty(&lp->lp_rtrq)) {
1226                         int msg2_cpt;
1227
1228                         msg2 = list_entry(lp->lp_rtrq.next,
1229                                           struct lnet_msg, msg_list);
1230                         list_del(&msg2->msg_list);
1231                         msg2_cpt = msg2->msg_rx_cpt;
1232                         spin_unlock(&lp->lp_lock);
1233                         /*
1234                          * messages on the lp_rtrq can be from any NID in
1235                          * the peer, which means they might have different
1236                          * cpts. We need to make sure we lock the right
1237                          * one.
1238                          */
1239                         if (msg2_cpt != msg->msg_rx_cpt) {
1240                                 lnet_net_unlock(msg->msg_rx_cpt);
1241                                 lnet_net_lock(msg2_cpt);
1242                         }
1243                         (void) lnet_post_routed_recv_locked(msg2, 1);
1244                         if (msg2_cpt != msg->msg_rx_cpt) {
1245                                 lnet_net_unlock(msg2_cpt);
1246                                 lnet_net_lock(msg->msg_rx_cpt);
1247                         }
1248                 } else {
1249                         spin_unlock(&lp->lp_lock);
1250                 }
1251         }
1252         if (rxni != NULL) {
1253                 msg->msg_rxni = NULL;
1254                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1255         }
1256         if (rxpeerni != NULL) {
1257                 msg->msg_rxpeer = NULL;
1258                 lnet_peer_ni_decref_locked(rxpeerni);
1259         }
1260 }
1261
1262 static int
1263 lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
1264 {
1265         if (p1->lpni_txqnob < p2->lpni_txqnob)
1266                 return 1;
1267
1268         if (p1->lpni_txqnob > p2->lpni_txqnob)
1269                 return -1;
1270
1271         if (p1->lpni_txcredits > p2->lpni_txcredits)
1272                 return 1;
1273
1274         if (p1->lpni_txcredits < p2->lpni_txcredits)
1275                 return -1;
1276
1277         return 0;
1278 }
1279
1280 static struct lnet_peer_ni *
1281 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1282                     struct lnet_peer *peer,
1283                     struct lnet_peer_ni *best_lpni,
1284                     struct lnet_peer_net *peer_net)
1285 {
1286         /*
1287          * Look at the peer NIs for the destination peer that connect
1288          * to the chosen net. If a peer_ni is preferred when using the
1289          * best_ni to communicate, we use that one. If there is no
1290          * preferred peer_ni, or there are multiple preferred peer_ni,
1291          * the available transmit credits are used. If the transmit
1292          * credits are equal, we round-robin over the peer_ni.
1293          */
1294         struct lnet_peer_ni *lpni = NULL;
1295         int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1296                 INT_MIN;
1297         int best_lpni_healthv = (best_lpni) ?
1298                 atomic_read(&best_lpni->lpni_healthv) : 0;
1299         bool preferred = false;
1300         bool ni_is_pref;
1301         int lpni_healthv;
1302
1303         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1304                 /*
1305                  * if the best_ni we've chosen aleady has this lpni
1306                  * preferred, then let's use it
1307                  */
1308                 if (best_ni) {
1309                         ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
1310                                                                 best_ni->ni_nid);
1311                         CDEBUG(D_NET, "%s ni_is_pref = %d\n",
1312                                libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
1313                 } else {
1314                         ni_is_pref = false;
1315                 }
1316
1317                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1318
1319                 if (best_lpni)
1320                         CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
1321                                 libcfs_nid2str(lpni->lpni_nid),
1322                                 lpni->lpni_txcredits, best_lpni_credits,
1323                                 lpni->lpni_seq, best_lpni->lpni_seq);
1324
1325                 /* pick the healthiest peer ni */
1326                 if (lpni_healthv < best_lpni_healthv) {
1327                         continue;
1328                 } else if (lpni_healthv > best_lpni_healthv) {
1329                         best_lpni_healthv = lpni_healthv;
1330                 /* if this is a preferred peer use it */
1331                 } else if (!preferred && ni_is_pref) {
1332                         preferred = true;
1333                 } else if (preferred && !ni_is_pref) {
1334                         /*
1335                          * this is not the preferred peer so let's ignore
1336                          * it.
1337                          */
1338                         continue;
1339                 } else if (lpni->lpni_txcredits < best_lpni_credits) {
1340                         /*
1341                          * We already have a peer that has more credits
1342                          * available than this one. No need to consider
1343                          * this peer further.
1344                          */
1345                         continue;
1346                 } else if (lpni->lpni_txcredits == best_lpni_credits) {
1347                         /*
1348                          * The best peer found so far and the current peer
1349                          * have the same number of available credits let's
1350                          * make sure to select between them using Round
1351                          * Robin
1352                          */
1353                         if (best_lpni) {
1354                                 if (best_lpni->lpni_seq <= lpni->lpni_seq)
1355                                         continue;
1356                         }
1357                 }
1358
1359                 best_lpni = lpni;
1360                 best_lpni_credits = lpni->lpni_txcredits;
1361         }
1362
1363         /* if we still can't find a peer ni then we can't reach it */
1364         if (!best_lpni) {
1365                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1366                         LNET_NIDNET(dst_nid);
1367                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1368                                 libcfs_net2str(net_id));
1369                 return NULL;
1370         }
1371
1372         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1373                libcfs_nid2str(best_lpni->lpni_nid));
1374
1375         return best_lpni;
1376 }
1377
1378 /*
1379  * Prerequisite: the best_ni should already be set in the sd
1380  * Find the best lpni.
1381  * If the net id is provided then restrict lpni selection on
1382  * that particular net.
1383  * Otherwise find any reachable lpni. When dealing with an MR
1384  * gateway and it has multiple lpnis which we can use
1385  * we want to select the best one from the list of reachable
1386  * ones.
1387  */
1388 static inline struct lnet_peer_ni *
1389 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1390                     struct lnet_peer *peer, __u32 net_id)
1391 {
1392         struct lnet_peer_net *peer_net;
1393         __u32 any_net = LNET_NIDNET(LNET_NID_ANY);
1394
1395         /* find the best_lpni on any local network */
1396         if (net_id == any_net) {
1397                 struct lnet_peer_ni *best_lpni = NULL;
1398                 struct lnet_peer_net *lpn;
1399                 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1400                         /* no net specified find any reachable peer ni */
1401                         if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1402                                 continue;
1403                         best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1404                                                         best_lpni, lpn);
1405                 }
1406
1407                 return best_lpni;
1408         }
1409         /* restrict on the specified net */
1410         peer_net = lnet_peer_get_net_locked(peer, net_id);
1411         if (peer_net)
1412                 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1413
1414         return NULL;
1415 }
1416
1417 /* Compare route priorities and hop counts */
1418 static int
1419 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1420 {
1421         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1422         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1423
1424         if (r1->lr_priority < r2->lr_priority)
1425                 return 1;
1426
1427         if (r1->lr_priority > r2->lr_priority)
1428                 return -1;
1429
1430         if (r1_hops < r2_hops)
1431                 return 1;
1432
1433         if (r1_hops > r2_hops)
1434                 return -1;
1435
1436         return 0;
1437 }
1438
1439 static struct lnet_route *
1440 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1441                        struct lnet_route **prev_route,
1442                        struct lnet_peer_ni **gwni)
1443 {
1444         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1445         struct lnet_route *best_route;
1446         struct lnet_route *last_route;
1447         struct lnet_route *route;
1448         int rc;
1449
1450         CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1451                libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1452
1453         best_route = last_route = NULL;
1454         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1455                 if (!lnet_is_route_alive(route))
1456                         continue;
1457
1458                 /*
1459                  * Restrict the selection of the router NI on the src_net
1460                  * provided. If the src_net is LNET_NID_ANY, then select
1461                  * the best interface available.
1462                  */
1463                 if (!best_route) {
1464                         lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1465                                                    route->lr_gateway,
1466                                                    src_net);
1467                         if (lpni) {
1468                                 best_route = last_route = route;
1469                                 best_gw_ni = lpni;
1470                         } else {
1471                                 CDEBUG(D_NET, "Gateway %s does not have a peer NI on net %s\n",
1472                                        libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1473                                        libcfs_net2str(src_net));
1474                         }
1475
1476                         continue;
1477                 }
1478
1479                 /* no protection on below fields, but it's harmless */
1480                 if (last_route->lr_seq - route->lr_seq < 0)
1481                         last_route = route;
1482
1483                 rc = lnet_compare_routes(route, best_route);
1484                 if (rc == -1)
1485                         continue;
1486
1487                 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1488                                            route->lr_gateway,
1489                                            src_net);
1490                 /* restrict the lpni on the src_net if specified */
1491                 if (!lpni) {
1492                         CDEBUG(D_NET, "Gateway %s does not have a peer NI on net %s\n",
1493                                libcfs_nid2str(route->lr_gateway->lp_primary_nid),
1494                                libcfs_net2str(src_net));
1495                         continue;
1496                 }
1497
1498                 if (rc == 1) {
1499                         best_route = route;
1500                         best_gw_ni = lpni;
1501                         continue;
1502                 }
1503
1504                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1505                 if (rc == -1)
1506                         continue;
1507
1508                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1509                         best_route = route;
1510                         best_gw_ni = lpni;
1511                         continue;
1512                 }
1513         }
1514
1515         *prev_route = last_route;
1516         *gwni = best_gw_ni;
1517
1518         return best_route;
1519 }
1520
1521 static struct lnet_ni *
1522 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1523                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1524                  int md_cpt)
1525 {
1526         struct lnet_ni *ni = NULL;
1527         unsigned int shortest_distance;
1528         int best_credits;
1529         int best_healthv;
1530
1531         /*
1532          * If there is no peer_ni that we can send to on this network,
1533          * then there is no point in looking for a new best_ni here.
1534         */
1535         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1536                 return best_ni;
1537
1538         if (best_ni == NULL) {
1539                 shortest_distance = UINT_MAX;
1540                 best_credits = INT_MIN;
1541                 best_healthv = 0;
1542         } else {
1543                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1544                                                      best_ni->ni_dev_cpt);
1545                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1546                 best_healthv = atomic_read(&best_ni->ni_healthv);
1547         }
1548
1549         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1550                 unsigned int distance;
1551                 int ni_credits;
1552                 int ni_healthv;
1553                 int ni_fatal;
1554
1555                 ni_credits = atomic_read(&ni->ni_tx_credits);
1556                 ni_healthv = atomic_read(&ni->ni_healthv);
1557                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1558
1559                 /*
1560                  * calculate the distance from the CPT on which
1561                  * the message memory is allocated to the CPT of
1562                  * the NI's physical device
1563                  */
1564                 distance = cfs_cpt_distance(lnet_cpt_table(),
1565                                             md_cpt,
1566                                             ni->ni_dev_cpt);
1567
1568                 CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
1569                        libcfs_nid2str(ni->ni_nid), ni_credits, distance,
1570                        ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
1571                         : "not seleced", best_credits, shortest_distance,
1572                         (best_ni) ? best_ni->ni_seq : 0);
1573
1574                 /*
1575                  * All distances smaller than the NUMA range
1576                  * are treated equally.
1577                  */
1578                 if (distance < lnet_numa_range)
1579                         distance = lnet_numa_range;
1580
1581                 /*
1582                  * Select on health, shorter distance, available
1583                  * credits, then round-robin.
1584                  */
1585                 if (ni_fatal) {
1586                         continue;
1587                 } else if (ni_healthv < best_healthv) {
1588                         continue;
1589                 } else if (ni_healthv > best_healthv) {
1590                         best_healthv = ni_healthv;
1591                         /*
1592                          * If we're going to prefer this ni because it's
1593                          * the healthiest, then we should set the
1594                          * shortest_distance in the algorithm in case
1595                          * there are multiple NIs with the same health but
1596                          * different distances.
1597                          */
1598                         if (distance < shortest_distance)
1599                                 shortest_distance = distance;
1600                 } else if (distance > shortest_distance) {
1601                         continue;
1602                 } else if (distance < shortest_distance) {
1603                         shortest_distance = distance;
1604                 } else if (ni_credits < best_credits) {
1605                         continue;
1606                 } else if (ni_credits == best_credits) {
1607                         if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1608                                 continue;
1609                 }
1610                 best_ni = ni;
1611                 best_credits = ni_credits;
1612         }
1613
1614         CDEBUG(D_NET, "selected best_ni %s\n",
1615                (best_ni) ? libcfs_nid2str(best_ni->ni_nid) : "no selection");
1616
1617         return best_ni;
1618 }
1619
1620 /*
1621  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1622  * because such traffic is required to perform discovery. We therefore
1623  * exclude all GET and PUT on that portal. We also exclude all ACK and
1624  * REPLY traffic, but that is because the portal is not tracked in the
1625  * message structure for these message types. We could restrict this
1626  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1627  */
1628 static bool
1629 lnet_msg_discovery(struct lnet_msg *msg)
1630 {
1631         if (msg->msg_type == LNET_MSG_PUT) {
1632                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1633                         return true;
1634         } else if (msg->msg_type == LNET_MSG_GET) {
1635                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1636                         return true;
1637         }
1638         return false;
1639 }
1640
1641 #define SRC_SPEC        0x0001
1642 #define SRC_ANY         0x0002
1643 #define LOCAL_DST       0x0004
1644 #define REMOTE_DST      0x0008
1645 #define MR_DST          0x0010
1646 #define NMR_DST         0x0020
1647 #define SND_RESP        0x0040
1648
1649 /* The following to defines are used for return codes */
1650 #define REPEAT_SEND     0x1000
1651 #define PASS_THROUGH    0x2000
1652
1653 /* The different cases lnet_select pathway needs to handle */
1654 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1655 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1656 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1657 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1658 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1659 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1660 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1661 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1662
1663 static int
1664 lnet_handle_lo_send(struct lnet_send_data *sd)
1665 {
1666         struct lnet_msg *msg = sd->sd_msg;
1667         int cpt = sd->sd_cpt;
1668
1669         /* No send credit hassles with LOLND */
1670         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1671         msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
1672         if (!msg->msg_routing)
1673                 msg->msg_hdr.src_nid =
1674                         cpu_to_le64(the_lnet.ln_loni->ni_nid);
1675         msg->msg_target.nid = the_lnet.ln_loni->ni_nid;
1676         lnet_msg_commit(msg, cpt);
1677         msg->msg_txni = the_lnet.ln_loni;
1678
1679         return LNET_CREDIT_OK;
1680 }
1681
1682 static int
1683 lnet_handle_send(struct lnet_send_data *sd)
1684 {
1685         struct lnet_ni *best_ni = sd->sd_best_ni;
1686         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1687         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1688         struct lnet_msg *msg = sd->sd_msg;
1689         int cpt2;
1690         __u32 send_case = sd->sd_send_case;
1691         int rc;
1692         __u32 routing = send_case & REMOTE_DST;
1693          struct lnet_rsp_tracker *rspt;
1694
1695         /*
1696          * Increment sequence number of the selected peer so that we
1697          * pick the next one in Round Robin.
1698          */
1699         best_lpni->lpni_seq++;
1700
1701         /*
1702          * grab a reference on the peer_ni so it sticks around even if
1703          * we need to drop and relock the lnet_net_lock below.
1704          */
1705         lnet_peer_ni_addref_locked(best_lpni);
1706
1707         /*
1708          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1709          * message. This ensures that we get a CPT that is correct for
1710          * the NI when the NI has been restricted to a subset of all CPTs.
1711          * If the selected CPT differs from the one currently locked, we
1712          * must unlock and relock the lnet_net_lock(), and then check whether
1713          * the configuration has changed. We don't have a hold on the best_ni
1714          * yet, and it may have vanished.
1715          */
1716         cpt2 = lnet_cpt_of_nid_locked(best_lpni->lpni_nid, best_ni);
1717         if (sd->sd_cpt != cpt2) {
1718                 __u32 seq = lnet_get_dlc_seq_locked();
1719                 lnet_net_unlock(sd->sd_cpt);
1720                 sd->sd_cpt = cpt2;
1721                 lnet_net_lock(sd->sd_cpt);
1722                 if (seq != lnet_get_dlc_seq_locked()) {
1723                         lnet_peer_ni_decref_locked(best_lpni);
1724                         return REPEAT_SEND;
1725                 }
1726         }
1727
1728         /*
1729          * store the best_lpni in the message right away to avoid having
1730          * to do the same operation under different conditions
1731          */
1732         msg->msg_txpeer = best_lpni;
1733         msg->msg_txni = best_ni;
1734
1735         /*
1736          * grab a reference for the best_ni since now it's in use in this
1737          * send. The reference will be dropped in lnet_finalize()
1738          */
1739         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1740
1741         /*
1742          * Always set the target.nid to the best peer picked. Either the
1743          * NID will be one of the peer NIDs selected, or the same NID as
1744          * what was originally set in the target or it will be the NID of
1745          * a router if this message should be routed
1746          */
1747         msg->msg_target.nid = msg->msg_txpeer->lpni_nid;
1748
1749         /*
1750          * lnet_msg_commit assigns the correct cpt to the message, which
1751          * is used to decrement the correct refcount on the ni when it's
1752          * time to return the credits
1753          */
1754         lnet_msg_commit(msg, sd->sd_cpt);
1755
1756         /*
1757          * If we are routing the message then we keep the src_nid that was
1758          * set by the originator. If we are not routing then we are the
1759          * originator and set it here.
1760          */
1761         if (!msg->msg_routing)
1762                 msg->msg_hdr.src_nid = cpu_to_le64(msg->msg_txni->ni_nid);
1763
1764         if (routing) {
1765                 msg->msg_target_is_router = 1;
1766                 msg->msg_target.pid = LNET_PID_LUSTRE;
1767                 /*
1768                  * since we're routing we want to ensure that the
1769                  * msg_hdr.dest_nid is set to the final destination. When
1770                  * the router receives this message it knows how to route
1771                  * it.
1772                  *
1773                  * final_dst_lpni is set at the beginning of the
1774                  * lnet_select_pathway() function and is never changed.
1775                  * It's safe to use it here.
1776                  */
1777                 msg->msg_hdr.dest_nid = cpu_to_le64(final_dst_lpni->lpni_nid);
1778         } else {
1779                 /*
1780                  * if we're not routing set the dest_nid to the best peer
1781                  * ni NID that we picked earlier in the algorithm.
1782                  */
1783                 msg->msg_hdr.dest_nid = cpu_to_le64(msg->msg_txpeer->lpni_nid);
1784         }
1785
1786         /*
1787          * if we have response tracker block update it with the next hop
1788          * nid
1789          */
1790         if (msg->msg_md) {
1791                 rspt = msg->msg_md->md_rspt_ptr;
1792                 if (rspt) {
1793                         rspt->rspt_next_hop_nid = msg->msg_txpeer->lpni_nid;
1794                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1795                                libcfs_nid2str(rspt->rspt_next_hop_nid));
1796                 }
1797         }
1798
1799         rc = lnet_post_send_locked(msg, 0);
1800
1801         if (!rc)
1802                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1803                        libcfs_nid2str(msg->msg_hdr.src_nid),
1804                        libcfs_nid2str(msg->msg_txni->ni_nid),
1805                        libcfs_nid2str(sd->sd_src_nid),
1806                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1807                        libcfs_nid2str(sd->sd_dst_nid),
1808                        libcfs_nid2str(msg->msg_txpeer->lpni_nid),
1809                        libcfs_nid2str(sd->sd_rtr_nid),
1810                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1811
1812         return rc;
1813 }
1814
1815 static inline void
1816 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1817                          struct lnet_msg *msg)
1818 {
1819         if (!lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1820                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1821                        libcfs_nid2str(lni->ni_nid),
1822                        libcfs_nid2str(lpni->lpni_nid));
1823                 lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
1824         }
1825 }
1826
1827 /*
1828  * Source Specified
1829  * Local Destination
1830  * non-mr peer
1831  *
1832  * use the source and destination NIDs as the pathway
1833  */
1834 static int
1835 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
1836 {
1837         /* the destination lpni is set before we get here. */
1838
1839         /* find local NI */
1840         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1841         if (!sd->sd_best_ni) {
1842                 CERROR("Can't send to %s: src %s is not a "
1843                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1844                                 libcfs_nid2str(sd->sd_src_nid));
1845                 return -EINVAL;
1846         }
1847
1848         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
1849
1850         return lnet_handle_send(sd);
1851 }
1852
1853 /*
1854  * Source Specified
1855  * Local Destination
1856  * MR Peer
1857  *
1858  * Don't run the selection algorithm on the peer NIs. By specifying the
1859  * local NID, we're also saying that we should always use the destination NID
1860  * provided. This handles the case where we should be using the same
1861  * destination NID for the all the messages which belong to the same RPC
1862  * request.
1863  */
1864 static int
1865 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
1866 {
1867         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
1868         if (!sd->sd_best_ni) {
1869                 CERROR("Can't send to %s: src %s is not a "
1870                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
1871                                 libcfs_nid2str(sd->sd_src_nid));
1872                 return -EINVAL;
1873         }
1874
1875         if (sd->sd_best_lpni &&
1876             sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
1877                 return lnet_handle_lo_send(sd);
1878         else if (sd->sd_best_lpni)
1879                 return lnet_handle_send(sd);
1880
1881         CERROR("can't send to %s. no NI on %s\n",
1882                libcfs_nid2str(sd->sd_dst_nid),
1883                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
1884
1885         return -EHOSTUNREACH;
1886 }
1887
1888 struct lnet_ni *
1889 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
1890                               struct lnet_peer *peer,
1891                               struct lnet_peer_net *peer_net,
1892                               int cpt,
1893                               bool incr_seq)
1894 {
1895         struct lnet_net *local_net;
1896         struct lnet_ni *best_ni;
1897
1898         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
1899         if (!local_net)
1900                 return NULL;
1901
1902         /*
1903          * Iterate through the NIs in this local Net and select
1904          * the NI to send from. The selection is determined by
1905          * these 3 criterion in the following priority:
1906          *      1. NUMA
1907          *      2. NI available credits
1908          *      3. Round Robin
1909          */
1910         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
1911                                    peer, peer_net, cpt);
1912
1913         if (incr_seq && best_ni)
1914                 best_ni->ni_seq++;
1915
1916         return best_ni;
1917 }
1918
1919 static int
1920 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
1921                              int cpt)
1922 {
1923         struct lnet_peer *peer;
1924         int rc;
1925
1926         lnet_peer_ni_addref_locked(lpni);
1927
1928         peer = lpni->lpni_peer_net->lpn_peer;
1929
1930         if (lnet_peer_gw_discovery(peer)) {
1931                 lnet_peer_ni_decref_locked(lpni);
1932                 return 0;
1933         }
1934
1935         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
1936                 lnet_peer_ni_decref_locked(lpni);
1937                 return 0;
1938         }
1939
1940         rc = lnet_discover_peer_locked(lpni, cpt, false);
1941         if (rc) {
1942                 lnet_peer_ni_decref_locked(lpni);
1943                 return rc;
1944         }
1945         /* The peer may have changed. */
1946         peer = lpni->lpni_peer_net->lpn_peer;
1947         spin_lock(&peer->lp_lock);
1948         if (lnet_peer_is_uptodate_locked(peer)) {
1949                 spin_unlock(&peer->lp_lock);
1950                 lnet_peer_ni_decref_locked(lpni);
1951                 return 0;
1952         }
1953         /* queue message and return */
1954         msg->msg_sending = 0;
1955         msg->msg_txpeer = NULL;
1956         list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
1957         spin_unlock(&peer->lp_lock);
1958
1959         lnet_peer_ni_decref_locked(lpni);
1960
1961         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
1962                msg, libcfs_nid2str(peer->lp_primary_nid));
1963
1964         return LNET_DC_WAIT;
1965 }
1966
1967 static int
1968 lnet_handle_find_routed_path(struct lnet_send_data *sd,
1969                              lnet_nid_t dst_nid,
1970                              struct lnet_peer_ni **gw_lpni,
1971                              struct lnet_peer **gw_peer)
1972 {
1973         int rc;
1974         __u32 local_lnet;
1975         struct lnet_peer *gw;
1976         struct lnet_peer *lp;
1977         struct lnet_peer_net *lpn;
1978         struct lnet_peer_net *best_lpn = NULL;
1979         struct lnet_remotenet *rnet, *best_rnet = NULL;
1980         struct lnet_route *best_route = NULL;
1981         struct lnet_route *last_route = NULL;
1982         struct lnet_peer_ni *lpni = NULL;
1983         struct lnet_peer_ni *gwni = NULL;
1984         bool route_found = false;
1985         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
1986                 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
1987                 LNET_NID_ANY;
1988
1989         CDEBUG(D_NET, "using src nid %s for route restriction\n",
1990                libcfs_nid2str(src_nid));
1991
1992         /* If a router nid was specified then we are replying to a GET or
1993          * sending an ACK. In this case we use the gateway associated with the
1994          * specified router nid.
1995          */
1996         if (sd->sd_rtr_nid != LNET_NID_ANY) {
1997                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
1998                 if (gwni) {
1999                         gw = gwni->lpni_peer_net->lpn_peer;
2000                         lnet_peer_ni_decref_locked(gwni);
2001                         if (gw->lp_rtr_refcount) {
2002                                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2003                                 route_found = true;
2004                         }
2005                 } else {
2006                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2007                                libcfs_nid2str(sd->sd_rtr_nid));
2008                 }
2009         }
2010
2011         if (!route_found) {
2012                 /* we've already looked up the initial lpni using dst_nid */
2013                 lpni = sd->sd_best_lpni;
2014                 /* the peer tree must be in existence */
2015                 LASSERT(lpni && lpni->lpni_peer_net &&
2016                         lpni->lpni_peer_net->lpn_peer);
2017                 lp = lpni->lpni_peer_net->lpn_peer;
2018
2019                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2020                         /* is this remote network reachable?  */
2021                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2022                         if (!rnet)
2023                                 continue;
2024
2025                         if (!best_lpn) {
2026                                 best_lpn = lpn;
2027                                 best_rnet = rnet;
2028                         }
2029
2030                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2031                                 continue;
2032
2033                         best_lpn = lpn;
2034                         best_rnet = rnet;
2035                 }
2036
2037                 if (!best_lpn) {
2038                         CERROR("peer %s has no available nets\n",
2039                                libcfs_nid2str(sd->sd_dst_nid));
2040                         return -EHOSTUNREACH;
2041                 }
2042
2043                 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2044                                                        sd->sd_dst_nid,
2045                                                        lp,
2046                                                        best_lpn->lpn_net_id);
2047                 if (!sd->sd_best_lpni) {
2048                         CERROR("peer %s is unreachable\n",
2049                                libcfs_nid2str(sd->sd_dst_nid));
2050                         return -EHOSTUNREACH;
2051                 }
2052
2053                 /*
2054                  * We're attempting to round robin over the remote peer
2055                  * NI's so update the final destination we selected
2056                  */
2057                 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2058
2059                 /*
2060                  * find the best route. Restrict the selection on the net of the
2061                  * local NI if we've already picked the local NI to send from.
2062                  * Otherwise, let's pick any route we can find and then find
2063                  * a local NI we can reach the route's gateway on. Any route we select
2064                  * will be reachable by virtue of the restriction we have when
2065                  * adding a route.
2066                  */
2067                 best_route = lnet_find_route_locked(best_rnet,
2068                                                     LNET_NIDNET(src_nid),
2069                                                     &last_route, &gwni);
2070
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                  * Increment the sequence number of the remote lpni so we
2091                  * can round robin over the different interfaces of the
2092                  * remote lpni
2093                  */
2094                 sd->sd_best_lpni->lpni_seq++;
2095         }
2096
2097         /*
2098          * Discover this gateway if it hasn't already been discovered.
2099          * This means we might delay the message until discovery has
2100          * completed
2101          */
2102         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2103         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2104         if (rc)
2105                 return rc;
2106
2107         if (!sd->sd_best_ni)
2108                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2109                                         lnet_peer_get_net_locked(gw,
2110                                                                  local_lnet),
2111                                         sd->sd_md_cpt,
2112                                         true);
2113
2114         if (!sd->sd_best_ni) {
2115                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2116                        libcfs_net2str(local_lnet),
2117                        libcfs_nid2str(sd->sd_src_nid));
2118                 return -EFAULT;
2119         }
2120
2121         *gw_lpni = gwni;
2122         *gw_peer = gw;
2123
2124         /*
2125          * increment the sequence numbers since now we're sure we're
2126          * going to use this path
2127          */
2128         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2129                 LASSERT(best_route && last_route);
2130                 best_route->lr_seq = last_route->lr_seq + 1;
2131                 best_lpn->lpn_seq++;
2132         }
2133
2134         return 0;
2135 }
2136
2137 /*
2138  * Handle two cases:
2139  *
2140  * Case 1:
2141  *  Source specified
2142  *  Remote destination
2143  *  Non-MR destination
2144  *
2145  * Case 2:
2146  *  Source specified
2147  *  Remote destination
2148  *  MR destination
2149  *
2150  * The handling of these two cases is similar. Even though the destination
2151  * can be MR or non-MR, we'll deal directly with the router.
2152  */
2153 static int
2154 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2155 {
2156         int rc;
2157         struct lnet_peer_ni *gw_lpni = NULL;
2158         struct lnet_peer *gw_peer = NULL;
2159
2160         /* find local NI */
2161         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2162         if (!sd->sd_best_ni) {
2163                 CERROR("Can't send to %s: src %s is not a "
2164                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2165                                 libcfs_nid2str(sd->sd_src_nid));
2166                 return -EINVAL;
2167         }
2168
2169         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2170                                      &gw_peer);
2171         if (rc)
2172                 return rc;
2173
2174         if (sd->sd_send_case & NMR_DST)
2175                 /*
2176                  * since the final destination is non-MR let's set its preferred
2177                  * NID before we send
2178                  */
2179                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2180                                          sd->sd_msg);
2181
2182         /*
2183          * We're going to send to the gw found so let's set its
2184          * info
2185          */
2186         sd->sd_peer = gw_peer;
2187         sd->sd_best_lpni = gw_lpni;
2188
2189         return lnet_handle_send(sd);
2190 }
2191
2192 struct lnet_ni *
2193 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2194                                bool discovery)
2195 {
2196         struct lnet_peer_net *peer_net = NULL;
2197         struct lnet_ni *best_ni = NULL;
2198         int lpn_healthv = 0;
2199
2200         /*
2201          * The peer can have multiple interfaces, some of them can be on
2202          * the local network and others on a routed network. We should
2203          * prefer the local network. However if the local network is not
2204          * available then we need to try the routed network
2205          */
2206
2207         /* go through all the peer nets and find the best_ni */
2208         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2209                 /*
2210                  * The peer's list of nets can contain non-local nets. We
2211                  * want to only examine the local ones.
2212                  */
2213                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2214                         continue;
2215
2216                 /* always select the lpn with the best health */
2217                 if (lpn_healthv <= peer_net->lpn_healthv)
2218                         lpn_healthv = peer_net->lpn_healthv;
2219                 else
2220                         continue;
2221
2222                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2223                                                         md_cpt, false);
2224
2225                 /*
2226                  * if this is a discovery message and lp_disc_net_id is
2227                  * specified then use that net to send the discovery on.
2228                  */
2229                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2230                     discovery)
2231                         break;
2232         }
2233
2234         if (best_ni)
2235                 /* increment sequence number so we can round robin */
2236                 best_ni->ni_seq++;
2237
2238         return best_ni;
2239 }
2240
2241 static struct lnet_ni *
2242 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2243 {
2244         struct lnet_ni *best_ni = NULL;
2245         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2246         struct lnet_peer_ni *lpni_entry;
2247
2248         /*
2249          * We must use a consistent source address when sending to a
2250          * non-MR peer. However, a non-MR peer can have multiple NIDs
2251          * on multiple networks, and we may even need to talk to this
2252          * peer on multiple networks -- certain types of
2253          * load-balancing configuration do this.
2254          *
2255          * So we need to pick the NI the peer prefers for this
2256          * particular network.
2257          */
2258         LASSERT(peer_net);
2259         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2260                             lpni_peer_nis) {
2261                 if (lpni_entry->lpni_pref_nnids == 0)
2262                         continue;
2263                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2264                 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2265                 break;
2266         }
2267
2268         return best_ni;
2269 }
2270
2271 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2272 static int
2273 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2274 {
2275         struct lnet_ni *best_ni = NULL;
2276         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2277
2278         /*
2279          * We must use a consistent source address when sending to a
2280          * non-MR peer. However, a non-MR peer can have multiple NIDs
2281          * on multiple networks, and we may even need to talk to this
2282          * peer on multiple networks -- certain types of
2283          * load-balancing configuration do this.
2284          *
2285          * So we need to pick the NI the peer prefers for this
2286          * particular network.
2287          */
2288
2289         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2290                                                        sd->sd_cpt);
2291
2292         /* if best_ni is still not set just pick one */
2293         if (!best_ni) {
2294                 best_ni =
2295                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2296                                                 sd->sd_best_lpni->lpni_peer_net,
2297                                                 sd->sd_md_cpt, true);
2298                 /* If there is no best_ni we don't have a route */
2299                 if (!best_ni) {
2300                         CERROR("no path to %s from net %s\n",
2301                                 libcfs_nid2str(best_lpni->lpni_nid),
2302                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2303                         return -EHOSTUNREACH;
2304                 }
2305         }
2306
2307         sd->sd_best_ni = best_ni;
2308
2309         /* Set preferred NI if necessary. */
2310         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2311
2312         return 0;
2313 }
2314
2315
2316 /*
2317  * Source not specified
2318  * Local destination
2319  * Non-MR Peer
2320  *
2321  * always use the same source NID for NMR peers
2322  * If we've talked to that peer before then we already have a preferred
2323  * source NI associated with it. Otherwise, we select a preferred local NI
2324  * and store it in the peer
2325  */
2326 static int
2327 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2328 {
2329         int rc = 0;
2330
2331         /* sd->sd_best_lpni is already set to the final destination */
2332
2333         /*
2334          * At this point we should've created the peer ni and peer. If we
2335          * can't find it, then something went wrong. Instead of assert
2336          * output a relevant message and fail the send
2337          */
2338         if (!sd->sd_best_lpni) {
2339                 CERROR("Internal fault. Unable to send msg %s to %s. "
2340                        "NID not known\n",
2341                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2342                        libcfs_nid2str(sd->sd_dst_nid));
2343                 return -EFAULT;
2344         }
2345
2346         if (sd->sd_msg->msg_routing) {
2347                 /* If I'm forwarding this message then I can choose any NI
2348                  * on the destination peer net
2349                  */
2350                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2351                                                                sd->sd_peer,
2352                                                                sd->sd_best_lpni->lpni_peer_net,
2353                                                                sd->sd_md_cpt,
2354                                                                true);
2355                 if (!sd->sd_best_ni) {
2356                         CERROR("Unable to forward message to %s. No local NI available\n",
2357                                libcfs_nid2str(sd->sd_dst_nid));
2358                         rc = -EHOSTUNREACH;
2359                 }
2360         } else
2361                 rc = lnet_select_preferred_best_ni(sd);
2362
2363         if (!rc)
2364                 rc = lnet_handle_send(sd);
2365
2366         return rc;
2367 }
2368
2369 static int
2370 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2371 {
2372         /*
2373          * NOTE we've already handled the remote peer case. So we only
2374          * need to worry about the local case here.
2375          *
2376          * if we're sending a response, ACK or reply, we need to send it
2377          * to the destination NID given to us. At this point we already
2378          * have the peer_ni we're suppose to send to, so just find the
2379          * best_ni on the peer net and use that. Since we're sending to an
2380          * MR peer then we can just run the selection algorithm on our
2381          * local NIs and pick the best one.
2382          */
2383         if (sd->sd_send_case & SND_RESP) {
2384                 sd->sd_best_ni =
2385                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2386                                                 sd->sd_best_lpni->lpni_peer_net,
2387                                                 sd->sd_md_cpt, true);
2388
2389                 if (!sd->sd_best_ni) {
2390                         /*
2391                          * We're not going to deal with not able to send
2392                          * a response to the provided final destination
2393                          */
2394                         CERROR("Can't send response to %s. "
2395                                "No local NI available\n",
2396                                 libcfs_nid2str(sd->sd_dst_nid));
2397                         return -EHOSTUNREACH;
2398                 }
2399
2400                 return lnet_handle_send(sd);
2401         }
2402
2403         /*
2404          * If we get here that means we're sending a fresh request, PUT or
2405          * GET, so we need to run our standard selection algorithm.
2406          * First find the best local interface that's on any of the peer's
2407          * networks.
2408          */
2409         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2410                                         sd->sd_md_cpt,
2411                                         lnet_msg_discovery(sd->sd_msg));
2412         if (sd->sd_best_ni) {
2413                 sd->sd_best_lpni =
2414                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2415                                       sd->sd_peer,
2416                                       sd->sd_best_ni->ni_net->net_id);
2417
2418                 /*
2419                  * if we're successful in selecting a peer_ni on the local
2420                  * network, then send to it. Otherwise fall through and
2421                  * try and see if we can reach it over another routed
2422                  * network
2423                  */
2424                 if (sd->sd_best_lpni &&
2425                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2426                         /*
2427                          * in case we initially started with a routed
2428                          * destination, let's reset to local
2429                          */
2430                         sd->sd_send_case &= ~REMOTE_DST;
2431                         sd->sd_send_case |= LOCAL_DST;
2432                         return lnet_handle_lo_send(sd);
2433                 } else if (sd->sd_best_lpni) {
2434                         /*
2435                          * in case we initially started with a routed
2436                          * destination, let's reset to local
2437                          */
2438                         sd->sd_send_case &= ~REMOTE_DST;
2439                         sd->sd_send_case |= LOCAL_DST;
2440                         return lnet_handle_send(sd);
2441                 }
2442
2443                 CERROR("Internal Error. Expected to have a best_lpni: "
2444                        "%s -> %s\n",
2445                        libcfs_nid2str(sd->sd_src_nid),
2446                        libcfs_nid2str(sd->sd_dst_nid));
2447
2448                 return -EFAULT;
2449         }
2450
2451         /*
2452          * Peer doesn't have a local network. Let's see if there is
2453          * a remote network we can reach it on.
2454          */
2455         return PASS_THROUGH;
2456 }
2457
2458 /*
2459  * Case 1:
2460  *      Source NID not specified
2461  *      Local destination
2462  *      MR peer
2463  *
2464  * Case 2:
2465  *      Source NID not speified
2466  *      Remote destination
2467  *      MR peer
2468  *
2469  * In both of these cases if we're sending a response, ACK or REPLY, then
2470  * we need to send to the destination NID provided.
2471  *
2472  * In the remote case let's deal with MR routers.
2473  *
2474  */
2475
2476 static int
2477 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2478 {
2479         int rc = 0;
2480         struct lnet_peer *gw_peer = NULL;
2481         struct lnet_peer_ni *gw_lpni = NULL;
2482
2483         /*
2484          * handle sending a response to a remote peer here so we don't
2485          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2486          */
2487         if (sd->sd_send_case & REMOTE_DST &&
2488             sd->sd_send_case & SND_RESP) {
2489                 struct lnet_peer_ni *gw;
2490                 struct lnet_peer *gw_peer;
2491
2492                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2493                                                   &gw_peer);
2494                 if (rc < 0) {
2495                         CERROR("Can't send response to %s. "
2496                                "No route available\n",
2497                                 libcfs_nid2str(sd->sd_dst_nid));
2498                         return -EHOSTUNREACH;
2499                 } else if (rc > 0) {
2500                         return rc;
2501                 }
2502
2503                 sd->sd_best_lpni = gw;
2504                 sd->sd_peer = gw_peer;
2505
2506                 return lnet_handle_send(sd);
2507         }
2508
2509         /*
2510          * Even though the NID for the peer might not be on a local network,
2511          * since the peer is MR there could be other interfaces on the
2512          * local network. In that case we'd still like to prefer the local
2513          * network over the routed network. If we're unable to do that
2514          * then we select the best router among the different routed networks,
2515          * and if the router is MR then we can deal with it as such.
2516          */
2517         rc = lnet_handle_any_mr_dsta(sd);
2518         if (rc != PASS_THROUGH)
2519                 return rc;
2520
2521         /*
2522          * Now that we must route to the destination, we must consider the
2523          * MR case, where the destination has multiple interfaces, some of
2524          * which we can route to and others we do not. For this reason we
2525          * need to select the destination which we can route to and if
2526          * there are multiple, we need to round robin.
2527          */
2528         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2529                                           &gw_peer);
2530         if (rc)
2531                 return rc;
2532
2533         sd->sd_send_case &= ~LOCAL_DST;
2534         sd->sd_send_case |= REMOTE_DST;
2535
2536         sd->sd_peer = gw_peer;
2537         sd->sd_best_lpni = gw_lpni;
2538
2539         return lnet_handle_send(sd);
2540 }
2541
2542 /*
2543  * Source not specified
2544  * Remote destination
2545  * Non-MR peer
2546  *
2547  * Must send to the specified peer NID using the same source NID that
2548  * we've used before. If it's the first time to talk to that peer then
2549  * find the source NI and assign it as preferred to that peer
2550  */
2551 static int
2552 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2553 {
2554         int rc;
2555         struct lnet_peer_ni *gw_lpni = NULL;
2556         struct lnet_peer *gw_peer = NULL;
2557
2558         /*
2559          * Let's see if we have a preferred NI to talk to this NMR peer
2560          */
2561         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2562                                                               sd->sd_cpt);
2563
2564         /*
2565          * find the router and that'll find the best NI if we didn't find
2566          * it already.
2567          */
2568         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2569                                           &gw_peer);
2570         if (rc)
2571                 return rc;
2572
2573         /*
2574          * set the best_ni we've chosen as the preferred one for
2575          * this peer
2576          */
2577         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2578
2579         /* we'll be sending to the gw */
2580         sd->sd_best_lpni = gw_lpni;
2581         sd->sd_peer = gw_peer;
2582
2583         return lnet_handle_send(sd);
2584 }
2585
2586 static int
2587 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2588 {
2589         /*
2590          * turn off the SND_RESP bit.
2591          * It will be checked in the case handling
2592          */
2593         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2594
2595         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2596                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2597                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2598                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2599                 libcfs_nid2str(sd->sd_dst_nid),
2600                 (send_case & LOCAL_DST) ? "local" : "routed");
2601
2602         switch (send_case) {
2603         /*
2604          * For all cases where the source is specified, we should always
2605          * use the destination NID, whether it's an MR destination or not,
2606          * since we're continuing a series of related messages for the
2607          * same RPC
2608          */
2609         case SRC_SPEC_LOCAL_NMR_DST:
2610                 return lnet_handle_spec_local_nmr_dst(sd);
2611         case SRC_SPEC_LOCAL_MR_DST:
2612                 return lnet_handle_spec_local_mr_dst(sd);
2613         case SRC_SPEC_ROUTER_NMR_DST:
2614         case SRC_SPEC_ROUTER_MR_DST:
2615                 return lnet_handle_spec_router_dst(sd);
2616         case SRC_ANY_LOCAL_NMR_DST:
2617                 return lnet_handle_any_local_nmr_dst(sd);
2618         case SRC_ANY_LOCAL_MR_DST:
2619         case SRC_ANY_ROUTER_MR_DST:
2620                 return lnet_handle_any_mr_dst(sd);
2621         case SRC_ANY_ROUTER_NMR_DST:
2622                 return lnet_handle_any_router_nmr_dst(sd);
2623         default:
2624                 CERROR("Unknown send case\n");
2625                 return -1;
2626         }
2627 }
2628
2629 static int
2630 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2631                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2632 {
2633         struct lnet_peer_ni *lpni;
2634         struct lnet_peer *peer;
2635         struct lnet_send_data send_data;
2636         int cpt, rc;
2637         int md_cpt;
2638         __u32 send_case = 0;
2639         bool final_hop;
2640         bool mr_forwarding_allowed;
2641
2642         memset(&send_data, 0, sizeof(send_data));
2643
2644         /*
2645          * get an initial CPT to use for locking. The idea here is not to
2646          * serialize the calls to select_pathway, so that as many
2647          * operations can run concurrently as possible. To do that we use
2648          * the CPT where this call is being executed. Later on when we
2649          * determine the CPT to use in lnet_message_commit, we switch the
2650          * lock and check if there was any configuration change.  If none,
2651          * then we proceed, if there is, then we restart the operation.
2652          */
2653         cpt = lnet_net_lock_current();
2654
2655         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2656         if (md_cpt == CFS_CPT_ANY)
2657                 md_cpt = cpt;
2658
2659 again:
2660
2661         /*
2662          * If we're being asked to send to the loopback interface, there
2663          * is no need to go through any selection. We can just shortcut
2664          * the entire process and send over lolnd
2665          */
2666         send_data.sd_msg = msg;
2667         send_data.sd_cpt = cpt;
2668         if (dst_nid == LNET_NID_LO_0) {
2669                 rc = lnet_handle_lo_send(&send_data);
2670                 lnet_net_unlock(cpt);
2671                 return rc;
2672         }
2673
2674         /*
2675          * find an existing peer_ni, or create one and mark it as having been
2676          * created due to network traffic. This call will create the
2677          * peer->peer_net->peer_ni tree.
2678          */
2679         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2680         if (IS_ERR(lpni)) {
2681                 lnet_net_unlock(cpt);
2682                 return PTR_ERR(lpni);
2683         }
2684
2685         /*
2686          * Cache the original src_nid and rtr_nid. If we need to resend the
2687          * message then we'll need to know whether the src_nid was originally
2688          * specified for this message. If it was originally specified,
2689          * then we need to keep using the same src_nid since it's
2690          * continuing the same sequence of messages. Similarly, rtr_nid will
2691          * affect our choice of next hop.
2692          */
2693         msg->msg_src_nid_param = src_nid;
2694         msg->msg_rtr_nid_param = rtr_nid;
2695
2696         /*
2697          * If necessary, perform discovery on the peer that owns this peer_ni.
2698          * Note, this can result in the ownership of this peer_ni changing
2699          * to another peer object.
2700          */
2701         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2702         if (rc) {
2703                 lnet_peer_ni_decref_locked(lpni);
2704                 lnet_net_unlock(cpt);
2705                 return rc;
2706         }
2707         lnet_peer_ni_decref_locked(lpni);
2708
2709         peer = lpni->lpni_peer_net->lpn_peer;
2710
2711         /*
2712          * Identify the different send cases
2713          */
2714         if (src_nid == LNET_NID_ANY)
2715                 send_case |= SRC_ANY;
2716         else
2717                 send_case |= SRC_SPEC;
2718
2719         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2720                 send_case |= LOCAL_DST;
2721         else
2722                 send_case |= REMOTE_DST;
2723
2724         final_hop = false;
2725         if (msg->msg_routing && (send_case & LOCAL_DST))
2726                 final_hop = true;
2727
2728         /* Determine whether to allow MR forwarding for this message.
2729          * NB: MR forwarding is allowed if the message originator and the
2730          * destination are both MR capable, and the destination lpni that was
2731          * originally chosen by the originator is unhealthy or down.
2732          * We check the MR capability of the destination further below
2733          */
2734         mr_forwarding_allowed = false;
2735         if (final_hop) {
2736                 struct lnet_peer *src_lp;
2737                 struct lnet_peer_ni *src_lpni;
2738
2739                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2740                                                   LNET_NID_ANY, cpt);
2741                 /* We don't fail the send if we hit any errors here. We'll just
2742                  * try to send it via non-multi-rail criteria
2743                  */
2744                 if (!IS_ERR(src_lpni)) {
2745                         src_lp = lpni->lpni_peer_net->lpn_peer;
2746                         if (lnet_peer_is_multi_rail(src_lp) &&
2747                             !lnet_is_peer_ni_alive(lpni))
2748                                 mr_forwarding_allowed = true;
2749
2750                 }
2751                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
2752                        mr_forwarding_allowed ? "allowed" : "not allowed");
2753         }
2754
2755         /*
2756          * Deal with the peer as NMR in the following cases:
2757          * 1. the peer is NMR
2758          * 2. We're trying to recover a specific peer NI
2759          * 3. I'm a router sending to the final destination and MR forwarding is
2760          *    not allowed for this message (as determined above).
2761          *    In this case the source of the message would've
2762          *    already selected the final destination so my job
2763          *    is to honor the selection.
2764          */
2765         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2766             (final_hop && !mr_forwarding_allowed))
2767                 send_case |= NMR_DST;
2768         else
2769                 send_case |= MR_DST;
2770
2771         if (lnet_msg_is_response(msg))
2772                 send_case |= SND_RESP;
2773
2774         /* assign parameters to the send_data */
2775         send_data.sd_rtr_nid = rtr_nid;
2776         send_data.sd_src_nid = src_nid;
2777         send_data.sd_dst_nid = dst_nid;
2778         send_data.sd_best_lpni = lpni;
2779         /*
2780          * keep a pointer to the final destination in case we're going to
2781          * route, so we'll need to access it later
2782          */
2783         send_data.sd_final_dst_lpni = lpni;
2784         send_data.sd_peer = peer;
2785         send_data.sd_md_cpt = md_cpt;
2786         send_data.sd_send_case = send_case;
2787
2788         rc = lnet_handle_send_case_locked(&send_data);
2789
2790         /*
2791          * Update the local cpt since send_data.sd_cpt might've been
2792          * updated as a result of calling lnet_handle_send_case_locked().
2793          */
2794         cpt = send_data.sd_cpt;
2795
2796         if (rc == REPEAT_SEND)
2797                 goto again;
2798
2799         lnet_net_unlock(cpt);
2800
2801         return rc;
2802 }
2803
2804 int
2805 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2806 {
2807         lnet_nid_t              dst_nid = msg->msg_target.nid;
2808         int                     rc;
2809
2810         /*
2811          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2812          * but we might want to use pre-determined router for ACK/REPLY
2813          * in the future
2814          */
2815         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2816         LASSERT(msg->msg_txpeer == NULL);
2817         LASSERT(msg->msg_txni == NULL);
2818         LASSERT(!msg->msg_sending);
2819         LASSERT(!msg->msg_target_is_router);
2820         LASSERT(!msg->msg_receiving);
2821
2822         msg->msg_sending = 1;
2823
2824         LASSERT(!msg->msg_tx_committed);
2825
2826         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2827         if (rc < 0) {
2828                 if (rc == -EHOSTUNREACH)
2829                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2830                 else
2831                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2832                 return rc;
2833         }
2834
2835         if (rc == LNET_CREDIT_OK)
2836                 lnet_ni_send(msg->msg_txni, msg);
2837
2838         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2839         return 0;
2840 }
2841
2842 enum lnet_mt_event_type {
2843         MT_TYPE_LOCAL_NI = 0,
2844         MT_TYPE_PEER_NI
2845 };
2846
2847 struct lnet_mt_event_info {
2848         enum lnet_mt_event_type mt_type;
2849         lnet_nid_t mt_nid;
2850 };
2851
2852 /* called with res_lock held */
2853 void
2854 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2855 {
2856         struct lnet_rsp_tracker *rspt;
2857
2858         /*
2859          * msg has a refcount on the MD so the MD is not going away.
2860          * The rspt queue for the cpt is protected by
2861          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2862          */
2863         if (!md->md_rspt_ptr)
2864                 return;
2865
2866         rspt = md->md_rspt_ptr;
2867
2868         /* debug code */
2869         LASSERT(rspt->rspt_cpt == cpt);
2870
2871         md->md_rspt_ptr = NULL;
2872
2873         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2874                 /*
2875                  * The monitor thread has invalidated this handle because the
2876                  * response timed out, but it failed to lookup the MD. That
2877                  * means this response tracker is on the zombie list. We can
2878                  * safely remove it under the resource lock (held by caller) and
2879                  * free the response tracker block.
2880                  */
2881                 list_del(&rspt->rspt_on_list);
2882                 lnet_rspt_free(rspt, cpt);
2883         } else {
2884                 /*
2885                  * invalidate the handle to indicate that a response has been
2886                  * received, which will then lead the monitor thread to clean up
2887                  * the rspt block.
2888                  */
2889                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2890         }
2891 }
2892
2893 void
2894 lnet_clean_zombie_rstqs(void)
2895 {
2896         struct lnet_rsp_tracker *rspt, *tmp;
2897         int i;
2898
2899         cfs_cpt_for_each(i, lnet_cpt_table()) {
2900                 list_for_each_entry_safe(rspt, tmp,
2901                                          the_lnet.ln_mt_zombie_rstqs[i],
2902                                          rspt_on_list) {
2903                         list_del(&rspt->rspt_on_list);
2904                         lnet_rspt_free(rspt, i);
2905                 }
2906         }
2907
2908         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2909 }
2910
2911 static void
2912 lnet_finalize_expired_responses(void)
2913 {
2914         struct lnet_libmd *md;
2915         struct lnet_rsp_tracker *rspt, *tmp;
2916         ktime_t now;
2917         int i;
2918
2919         if (the_lnet.ln_mt_rstq == NULL)
2920                 return;
2921
2922         cfs_cpt_for_each(i, lnet_cpt_table()) {
2923                 LIST_HEAD(local_queue);
2924
2925                 lnet_net_lock(i);
2926                 if (!the_lnet.ln_mt_rstq[i]) {
2927                         lnet_net_unlock(i);
2928                         continue;
2929                 }
2930                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2931                 lnet_net_unlock(i);
2932
2933                 now = ktime_get();
2934
2935                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2936                         /*
2937                          * The rspt mdh will be invalidated when a response
2938                          * is received or whenever we want to discard the
2939                          * block the monitor thread will walk the queue
2940                          * and clean up any rsts with an invalid mdh.
2941                          * The monitor thread will walk the queue until
2942                          * the first unexpired rspt block. This means that
2943                          * some rspt blocks which received their
2944                          * corresponding responses will linger in the
2945                          * queue until they are cleaned up eventually.
2946                          */
2947                         lnet_res_lock(i);
2948                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2949                                 lnet_res_unlock(i);
2950                                 list_del(&rspt->rspt_on_list);
2951                                 lnet_rspt_free(rspt, i);
2952                                 continue;
2953                         }
2954
2955                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2956                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2957                                 struct lnet_peer_ni *lpni;
2958                                 lnet_nid_t nid;
2959
2960                                 md = lnet_handle2md(&rspt->rspt_mdh);
2961                                 if (!md) {
2962                                         /* MD has been queued for unlink, but
2963                                          * rspt hasn't been detached (Note we've
2964                                          * checked above that the rspt_mdh is
2965                                          * valid). Since we cannot lookup the MD
2966                                          * we're unable to detach the rspt
2967                                          * ourselves. Thus, move the rspt to the
2968                                          * zombie list where we'll wait for
2969                                          * either:
2970                                          *   1. The remaining operations on the
2971                                          *   MD to complete. In this case the
2972                                          *   final operation will result in
2973                                          *   lnet_msg_detach_md()->
2974                                          *   lnet_detach_rsp_tracker() where
2975                                          *   we will clean up this response
2976                                          *   tracker.
2977                                          *   2. LNet to shutdown. In this case
2978                                          *   we'll wait until after all LND Nets
2979                                          *   have shutdown and then we can
2980                                          *   safely free any remaining response
2981                                          *   tracker blocks on the zombie list.
2982                                          * Note: We need to hold the resource
2983                                          * lock when adding to the zombie list
2984                                          * because we may have concurrent access
2985                                          * with lnet_detach_rsp_tracker().
2986                                          */
2987                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2988                                         list_move(&rspt->rspt_on_list,
2989                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2990                                         lnet_res_unlock(i);
2991                                         continue;
2992                                 }
2993                                 LASSERT(md->md_rspt_ptr == rspt);
2994                                 md->md_rspt_ptr = NULL;
2995                                 lnet_res_unlock(i);
2996
2997                                 LNetMDUnlink(rspt->rspt_mdh);
2998
2999                                 nid = rspt->rspt_next_hop_nid;
3000
3001                                 list_del(&rspt->rspt_on_list);
3002                                 lnet_rspt_free(rspt, i);
3003
3004                                 /* If we're shutting down we just want to clean
3005                                  * up the rspt blocks
3006                                  */
3007                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3008                                         continue;
3009
3010                                 lnet_net_lock(i);
3011                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3012                                 lnet_net_unlock(i);
3013
3014                                 CDEBUG(D_NET,
3015                                        "Response timeout: md = %p: nid = %s\n",
3016                                        md, libcfs_nid2str(nid));
3017
3018                                 /*
3019                                  * If there is a timeout on the response
3020                                  * from the next hop decrement its health
3021                                  * value so that we don't use it
3022                                  */
3023                                 lnet_net_lock(0);
3024                                 lpni = lnet_find_peer_ni_locked(nid);
3025                                 if (lpni) {
3026                                         lnet_handle_remote_failure_locked(lpni);
3027                                         lnet_peer_ni_decref_locked(lpni);
3028                                 }
3029                                 lnet_net_unlock(0);
3030                         } else {
3031                                 lnet_res_unlock(i);
3032                                 break;
3033                         }
3034                 }
3035
3036                 if (!list_empty(&local_queue)) {
3037                         lnet_net_lock(i);
3038                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3039                         lnet_net_unlock(i);
3040                 }
3041         }
3042 }
3043
3044 static void
3045 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3046 {
3047         struct lnet_msg *msg;
3048
3049         while (!list_empty(resendq)) {
3050                 struct lnet_peer_ni *lpni;
3051
3052                 msg = list_entry(resendq->next, struct lnet_msg,
3053                                  msg_list);
3054
3055                 list_del_init(&msg->msg_list);
3056
3057                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3058                 if (!lpni) {
3059                         lnet_net_unlock(cpt);
3060                         CERROR("Expected that a peer is already created for %s\n",
3061                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3062                         msg->msg_no_resend = true;
3063                         lnet_finalize(msg, -EFAULT);
3064                         lnet_net_lock(cpt);
3065                 } else {
3066                         int rc;
3067
3068                         lnet_peer_ni_decref_locked(lpni);
3069
3070                         lnet_net_unlock(cpt);
3071                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3072                                libcfs_nid2str(msg->msg_src_nid_param),
3073                                libcfs_id2str(msg->msg_target),
3074                                lnet_msgtyp2str(msg->msg_type),
3075                                msg->msg_recovery,
3076                                msg->msg_retry_count);
3077                         rc = lnet_send(msg->msg_src_nid_param, msg,
3078                                        msg->msg_rtr_nid_param);
3079                         if (rc) {
3080                                 CERROR("Error sending %s to %s: %d\n",
3081                                        lnet_msgtyp2str(msg->msg_type),
3082                                        libcfs_id2str(msg->msg_target), rc);
3083                                 msg->msg_no_resend = true;
3084                                 lnet_finalize(msg, rc);
3085                         }
3086                         lnet_net_lock(cpt);
3087                         if (!rc)
3088                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3089                 }
3090         }
3091 }
3092
3093 static void
3094 lnet_resend_pending_msgs(void)
3095 {
3096         int i;
3097
3098         cfs_cpt_for_each(i, lnet_cpt_table()) {
3099                 lnet_net_lock(i);
3100                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3101                 lnet_net_unlock(i);
3102         }
3103 }
3104
3105 /* called with cpt and ni_lock held */
3106 static void
3107 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3108 {
3109         struct lnet_handle_md recovery_mdh;
3110
3111         LNetInvalidateMDHandle(&recovery_mdh);
3112
3113         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3114             force) {
3115                 recovery_mdh = ni->ni_ping_mdh;
3116                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3117         }
3118         lnet_ni_unlock(ni);
3119         lnet_net_unlock(cpt);
3120         if (!LNetMDHandleIsInvalid(recovery_mdh))
3121                 LNetMDUnlink(recovery_mdh);
3122         lnet_net_lock(cpt);
3123         lnet_ni_lock(ni);
3124 }
3125
3126 static void
3127 lnet_recover_local_nis(void)
3128 {
3129         struct lnet_mt_event_info *ev_info;
3130         LIST_HEAD(processed_list);
3131         LIST_HEAD(local_queue);
3132         struct lnet_handle_md mdh;
3133         struct lnet_ni *tmp;
3134         struct lnet_ni *ni;
3135         lnet_nid_t nid;
3136         int healthv;
3137         int rc;
3138
3139         /*
3140          * splice the recovery queue on a local queue. We will iterate
3141          * through the local queue and update it as needed. Once we're
3142          * done with the traversal, we'll splice the local queue back on
3143          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3144          * will be traversed in the next iteration.
3145          */
3146         lnet_net_lock(0);
3147         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3148                          &local_queue);
3149         lnet_net_unlock(0);
3150
3151         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3152                 /*
3153                  * if an NI is being deleted or it is now healthy, there
3154                  * is no need to keep it around in the recovery queue.
3155                  * The monitor thread is the only thread responsible for
3156                  * removing the NI from the recovery queue.
3157                  * Multiple threads can be adding NIs to the recovery
3158                  * queue.
3159                  */
3160                 healthv = atomic_read(&ni->ni_healthv);
3161
3162                 lnet_net_lock(0);
3163                 lnet_ni_lock(ni);
3164                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3165                     healthv == LNET_MAX_HEALTH_VALUE) {
3166                         list_del_init(&ni->ni_recovery);
3167                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3168                         lnet_ni_unlock(ni);
3169                         lnet_ni_decref_locked(ni, 0);
3170                         lnet_net_unlock(0);
3171                         continue;
3172                 }
3173
3174                 /*
3175                  * if the local NI failed recovery we must unlink the md.
3176                  * But we want to keep the local_ni on the recovery queue
3177                  * so we can continue the attempts to recover it.
3178                  */
3179                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3180                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3181                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3182                 }
3183
3184                 lnet_ni_unlock(ni);
3185                 lnet_net_unlock(0);
3186
3187
3188                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3189                        libcfs_nid2str(ni->ni_nid));
3190
3191                 lnet_ni_lock(ni);
3192                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3193                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3194                         lnet_ni_unlock(ni);
3195
3196                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3197                         if (!ev_info) {
3198                                 CERROR("out of memory. Can't recover %s\n",
3199                                        libcfs_nid2str(ni->ni_nid));
3200                                 lnet_ni_lock(ni);
3201                                 ni->ni_recovery_state &=
3202                                   ~LNET_NI_RECOVERY_PENDING;
3203                                 lnet_ni_unlock(ni);
3204                                 continue;
3205                         }
3206
3207                         mdh = ni->ni_ping_mdh;
3208                         /*
3209                          * Invalidate the ni mdh in case it's deleted.
3210                          * We'll unlink the mdh in this case below.
3211                          */
3212                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3213                         nid = ni->ni_nid;
3214
3215                         /*
3216                          * remove the NI from the local queue and drop the
3217                          * reference count to it while we're recovering
3218                          * it. The reason for that, is that the NI could
3219                          * be deleted, and the way the code is structured
3220                          * is if we don't drop the NI, then the deletion
3221                          * code will enter a loop waiting for the
3222                          * reference count to be removed while holding the
3223                          * ln_mutex_lock(). When we look up the peer to
3224                          * send to in lnet_select_pathway() we will try to
3225                          * lock the ln_mutex_lock() as well, leading to
3226                          * a deadlock. By dropping the refcount and
3227                          * removing it from the list, we allow for the NI
3228                          * to be removed, then we use the cached NID to
3229                          * look it up again. If it's gone, then we just
3230                          * continue examining the rest of the queue.
3231                          */
3232                         lnet_net_lock(0);
3233                         list_del_init(&ni->ni_recovery);
3234                         lnet_ni_decref_locked(ni, 0);
3235                         lnet_net_unlock(0);
3236
3237                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3238                         ev_info->mt_nid = nid;
3239                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3240                                             ev_info, the_lnet.ln_mt_handler,
3241                                             true);
3242                         /* lookup the nid again */
3243                         lnet_net_lock(0);
3244                         ni = lnet_nid2ni_locked(nid, 0);
3245                         if (!ni) {
3246                                 /*
3247                                  * the NI has been deleted when we dropped
3248                                  * the ref count
3249                                  */
3250                                 lnet_net_unlock(0);
3251                                 LNetMDUnlink(mdh);
3252                                 continue;
3253                         }
3254                         /*
3255                          * Same note as in lnet_recover_peer_nis(). When
3256                          * we're sending the ping, the NI is free to be
3257                          * deleted or manipulated. By this point it
3258                          * could've been added back on the recovery queue,
3259                          * and a refcount taken on it.
3260                          * So we can't just add it blindly again or we'll
3261                          * corrupt the queue. We must check under lock if
3262                          * it's not on any list and if not then add it
3263                          * to the processed list, which will eventually be
3264                          * spliced back on to the recovery queue.
3265                          */
3266                         ni->ni_ping_mdh = mdh;
3267                         if (list_empty(&ni->ni_recovery)) {
3268                                 list_add_tail(&ni->ni_recovery, &processed_list);
3269                                 lnet_ni_addref_locked(ni, 0);
3270                         }
3271                         lnet_net_unlock(0);
3272
3273                         lnet_ni_lock(ni);
3274                         if (rc)
3275                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3276                 }
3277                 lnet_ni_unlock(ni);
3278         }
3279
3280         /*
3281          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3282          * reexamined in the next iteration.
3283          */
3284         list_splice_init(&processed_list, &local_queue);
3285         lnet_net_lock(0);
3286         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3287         lnet_net_unlock(0);
3288 }
3289
3290 static int
3291 lnet_resendqs_create(void)
3292 {
3293         struct list_head **resendqs;
3294         resendqs = lnet_create_array_of_queues();
3295
3296         if (!resendqs)
3297                 return -ENOMEM;
3298
3299         lnet_net_lock(LNET_LOCK_EX);
3300         the_lnet.ln_mt_resendqs = resendqs;
3301         lnet_net_unlock(LNET_LOCK_EX);
3302
3303         return 0;
3304 }
3305
3306 static void
3307 lnet_clean_local_ni_recoveryq(void)
3308 {
3309         struct lnet_ni *ni;
3310
3311         /* This is only called when the monitor thread has stopped */
3312         lnet_net_lock(0);
3313
3314         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3315                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3316                                 struct lnet_ni, ni_recovery);
3317                 list_del_init(&ni->ni_recovery);
3318                 lnet_ni_lock(ni);
3319                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3320                 lnet_ni_unlock(ni);
3321                 lnet_ni_decref_locked(ni, 0);
3322         }
3323
3324         lnet_net_unlock(0);
3325 }
3326
3327 static void
3328 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3329                                      bool force)
3330 {
3331         struct lnet_handle_md recovery_mdh;
3332
3333         LNetInvalidateMDHandle(&recovery_mdh);
3334
3335         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3336                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3337                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3338         }
3339         spin_unlock(&lpni->lpni_lock);
3340         lnet_net_unlock(cpt);
3341         if (!LNetMDHandleIsInvalid(recovery_mdh))
3342                 LNetMDUnlink(recovery_mdh);
3343         lnet_net_lock(cpt);
3344         spin_lock(&lpni->lpni_lock);
3345 }
3346
3347 static void
3348 lnet_clean_peer_ni_recoveryq(void)
3349 {
3350         struct lnet_peer_ni *lpni, *tmp;
3351
3352         lnet_net_lock(LNET_LOCK_EX);
3353
3354         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3355                                  lpni_recovery) {
3356                 list_del_init(&lpni->lpni_recovery);
3357                 spin_lock(&lpni->lpni_lock);
3358                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3359                 spin_unlock(&lpni->lpni_lock);
3360                 lnet_peer_ni_decref_locked(lpni);
3361         }
3362
3363         lnet_net_unlock(LNET_LOCK_EX);
3364 }
3365
3366 static void
3367 lnet_clean_resendqs(void)
3368 {
3369         struct lnet_msg *msg, *tmp;
3370         LIST_HEAD(msgs);
3371         int i;
3372
3373         cfs_cpt_for_each(i, lnet_cpt_table()) {
3374                 lnet_net_lock(i);
3375                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3376                 lnet_net_unlock(i);
3377                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3378                         list_del_init(&msg->msg_list);
3379                         msg->msg_no_resend = true;
3380                         lnet_finalize(msg, -ESHUTDOWN);
3381                 }
3382         }
3383
3384         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3385 }
3386
3387 static void
3388 lnet_recover_peer_nis(void)
3389 {
3390         struct lnet_mt_event_info *ev_info;
3391         LIST_HEAD(processed_list);
3392         LIST_HEAD(local_queue);
3393         struct lnet_handle_md mdh;
3394         struct lnet_peer_ni *lpni;
3395         struct lnet_peer_ni *tmp;
3396         lnet_nid_t nid;
3397         int healthv;
3398         int rc;
3399
3400         /*
3401          * Always use cpt 0 for locking across all interactions with
3402          * ln_mt_peerNIRecovq
3403          */
3404         lnet_net_lock(0);
3405         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3406                          &local_queue);
3407         lnet_net_unlock(0);
3408
3409         list_for_each_entry_safe(lpni, tmp, &local_queue,
3410                                  lpni_recovery) {
3411                 /*
3412                  * The same protection strategy is used here as is in the
3413                  * local recovery case.
3414                  */
3415                 lnet_net_lock(0);
3416                 healthv = atomic_read(&lpni->lpni_healthv);
3417                 spin_lock(&lpni->lpni_lock);
3418                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3419                     healthv == LNET_MAX_HEALTH_VALUE) {
3420                         list_del_init(&lpni->lpni_recovery);
3421                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3422                         spin_unlock(&lpni->lpni_lock);
3423                         lnet_peer_ni_decref_locked(lpni);
3424                         lnet_net_unlock(0);
3425                         continue;
3426                 }
3427
3428                 /*
3429                  * If the peer NI has failed recovery we must unlink the
3430                  * md. But we want to keep the peer ni on the recovery
3431                  * queue so we can try to continue recovering it
3432                  */
3433                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3434                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3435                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3436                 }
3437
3438                 spin_unlock(&lpni->lpni_lock);
3439                 lnet_net_unlock(0);
3440
3441                 /*
3442                  * NOTE: we're racing with peer deletion from user space.
3443                  * It's possible that a peer is deleted after we check its
3444                  * state. In this case the recovery can create a new peer
3445                  */
3446                 spin_lock(&lpni->lpni_lock);
3447                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3448                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3449                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3450                         spin_unlock(&lpni->lpni_lock);
3451
3452                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3453                         if (!ev_info) {
3454                                 CERROR("out of memory. Can't recover %s\n",
3455                                        libcfs_nid2str(lpni->lpni_nid));
3456                                 spin_lock(&lpni->lpni_lock);
3457                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3458                                 spin_unlock(&lpni->lpni_lock);
3459                                 continue;
3460                         }
3461
3462                         /* look at the comments in lnet_recover_local_nis() */
3463                         mdh = lpni->lpni_recovery_ping_mdh;
3464                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3465                         nid = lpni->lpni_nid;
3466                         lnet_net_lock(0);
3467                         list_del_init(&lpni->lpni_recovery);
3468                         lnet_peer_ni_decref_locked(lpni);
3469                         lnet_net_unlock(0);
3470
3471                         ev_info->mt_type = MT_TYPE_PEER_NI;
3472                         ev_info->mt_nid = nid;
3473                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3474                                             ev_info, the_lnet.ln_mt_handler,
3475                                             true);
3476                         lnet_net_lock(0);
3477                         /*
3478                          * lnet_find_peer_ni_locked() grabs a refcount for
3479                          * us. No need to take it explicitly.
3480                          */
3481                         lpni = lnet_find_peer_ni_locked(nid);
3482                         if (!lpni) {
3483                                 lnet_net_unlock(0);
3484                                 LNetMDUnlink(mdh);
3485                                 continue;
3486                         }
3487
3488                         lpni->lpni_recovery_ping_mdh = mdh;
3489                         /*
3490                          * While we're unlocked the lpni could've been
3491                          * readded on the recovery queue. In this case we
3492                          * don't need to add it to the local queue, since
3493                          * it's already on there and the thread that added
3494                          * it would've incremented the refcount on the
3495                          * peer, which means we need to decref the refcount
3496                          * that was implicitly grabbed by find_peer_ni_locked.
3497                          * Otherwise, if the lpni is still not on
3498                          * the recovery queue, then we'll add it to the
3499                          * processed list.
3500                          */
3501                         if (list_empty(&lpni->lpni_recovery))
3502                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3503                         else
3504                                 lnet_peer_ni_decref_locked(lpni);
3505                         lnet_net_unlock(0);
3506
3507                         spin_lock(&lpni->lpni_lock);
3508                         if (rc)
3509                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3510                 }
3511                 spin_unlock(&lpni->lpni_lock);
3512         }
3513
3514         list_splice_init(&processed_list, &local_queue);
3515         lnet_net_lock(0);
3516         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3517         lnet_net_unlock(0);
3518 }
3519
3520 static int
3521 lnet_monitor_thread(void *arg)
3522 {
3523         time64_t recovery_timeout = 0;
3524         time64_t rsp_timeout = 0;
3525         int interval;
3526         time64_t now;
3527
3528         wait_for_completion(&the_lnet.ln_started);
3529         /*
3530          * The monitor thread takes care of the following:
3531          *  1. Checks the aliveness of routers
3532          *  2. Checks if there are messages on the resend queue to resend
3533          *     them.
3534          *  3. Check if there are any NIs on the local recovery queue and
3535          *     pings them
3536          *  4. Checks if there are any NIs on the remote recovery queue
3537          *     and pings them.
3538          */
3539         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3540                 now = ktime_get_real_seconds();
3541
3542                 if (lnet_router_checker_active())
3543                         lnet_check_routers();
3544
3545                 lnet_resend_pending_msgs();
3546
3547                 if (now >= rsp_timeout) {
3548                         lnet_finalize_expired_responses();
3549                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3550                 }
3551
3552                 if (now >= recovery_timeout) {
3553                         lnet_recover_local_nis();
3554                         lnet_recover_peer_nis();
3555                         recovery_timeout = now + lnet_recovery_interval;
3556                 }
3557
3558                 /*
3559                  * TODO do we need to check if we should sleep without
3560                  * timeout?  Technically, an active system will always
3561                  * have messages in flight so this check will always
3562                  * evaluate to false. And on an idle system do we care
3563                  * if we wake up every 1 second? Although, we've seen
3564                  * cases where we get a complaint that an idle thread
3565                  * is waking up unnecessarily.
3566                  *
3567                  * Take into account the current net_count when you wake
3568                  * up for alive router checking, since we need to check
3569                  * possibly as many networks as we have configured.
3570                  */
3571                 interval = min(lnet_recovery_interval,
3572                                min((unsigned int) alive_router_check_interval /
3573                                         lnet_current_net_count,
3574                                    lnet_transaction_timeout / 2));
3575                 wait_for_completion_interruptible_timeout(
3576                         &the_lnet.ln_mt_wait_complete,
3577                         cfs_time_seconds(interval));
3578                 /* Must re-init the completion before testing anything,
3579                  * including ln_mt_state.
3580                  */
3581                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3582         }
3583
3584         /* Shutting down */
3585         lnet_net_lock(LNET_LOCK_EX);
3586         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3587         lnet_net_unlock(LNET_LOCK_EX);
3588
3589         /* signal that the monitor thread is exiting */
3590         up(&the_lnet.ln_mt_signal);
3591
3592         return 0;
3593 }
3594
3595 /*
3596  * lnet_send_ping
3597  * Sends a ping.
3598  * Returns == 0 if success
3599  * Returns > 0 if LNetMDBind or prior fails
3600  * Returns < 0 if LNetGet fails
3601  */
3602 int
3603 lnet_send_ping(lnet_nid_t dest_nid,
3604                struct lnet_handle_md *mdh, int nnis,
3605                void *user_data, lnet_handler_t handler, bool recovery)
3606 {
3607         struct lnet_md md = { NULL };
3608         struct lnet_process_id id;
3609         struct lnet_ping_buffer *pbuf;
3610         int rc;
3611
3612         if (dest_nid == LNET_NID_ANY) {
3613                 rc = -EHOSTUNREACH;
3614                 goto fail_error;
3615         }
3616
3617         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3618         if (!pbuf) {
3619                 rc = ENOMEM;
3620                 goto fail_error;
3621         }
3622
3623         /* initialize md content */
3624         md.start     = &pbuf->pb_info;
3625         md.length    = LNET_PING_INFO_SIZE(nnis);
3626         md.threshold = 2; /* GET/REPLY */
3627         md.max_size  = 0;
3628         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3629         md.user_ptr  = user_data;
3630         md.handler   = handler;
3631
3632         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3633         if (rc) {
3634                 lnet_ping_buffer_decref(pbuf);
3635                 CERROR("Can't bind MD: %d\n", rc);
3636                 rc = -rc; /* change the rc to positive */
3637                 goto fail_error;
3638         }
3639         id.pid = LNET_PID_LUSTRE;
3640         id.nid = dest_nid;
3641
3642         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3643                      LNET_RESERVED_PORTAL,
3644                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3645
3646         if (rc)
3647                 goto fail_unlink_md;
3648
3649         return 0;
3650
3651 fail_unlink_md:
3652         LNetMDUnlink(*mdh);
3653         LNetInvalidateMDHandle(mdh);
3654 fail_error:
3655         return rc;
3656 }
3657
3658 static void
3659 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3660                            int status, bool send, bool unlink_event)
3661 {
3662         lnet_nid_t nid = ev_info->mt_nid;
3663
3664         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3665                 struct lnet_ni *ni;
3666
3667                 lnet_net_lock(0);
3668                 ni = lnet_nid2ni_locked(nid, 0);
3669                 if (!ni) {
3670                         lnet_net_unlock(0);
3671                         return;
3672                 }
3673                 lnet_ni_lock(ni);
3674                 if (!send || (send && status != 0))
3675                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3676                 if (status)
3677                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3678                 lnet_ni_unlock(ni);
3679                 lnet_net_unlock(0);
3680
3681                 if (status != 0) {
3682                         CERROR("local NI (%s) recovery failed with %d\n",
3683                                libcfs_nid2str(nid), status);
3684                         return;
3685                 }
3686                 /*
3687                  * need to increment healthv for the ni here, because in
3688                  * the lnet_finalize() path we don't have access to this
3689                  * NI. And in order to get access to it, we'll need to
3690                  * carry forward too much information.
3691                  * In the peer case, it'll naturally be incremented
3692                  */
3693                 if (!unlink_event)
3694                         lnet_inc_healthv(&ni->ni_healthv,
3695                                          lnet_health_sensitivity);
3696         } else {
3697                 struct lnet_peer_ni *lpni;
3698                 int cpt;
3699
3700                 cpt = lnet_net_lock_current();
3701                 lpni = lnet_find_peer_ni_locked(nid);
3702                 if (!lpni) {
3703                         lnet_net_unlock(cpt);
3704                         return;
3705                 }
3706                 spin_lock(&lpni->lpni_lock);
3707                 if (!send || (send && status != 0))
3708                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3709                 if (status)
3710                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3711                 spin_unlock(&lpni->lpni_lock);
3712                 lnet_peer_ni_decref_locked(lpni);
3713                 lnet_net_unlock(cpt);
3714
3715                 if (status != 0)
3716                         CERROR("peer NI (%s) recovery failed with %d\n",
3717                                libcfs_nid2str(nid), status);
3718         }
3719 }
3720
3721 void
3722 lnet_mt_event_handler(struct lnet_event *event)
3723 {
3724         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3725         struct lnet_ping_buffer *pbuf;
3726
3727         /* TODO: remove assert */
3728         LASSERT(event->type == LNET_EVENT_REPLY ||
3729                 event->type == LNET_EVENT_SEND ||
3730                 event->type == LNET_EVENT_UNLINK);
3731
3732         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3733                event->status);
3734
3735         switch (event->type) {
3736         case LNET_EVENT_UNLINK:
3737                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3738                        libcfs_nid2str(ev_info->mt_nid));
3739                 /* fallthrough */
3740         case LNET_EVENT_REPLY:
3741                 lnet_handle_recovery_reply(ev_info, event->status, false,
3742                                            event->type == LNET_EVENT_UNLINK);
3743                 break;
3744         case LNET_EVENT_SEND:
3745                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3746                                libcfs_nid2str(ev_info->mt_nid),
3747                                (event->status) ? "unsuccessfully" :
3748                                "successfully", event->status);
3749                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3750                 break;
3751         default:
3752                 CERROR("Unexpected event: %d\n", event->type);
3753                 break;
3754         }
3755         if (event->unlinked) {
3756                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3757                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
3758                 lnet_ping_buffer_decref(pbuf);
3759         }
3760 }
3761
3762 static int
3763 lnet_rsp_tracker_create(void)
3764 {
3765         struct list_head **rstqs;
3766         rstqs = lnet_create_array_of_queues();
3767
3768         if (!rstqs)
3769                 return -ENOMEM;
3770
3771         the_lnet.ln_mt_rstq = rstqs;
3772
3773         return 0;
3774 }
3775
3776 static void
3777 lnet_rsp_tracker_clean(void)
3778 {
3779         lnet_finalize_expired_responses();
3780
3781         cfs_percpt_free(the_lnet.ln_mt_rstq);
3782         the_lnet.ln_mt_rstq = NULL;
3783 }
3784
3785 int lnet_monitor_thr_start(void)
3786 {
3787         int rc = 0;
3788         struct task_struct *task;
3789
3790         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3791                 return -EALREADY;
3792
3793         rc = lnet_resendqs_create();
3794         if (rc)
3795                 return rc;
3796
3797         rc = lnet_rsp_tracker_create();
3798         if (rc)
3799                 goto clean_queues;
3800
3801         sema_init(&the_lnet.ln_mt_signal, 0);
3802
3803         lnet_net_lock(LNET_LOCK_EX);
3804         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3805         lnet_net_unlock(LNET_LOCK_EX);
3806         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3807         if (IS_ERR(task)) {
3808                 rc = PTR_ERR(task);
3809                 CERROR("Can't start monitor thread: %d\n", rc);
3810                 goto clean_thread;
3811         }
3812
3813         return 0;
3814
3815 clean_thread:
3816         lnet_net_lock(LNET_LOCK_EX);
3817         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3818         lnet_net_unlock(LNET_LOCK_EX);
3819         /* block until event callback signals exit */
3820         down(&the_lnet.ln_mt_signal);
3821         /* clean up */
3822         lnet_net_lock(LNET_LOCK_EX);
3823         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3824         lnet_net_unlock(LNET_LOCK_EX);
3825         lnet_rsp_tracker_clean();
3826         lnet_clean_local_ni_recoveryq();
3827         lnet_clean_peer_ni_recoveryq();
3828         lnet_clean_resendqs();
3829         the_lnet.ln_mt_handler = NULL;
3830         return rc;
3831 clean_queues:
3832         lnet_rsp_tracker_clean();
3833         lnet_clean_local_ni_recoveryq();
3834         lnet_clean_peer_ni_recoveryq();
3835         lnet_clean_resendqs();
3836         return rc;
3837 }
3838
3839 void lnet_monitor_thr_stop(void)
3840 {
3841         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3842                 return;
3843
3844         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3845         lnet_net_lock(LNET_LOCK_EX);
3846         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3847         lnet_net_unlock(LNET_LOCK_EX);
3848
3849         /* tell the monitor thread that we're shutting down */
3850         complete(&the_lnet.ln_mt_wait_complete);
3851
3852         /* block until monitor thread signals that it's done */
3853         down(&the_lnet.ln_mt_signal);
3854         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3855
3856         /* perform cleanup tasks */
3857         lnet_rsp_tracker_clean();
3858         lnet_clean_local_ni_recoveryq();
3859         lnet_clean_peer_ni_recoveryq();
3860         lnet_clean_resendqs();
3861 }
3862
3863 void
3864 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3865                   __u32 msg_type)
3866 {
3867         lnet_net_lock(cpt);
3868         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3869         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3870         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3871         lnet_net_unlock(cpt);
3872
3873         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3874 }
3875
3876 static void
3877 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3878 {
3879         struct lnet_hdr *hdr = &msg->msg_hdr;
3880
3881         if (msg->msg_wanted != 0)
3882                 lnet_setpayloadbuffer(msg);
3883
3884         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3885
3886         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3887          * it back into the ACK during lnet_finalize() */
3888         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3889                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3890
3891         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3892                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3893 }
3894
3895 static int
3896 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3897 {
3898         struct lnet_hdr         *hdr = &msg->msg_hdr;
3899         struct lnet_match_info  info;
3900         int                     rc;
3901         bool                    ready_delay;
3902
3903         /* Convert put fields to host byte order */
3904         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3905         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3906         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3907
3908         /* Primary peer NID. */
3909         info.mi_id.nid  = msg->msg_initiator;
3910         info.mi_id.pid  = hdr->src_pid;
3911         info.mi_opc     = LNET_MD_OP_PUT;
3912         info.mi_portal  = hdr->msg.put.ptl_index;
3913         info.mi_rlength = hdr->payload_length;
3914         info.mi_roffset = hdr->msg.put.offset;
3915         info.mi_mbits   = hdr->msg.put.match_bits;
3916         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3917
3918         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3919         ready_delay = msg->msg_rx_ready_delay;
3920
3921  again:
3922         rc = lnet_ptl_match_md(&info, msg);
3923         switch (rc) {
3924         default:
3925                 LBUG();
3926
3927         case LNET_MATCHMD_OK:
3928                 lnet_recv_put(ni, msg);
3929                 return 0;
3930
3931         case LNET_MATCHMD_NONE:
3932                 if (ready_delay)
3933                         /* no eager_recv or has already called it, should
3934                          * have been attached on delayed list */
3935                         return 0;
3936
3937                 rc = lnet_ni_eager_recv(ni, msg);
3938                 if (rc == 0) {
3939                         ready_delay = true;
3940                         goto again;
3941                 }
3942                 /* fall through */
3943
3944         case LNET_MATCHMD_DROP:
3945                 CNETERR("Dropping PUT from %s portal %d match %llu"
3946                         " offset %d length %d: %d\n",
3947                         libcfs_id2str(info.mi_id), info.mi_portal,
3948                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3949
3950                 return -ENOENT; /* -ve: OK but no match */
3951         }
3952 }
3953
3954 static int
3955 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3956 {
3957         struct lnet_match_info info;
3958         struct lnet_hdr *hdr = &msg->msg_hdr;
3959         struct lnet_process_id source_id;
3960         struct lnet_handle_wire reply_wmd;
3961         int rc;
3962
3963         /* Convert get fields to host byte order */
3964         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3965         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3966         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3967         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3968
3969         source_id.nid = hdr->src_nid;
3970         source_id.pid = hdr->src_pid;
3971         /* Primary peer NID */
3972         info.mi_id.nid  = msg->msg_initiator;
3973         info.mi_id.pid  = hdr->src_pid;
3974         info.mi_opc     = LNET_MD_OP_GET;
3975         info.mi_portal  = hdr->msg.get.ptl_index;
3976         info.mi_rlength = hdr->msg.get.sink_length;
3977         info.mi_roffset = hdr->msg.get.src_offset;
3978         info.mi_mbits   = hdr->msg.get.match_bits;
3979         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3980
3981         rc = lnet_ptl_match_md(&info, msg);
3982         if (rc == LNET_MATCHMD_DROP) {
3983                 CNETERR("Dropping GET from %s portal %d match %llu"
3984                         " offset %d length %d\n",
3985                         libcfs_id2str(info.mi_id), info.mi_portal,
3986                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3987                 return -ENOENT; /* -ve: OK but no match */
3988         }
3989
3990         LASSERT(rc == LNET_MATCHMD_OK);
3991
3992         lnet_build_msg_event(msg, LNET_EVENT_GET);
3993
3994         reply_wmd = hdr->msg.get.return_wmd;
3995
3996         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3997                        msg->msg_offset, msg->msg_wanted);
3998
3999         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4000
4001         if (rdma_get) {
4002                 /* The LND completes the REPLY from her recv procedure */
4003                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4004                              msg->msg_offset, msg->msg_len, msg->msg_len);
4005                 return 0;
4006         }
4007
4008         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4009         msg->msg_receiving = 0;
4010
4011         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
4012         if (rc < 0) {
4013                 /* didn't get as far as lnet_ni_send() */
4014                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4015                        libcfs_nid2str(ni->ni_nid),
4016                        libcfs_id2str(info.mi_id), rc);
4017
4018                 lnet_finalize(msg, rc);
4019         }
4020
4021         return 0;
4022 }
4023
4024 static int
4025 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4026 {
4027         void *private = msg->msg_private;
4028         struct lnet_hdr *hdr = &msg->msg_hdr;
4029         struct lnet_process_id src = {0};
4030         struct lnet_libmd *md;
4031         unsigned int rlength;
4032         unsigned int mlength;
4033         int cpt;
4034
4035         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4036         lnet_res_lock(cpt);
4037
4038         src.nid = hdr->src_nid;
4039         src.pid = hdr->src_pid;
4040
4041         /* NB handles only looked up by creator (no flips) */
4042         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4043         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4044                 CNETERR("%s: Dropping REPLY from %s for %s "
4045                         "MD %#llx.%#llx\n",
4046                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4047                         (md == NULL) ? "invalid" : "inactive",
4048                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4049                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4050                 if (md != NULL && md->md_me != NULL)
4051                         CERROR("REPLY MD also attached to portal %d\n",
4052                                md->md_me->me_portal);
4053
4054                 lnet_res_unlock(cpt);
4055                 return -ENOENT; /* -ve: OK but no match */
4056         }
4057
4058         LASSERT(md->md_offset == 0);
4059
4060         rlength = hdr->payload_length;
4061         mlength = min(rlength, md->md_length);
4062
4063         if (mlength < rlength &&
4064             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4065                 CNETERR("%s: Dropping REPLY from %s length %d "
4066                         "for MD %#llx would overflow (%d)\n",
4067                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4068                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4069                         mlength);
4070                 lnet_res_unlock(cpt);
4071                 return -ENOENT; /* -ve: OK but no match */
4072         }
4073
4074         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4075                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4076                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4077
4078         lnet_msg_attach_md(msg, md, 0, mlength);
4079
4080         if (mlength != 0)
4081                 lnet_setpayloadbuffer(msg);
4082
4083         lnet_res_unlock(cpt);
4084
4085         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4086
4087         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4088         return 0;
4089 }
4090
4091 static int
4092 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4093 {
4094         struct lnet_hdr *hdr = &msg->msg_hdr;
4095         struct lnet_process_id src = {0};
4096         struct lnet_libmd *md;
4097         int cpt;
4098
4099         src.nid = hdr->src_nid;
4100         src.pid = hdr->src_pid;
4101
4102         /* Convert ack fields to host byte order */
4103         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4104         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4105
4106         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4107         lnet_res_lock(cpt);
4108
4109         /* NB handles only looked up by creator (no flips) */
4110         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4111         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4112                 /* Don't moan; this is expected */
4113                 CDEBUG(D_NET,
4114                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4115                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4116                        (md == NULL) ? "invalid" : "inactive",
4117                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4118                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4119                 if (md != NULL && md->md_me != NULL)
4120                         CERROR("Source MD also attached to portal %d\n",
4121                                md->md_me->me_portal);
4122
4123                 lnet_res_unlock(cpt);
4124                 return -ENOENT;                  /* -ve! */
4125         }
4126
4127         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4128                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4129                hdr->msg.ack.dst_wmd.wh_object_cookie);
4130
4131         lnet_msg_attach_md(msg, md, 0, 0);
4132
4133         lnet_res_unlock(cpt);
4134
4135         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4136
4137         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4138         return 0;
4139 }
4140
4141 /**
4142  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4143  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4144  * \retval -ve                  error code
4145  */
4146 int
4147 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4148 {
4149         int     rc = 0;
4150
4151         if (!the_lnet.ln_routing)
4152                 return -ECANCELED;
4153
4154         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4155             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4156                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4157                         msg->msg_rx_ready_delay = 1;
4158                 } else {
4159                         lnet_net_unlock(msg->msg_rx_cpt);
4160                         rc = lnet_ni_eager_recv(ni, msg);
4161                         lnet_net_lock(msg->msg_rx_cpt);
4162                 }
4163         }
4164
4165         if (rc == 0)
4166                 rc = lnet_post_routed_recv_locked(msg, 0);
4167         return rc;
4168 }
4169
4170 int
4171 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4172 {
4173         int     rc;
4174
4175         switch (msg->msg_type) {
4176         case LNET_MSG_ACK:
4177                 rc = lnet_parse_ack(ni, msg);
4178                 break;
4179         case LNET_MSG_PUT:
4180                 rc = lnet_parse_put(ni, msg);
4181                 break;
4182         case LNET_MSG_GET:
4183                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4184                 break;
4185         case LNET_MSG_REPLY:
4186                 rc = lnet_parse_reply(ni, msg);
4187                 break;
4188         default: /* prevent an unused label if !kernel */
4189                 LASSERT(0);
4190                 return -EPROTO;
4191         }
4192
4193         LASSERT(rc == 0 || rc == -ENOENT);
4194         return rc;
4195 }
4196
4197 char *
4198 lnet_msgtyp2str (int type)
4199 {
4200         switch (type) {
4201         case LNET_MSG_ACK:
4202                 return ("ACK");
4203         case LNET_MSG_PUT:
4204                 return ("PUT");
4205         case LNET_MSG_GET:
4206                 return ("GET");
4207         case LNET_MSG_REPLY:
4208                 return ("REPLY");
4209         case LNET_MSG_HELLO:
4210                 return ("HELLO");
4211         default:
4212                 return ("<UNKNOWN>");
4213         }
4214 }
4215
4216 void
4217 lnet_print_hdr(struct lnet_hdr *hdr)
4218 {
4219         struct lnet_process_id src = {
4220                 .nid = hdr->src_nid,
4221                 .pid = hdr->src_pid,
4222         };
4223         struct lnet_process_id dst = {
4224                 .nid = hdr->dest_nid,
4225                 .pid = hdr->dest_pid,
4226         };
4227         char *type_str = lnet_msgtyp2str(hdr->type);
4228
4229         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4230         CWARN("    From %s\n", libcfs_id2str(src));
4231         CWARN("    To   %s\n", libcfs_id2str(dst));
4232
4233         switch (hdr->type) {
4234         default:
4235                 break;
4236
4237         case LNET_MSG_PUT:
4238                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4239                       "match bits %llu\n",
4240                       hdr->msg.put.ptl_index,
4241                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4242                       hdr->msg.put.ack_wmd.wh_object_cookie,
4243                       hdr->msg.put.match_bits);
4244                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4245                       hdr->payload_length, hdr->msg.put.offset,
4246                       hdr->msg.put.hdr_data);
4247                 break;
4248
4249         case LNET_MSG_GET:
4250                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4251                       "match bits %llu\n", hdr->msg.get.ptl_index,
4252                       hdr->msg.get.return_wmd.wh_interface_cookie,
4253                       hdr->msg.get.return_wmd.wh_object_cookie,
4254                       hdr->msg.get.match_bits);
4255                 CWARN("    Length %d, src offset %d\n",
4256                       hdr->msg.get.sink_length,
4257                       hdr->msg.get.src_offset);
4258                 break;
4259
4260         case LNET_MSG_ACK:
4261                 CWARN("    dst md %#llx.%#llx, "
4262                       "manipulated length %d\n",
4263                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4264                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4265                       hdr->msg.ack.mlength);
4266                 break;
4267
4268         case LNET_MSG_REPLY:
4269                 CWARN("    dst md %#llx.%#llx, "
4270                       "length %d\n",
4271                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4272                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4273                       hdr->payload_length);
4274         }
4275
4276 }
4277
4278 int
4279 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4280            void *private, int rdma_req)
4281 {
4282         struct lnet_peer_ni *lpni;
4283         struct lnet_msg *msg;
4284         __u32 payload_length;
4285         lnet_pid_t dest_pid;
4286         lnet_nid_t dest_nid;
4287         lnet_nid_t src_nid;
4288         bool push = false;
4289         int for_me;
4290         __u32 type;
4291         int rc = 0;
4292         int cpt;
4293
4294         LASSERT (!in_interrupt ());
4295
4296         type = le32_to_cpu(hdr->type);
4297         src_nid = le64_to_cpu(hdr->src_nid);
4298         dest_nid = le64_to_cpu(hdr->dest_nid);
4299         dest_pid = le32_to_cpu(hdr->dest_pid);
4300         payload_length = le32_to_cpu(hdr->payload_length);
4301
4302         for_me = (ni->ni_nid == dest_nid);
4303         cpt = lnet_cpt_of_nid(from_nid, ni);
4304
4305         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4306                 libcfs_nid2str(dest_nid),
4307                 libcfs_nid2str(ni->ni_nid),
4308                 libcfs_nid2str(src_nid),
4309                 lnet_msgtyp2str(type),
4310                 (for_me) ? "for me" : "routed");
4311
4312         switch (type) {
4313         case LNET_MSG_ACK:
4314         case LNET_MSG_GET:
4315                 if (payload_length > 0) {
4316                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4317                                libcfs_nid2str(from_nid),
4318                                libcfs_nid2str(src_nid),
4319                                lnet_msgtyp2str(type), payload_length);
4320                         return -EPROTO;
4321                 }
4322                 break;
4323
4324         case LNET_MSG_PUT:
4325         case LNET_MSG_REPLY:
4326                 if (payload_length >
4327                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4328                         CERROR("%s, src %s: bad %s payload %d "
4329                                "(%d max expected)\n",
4330                                libcfs_nid2str(from_nid),
4331                                libcfs_nid2str(src_nid),
4332                                lnet_msgtyp2str(type),
4333                                payload_length,
4334                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4335                         return -EPROTO;
4336                 }
4337                 break;
4338
4339         default:
4340                 CERROR("%s, src %s: Bad message type 0x%x\n",
4341                        libcfs_nid2str(from_nid),
4342                        libcfs_nid2str(src_nid), type);
4343                 return -EPROTO;
4344         }
4345
4346         if (the_lnet.ln_routing &&
4347             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4348                 lnet_ni_lock(ni);
4349                 spin_lock(&ni->ni_net->net_lock);
4350                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4351                 spin_unlock(&ni->ni_net->net_lock);
4352                 if (ni->ni_status != NULL &&
4353                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4354                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4355                         push = true;
4356                 }
4357                 lnet_ni_unlock(ni);
4358         }
4359
4360         if (push)
4361                 lnet_push_update_to_peers(1);
4362
4363         /* Regard a bad destination NID as a protocol error.  Senders should
4364          * know what they're doing; if they don't they're misconfigured, buggy
4365          * or malicious so we chop them off at the knees :) */
4366
4367         if (!for_me) {
4368                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4369                         /* should have gone direct */
4370                         CERROR("%s, src %s: Bad dest nid %s "
4371                                "(should have been sent direct)\n",
4372                                 libcfs_nid2str(from_nid),
4373                                 libcfs_nid2str(src_nid),
4374                                 libcfs_nid2str(dest_nid));
4375                         return -EPROTO;
4376                 }
4377
4378                 if (lnet_islocalnid(dest_nid)) {
4379                         /* dest is another local NI; sender should have used
4380                          * this node's NID on its own network */
4381                         CERROR("%s, src %s: Bad dest nid %s "
4382                                "(it's my nid but on a different network)\n",
4383                                 libcfs_nid2str(from_nid),
4384                                 libcfs_nid2str(src_nid),
4385                                 libcfs_nid2str(dest_nid));
4386                         return -EPROTO;
4387                 }
4388
4389                 if (rdma_req && type == LNET_MSG_GET) {
4390                         CERROR("%s, src %s: Bad optimized GET for %s "
4391                                "(final destination must be me)\n",
4392                                 libcfs_nid2str(from_nid),
4393                                 libcfs_nid2str(src_nid),
4394                                 libcfs_nid2str(dest_nid));
4395                         return -EPROTO;
4396                 }
4397
4398                 if (!the_lnet.ln_routing) {
4399                         CERROR("%s, src %s: Dropping message for %s "
4400                                "(routing not enabled)\n",
4401                                 libcfs_nid2str(from_nid),
4402                                 libcfs_nid2str(src_nid),
4403                                 libcfs_nid2str(dest_nid));
4404                         goto drop;
4405                 }
4406         }
4407
4408         /* Message looks OK; we're not going to return an error, so we MUST
4409          * call back lnd_recv() come what may... */
4410
4411         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4412             fail_peer(src_nid, 0)) {                    /* shall we now? */
4413                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4414                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4415                        lnet_msgtyp2str(type));
4416                 goto drop;
4417         }
4418
4419         if (!list_empty(&the_lnet.ln_drop_rules) &&
4420             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4421                 CDEBUG(D_NET,
4422                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4423                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4424                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4425                 goto drop;
4426         }
4427
4428         if (lnet_drop_asym_route && for_me &&
4429             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4430                 struct lnet_net *net;
4431                 struct lnet_remotenet *rnet;
4432                 bool found = true;
4433
4434                 /* we are dealing with a routed message,
4435                  * so see if route to reach src_nid goes through from_nid
4436                  */
4437                 lnet_net_lock(cpt);
4438                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4439                 if (!net) {
4440                         lnet_net_unlock(cpt);
4441                         CERROR("net %s not found\n",
4442                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4443                         return -EPROTO;
4444                 }
4445
4446                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4447                 if (rnet) {
4448                         struct lnet_peer *gw = NULL;
4449                         struct lnet_peer_ni *lpni = NULL;
4450                         struct lnet_route *route;
4451
4452                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4453                                 found = false;
4454                                 gw = route->lr_gateway;
4455                                 if (route->lr_lnet != net->net_id)
4456                                         continue;
4457                                 /*
4458                                  * if the nid is one of the gateway's NIDs
4459                                  * then this is a valid gateway
4460                                  */
4461                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4462                                                 NULL, lpni)) != NULL) {
4463                                         if (lpni->lpni_nid == from_nid) {
4464                                                 found = true;
4465                                                 break;
4466                                         }
4467                                 }
4468                         }
4469                 }
4470                 lnet_net_unlock(cpt);
4471                 if (!found) {
4472                         /* we would not use from_nid to route a message to
4473                          * src_nid
4474                          * => asymmetric routing detected but forbidden
4475                          */
4476                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4477                                libcfs_nid2str(from_nid),
4478                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4479                         goto drop;
4480                 }
4481         }
4482
4483         msg = lnet_msg_alloc();
4484         if (msg == NULL) {
4485                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4486                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4487                        lnet_msgtyp2str(type));
4488                 goto drop;
4489         }
4490
4491         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4492          * pointers NULL etc */
4493
4494         msg->msg_type = type;
4495         msg->msg_private = private;
4496         msg->msg_receiving = 1;
4497         msg->msg_rdma_get = rdma_req;
4498         msg->msg_len = msg->msg_wanted = payload_length;
4499         msg->msg_offset = 0;
4500         msg->msg_hdr = *hdr;
4501         /* for building message event */
4502         msg->msg_from = from_nid;
4503         if (!for_me) {
4504                 msg->msg_target.pid     = dest_pid;
4505                 msg->msg_target.nid     = dest_nid;
4506                 msg->msg_routing        = 1;
4507
4508         } else {
4509                 /* convert common msg->hdr fields to host byteorder */
4510                 msg->msg_hdr.type       = type;
4511                 msg->msg_hdr.src_nid    = src_nid;
4512                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4513                 msg->msg_hdr.dest_nid   = dest_nid;
4514                 msg->msg_hdr.dest_pid   = dest_pid;
4515                 msg->msg_hdr.payload_length = payload_length;
4516         }
4517
4518         lnet_net_lock(cpt);
4519         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4520         if (IS_ERR(lpni)) {
4521                 lnet_net_unlock(cpt);
4522                 CERROR("%s, src %s: Dropping %s "
4523                        "(error %ld looking up sender)\n",
4524                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4525                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4526                 lnet_msg_free(msg);
4527                 if (rc == -ESHUTDOWN)
4528                         /* We are shutting down.  Don't do anything more */
4529                         return 0;
4530                 goto drop;
4531         }
4532
4533         if (the_lnet.ln_routing)
4534                 lpni->lpni_last_alive = ktime_get_seconds();
4535
4536         msg->msg_rxpeer = lpni;
4537         msg->msg_rxni = ni;
4538         lnet_ni_addref_locked(ni, cpt);
4539         /* Multi-Rail: Primary NID of source. */
4540         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4541
4542         /*
4543          * mark the status of this lpni as UP since we received a message
4544          * from it. The ping response reports back the ns_status which is
4545          * marked on the remote as up or down and we cache it here.
4546          */
4547         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4548
4549         lnet_msg_commit(msg, cpt);
4550
4551         /* message delay simulation */
4552         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4553                      lnet_delay_rule_match_locked(hdr, msg))) {
4554                 lnet_net_unlock(cpt);
4555                 return 0;
4556         }
4557
4558         if (!for_me) {
4559                 rc = lnet_parse_forward_locked(ni, msg);
4560                 lnet_net_unlock(cpt);
4561
4562                 if (rc < 0)
4563                         goto free_drop;
4564
4565                 if (rc == LNET_CREDIT_OK) {
4566                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4567                                      0, payload_length, payload_length);
4568                 }
4569                 return 0;
4570         }
4571
4572         lnet_net_unlock(cpt);
4573
4574         rc = lnet_parse_local(ni, msg);
4575         if (rc != 0)
4576                 goto free_drop;
4577         return 0;
4578
4579  free_drop:
4580         LASSERT(msg->msg_md == NULL);
4581         lnet_finalize(msg, rc);
4582
4583  drop:
4584         lnet_drop_message(ni, cpt, private, payload_length, type);
4585         return 0;
4586 }
4587 EXPORT_SYMBOL(lnet_parse);
4588
4589 void
4590 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4591 {
4592         while (!list_empty(head)) {
4593                 struct lnet_process_id id = {0};
4594                 struct lnet_msg *msg;
4595
4596                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4597                 list_del(&msg->msg_list);
4598
4599                 id.nid = msg->msg_hdr.src_nid;
4600                 id.pid = msg->msg_hdr.src_pid;
4601
4602                 LASSERT(msg->msg_md == NULL);
4603                 LASSERT(msg->msg_rx_delayed);
4604                 LASSERT(msg->msg_rxpeer != NULL);
4605                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4606
4607                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4608                       " offset %d length %d: %s\n",
4609                       libcfs_id2str(id),
4610                       msg->msg_hdr.msg.put.ptl_index,
4611                       msg->msg_hdr.msg.put.match_bits,
4612                       msg->msg_hdr.msg.put.offset,
4613                       msg->msg_hdr.payload_length, reason);
4614
4615                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4616                  * called lnet_drop_message(), so I just hang onto msg as well
4617                  * until that's done */
4618
4619                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4620                                   msg->msg_private, msg->msg_len,
4621                                   msg->msg_type);
4622
4623                 msg->msg_no_resend = true;
4624                 /*
4625                  * NB: message will not generate event because w/o attached MD,
4626                  * but we still should give error code so lnet_msg_decommit()
4627                  * can skip counters operations and other checks.
4628                  */
4629                 lnet_finalize(msg, -ENOENT);
4630         }
4631 }
4632
4633 void
4634 lnet_recv_delayed_msg_list(struct list_head *head)
4635 {
4636         while (!list_empty(head)) {
4637                 struct lnet_msg *msg;
4638                 struct lnet_process_id id;
4639
4640                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4641                 list_del(&msg->msg_list);
4642
4643                 /* md won't disappear under me, since each msg
4644                  * holds a ref on it */
4645
4646                 id.nid = msg->msg_hdr.src_nid;
4647                 id.pid = msg->msg_hdr.src_pid;
4648
4649                 LASSERT(msg->msg_rx_delayed);
4650                 LASSERT(msg->msg_md != NULL);
4651                 LASSERT(msg->msg_rxpeer != NULL);
4652                 LASSERT(msg->msg_rxni != NULL);
4653                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4654
4655                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4656                        "match %llu offset %d length %d.\n",
4657                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4658                         msg->msg_hdr.msg.put.match_bits,
4659                         msg->msg_hdr.msg.put.offset,
4660                         msg->msg_hdr.payload_length);
4661
4662                 lnet_recv_put(msg->msg_rxni, msg);
4663         }
4664 }
4665
4666 static void
4667 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4668                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4669 {
4670         s64 timeout_ns;
4671         struct lnet_rsp_tracker *local_rspt;
4672
4673         /*
4674          * MD has a refcount taken by message so it's not going away.
4675          * The MD however can be looked up. We need to secure the access
4676          * to the md_rspt_ptr by taking the res_lock.
4677          * The rspt can be accessed without protection up to when it gets
4678          * added to the list.
4679          */
4680
4681         lnet_res_lock(cpt);
4682         local_rspt = md->md_rspt_ptr;
4683         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4684         if (local_rspt != NULL) {
4685                 /*
4686                  * we already have an rspt attached to the md, so we'll
4687                  * update the deadline on that one.
4688                  */
4689                 lnet_rspt_free(rspt, cpt);
4690         } else {
4691                 /* new md */
4692                 rspt->rspt_mdh = mdh;
4693                 rspt->rspt_cpt = cpt;
4694                 /* store the rspt so we can access it when we get the REPLY */
4695                 md->md_rspt_ptr = rspt;
4696                 local_rspt = rspt;
4697         }
4698         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4699
4700         /*
4701          * add to the list of tracked responses. It's added to tail of the
4702          * list in order to expire all the older entries first.
4703          */
4704         lnet_net_lock(cpt);
4705         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4706         lnet_net_unlock(cpt);
4707         lnet_res_unlock(cpt);
4708 }
4709
4710 /**
4711  * Initiate an asynchronous PUT operation.
4712  *
4713  * There are several events associated with a PUT: completion of the send on
4714  * the initiator node (LNET_EVENT_SEND), and when the send completes
4715  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4716  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4717  * used at the target node to indicate the completion of incoming data
4718  * delivery.
4719  *
4720  * The local events will be logged in the EQ associated with the MD pointed to
4721  * by \a mdh handle. Using a MD without an associated EQ results in these
4722  * events being discarded. In this case, the caller must have another
4723  * mechanism (e.g., a higher level protocol) for determining when it is safe
4724  * to modify the memory region associated with the MD.
4725  *
4726  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4727  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4728  *
4729  * \param self Indicates the NID of a local interface through which to send
4730  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4731  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4732  * must be "free floating" (See LNetMDBind()).
4733  * \param ack Controls whether an acknowledgment is requested.
4734  * Acknowledgments are only sent when they are requested by the initiating
4735  * process and the target MD enables them.
4736  * \param target A process identifier for the target process.
4737  * \param portal The index in the \a target's portal table.
4738  * \param match_bits The match bits to use for MD selection at the target
4739  * process.
4740  * \param offset The offset into the target MD (only used when the target
4741  * MD has the LNET_MD_MANAGE_REMOTE option set).
4742  * \param hdr_data 64 bits of user data that can be included in the message
4743  * header. This data is written to an event queue entry at the target if an
4744  * EQ is present on the matching MD.
4745  *
4746  * \retval  0      Success, and only in this case events will be generated
4747  * and logged to EQ (if it exists).
4748  * \retval -EIO    Simulated failure.
4749  * \retval -ENOMEM Memory allocation failure.
4750  * \retval -ENOENT Invalid MD object.
4751  *
4752  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4753  */
4754 int
4755 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4756         struct lnet_process_id target, unsigned int portal,
4757         __u64 match_bits, unsigned int offset,
4758         __u64 hdr_data)
4759 {
4760         struct lnet_msg *msg;
4761         struct lnet_libmd *md;
4762         int cpt;
4763         int rc;
4764         struct lnet_rsp_tracker *rspt = NULL;
4765
4766         LASSERT(the_lnet.ln_refcount > 0);
4767
4768         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4769             fail_peer(target.nid, 1)) {                 /* shall we now? */
4770                 CERROR("Dropping PUT to %s: simulated failure\n",
4771                        libcfs_id2str(target));
4772                 return -EIO;
4773         }
4774
4775         msg = lnet_msg_alloc();
4776         if (msg == NULL) {
4777                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4778                        libcfs_id2str(target));
4779                 return -ENOMEM;
4780         }
4781         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4782
4783         cpt = lnet_cpt_of_cookie(mdh.cookie);
4784
4785         if (ack == LNET_ACK_REQ) {
4786                 rspt = lnet_rspt_alloc(cpt);
4787                 if (!rspt) {
4788                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4789                                 libcfs_id2str(target));
4790                         return -ENOMEM;
4791                 }
4792                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4793         }
4794
4795         lnet_res_lock(cpt);
4796
4797         md = lnet_handle2md(&mdh);
4798         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4799                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4800                        match_bits, portal, libcfs_id2str(target),
4801                        md == NULL ? -1 : md->md_threshold);
4802                 if (md != NULL && md->md_me != NULL)
4803                         CERROR("Source MD also attached to portal %d\n",
4804                                md->md_me->me_portal);
4805                 lnet_res_unlock(cpt);
4806
4807                 lnet_rspt_free(rspt, cpt);
4808                 lnet_msg_free(msg);
4809                 return -ENOENT;
4810         }
4811
4812         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4813
4814         lnet_msg_attach_md(msg, md, 0, 0);
4815
4816         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4817
4818         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4819         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4820         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4821         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4822
4823         /* NB handles only looked up by creator (no flips) */
4824         if (ack == LNET_ACK_REQ) {
4825                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4826                         the_lnet.ln_interface_cookie;
4827                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4828                         md->md_lh.lh_cookie;
4829         } else {
4830                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4831                         LNET_WIRE_HANDLE_COOKIE_NONE;
4832                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4833                         LNET_WIRE_HANDLE_COOKIE_NONE;
4834         }
4835
4836         lnet_res_unlock(cpt);
4837
4838         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4839
4840         if (ack == LNET_ACK_REQ)
4841                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4842
4843         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4844                                  CFS_FAIL_ONCE))
4845                 rc = -EIO;
4846         else
4847                 rc = lnet_send(self, msg, LNET_NID_ANY);
4848
4849         if (rc != 0) {
4850                 CNETERR("Error sending PUT to %s: %d\n",
4851                         libcfs_id2str(target), rc);
4852                 msg->msg_no_resend = true;
4853                 lnet_finalize(msg, rc);
4854         }
4855
4856         /* completion will be signalled by an event */
4857         return 0;
4858 }
4859 EXPORT_SYMBOL(LNetPut);
4860
4861 /*
4862  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4863  * returns a msg for the LND to pass to lnet_finalize() when the sink
4864  * data has been received.
4865  *
4866  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4867  * lnet_finalize() is called on it, so the LND must call this first
4868  */
4869 struct lnet_msg *
4870 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4871 {
4872         struct lnet_msg *msg = lnet_msg_alloc();
4873         struct lnet_libmd *getmd = getmsg->msg_md;
4874         struct lnet_process_id peer_id = getmsg->msg_target;
4875         int cpt;
4876
4877         LASSERT(!getmsg->msg_target_is_router);
4878         LASSERT(!getmsg->msg_routing);
4879
4880         if (msg == NULL) {
4881                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4882                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4883                 goto drop;
4884         }
4885
4886         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4887         lnet_res_lock(cpt);
4888
4889         LASSERT(getmd->md_refcount > 0);
4890
4891         if (getmd->md_threshold == 0) {
4892                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4893                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4894                         getmd);
4895                 lnet_res_unlock(cpt);
4896                 goto drop;
4897         }
4898
4899         LASSERT(getmd->md_offset == 0);
4900
4901         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4902                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4903
4904         /* setup information for lnet_build_msg_event */
4905         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4906         msg->msg_from = peer_id.nid;
4907         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4908         msg->msg_hdr.src_nid = peer_id.nid;
4909         msg->msg_hdr.payload_length = getmd->md_length;
4910         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4911
4912         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4913         lnet_res_unlock(cpt);
4914
4915         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4916
4917         lnet_net_lock(cpt);
4918         lnet_msg_commit(msg, cpt);
4919         lnet_net_unlock(cpt);
4920
4921         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4922
4923         return msg;
4924
4925  drop:
4926         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4927
4928         lnet_net_lock(cpt);
4929         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4930         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4931         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4932                 getmd->md_length;
4933         lnet_net_unlock(cpt);
4934
4935         if (msg != NULL)
4936                 lnet_msg_free(msg);
4937
4938         return NULL;
4939 }
4940 EXPORT_SYMBOL(lnet_create_reply_msg);
4941
4942 void
4943 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4944                        unsigned int len)
4945 {
4946         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4947          * completed and I know it. */
4948         LASSERT(reply != NULL);
4949         LASSERT(reply->msg_type == LNET_MSG_GET);
4950         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4951
4952         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4953          * the end of my buffer, I might as well be dead. */
4954         LASSERT(len <= reply->msg_ev.mlength);
4955
4956         reply->msg_ev.mlength = len;
4957 }
4958 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4959
4960 /**
4961  * Initiate an asynchronous GET operation.
4962  *
4963  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4964  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4965  * the target node in the REPLY has been written to local MD.
4966  *
4967  * On the target node, an LNET_EVENT_GET is logged when the GET request
4968  * arrives and is accepted into a MD.
4969  *
4970  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4971  * \param mdh A handle for the MD that describes the memory into which the
4972  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4973  *
4974  * \retval  0      Success, and only in this case events will be generated
4975  * and logged to EQ (if it exists) of the MD.
4976  * \retval -EIO    Simulated failure.
4977  * \retval -ENOMEM Memory allocation failure.
4978  * \retval -ENOENT Invalid MD object.
4979  */
4980 int
4981 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4982         struct lnet_process_id target, unsigned int portal,
4983         __u64 match_bits, unsigned int offset, bool recovery)
4984 {
4985         struct lnet_msg *msg;
4986         struct lnet_libmd *md;
4987         struct lnet_rsp_tracker *rspt;
4988         int cpt;
4989         int rc;
4990
4991         LASSERT(the_lnet.ln_refcount > 0);
4992
4993         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4994             fail_peer(target.nid, 1))                   /* shall we now? */
4995         {
4996                 CERROR("Dropping GET to %s: simulated failure\n",
4997                        libcfs_id2str(target));
4998                 return -EIO;
4999         }
5000
5001         msg = lnet_msg_alloc();
5002         if (!msg) {
5003                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5004                        libcfs_id2str(target));
5005                 return -ENOMEM;
5006         }
5007
5008         cpt = lnet_cpt_of_cookie(mdh.cookie);
5009
5010         rspt = lnet_rspt_alloc(cpt);
5011         if (!rspt) {
5012                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5013                        libcfs_id2str(target));
5014                 return -ENOMEM;
5015         }
5016         INIT_LIST_HEAD(&rspt->rspt_on_list);
5017
5018         msg->msg_recovery = recovery;
5019
5020         lnet_res_lock(cpt);
5021
5022         md = lnet_handle2md(&mdh);
5023         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5024                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5025                        match_bits, portal, libcfs_id2str(target),
5026                        md == NULL ? -1 : md->md_threshold);
5027                 if (md != NULL && md->md_me != NULL)
5028                         CERROR("REPLY MD also attached to portal %d\n",
5029                                md->md_me->me_portal);
5030
5031                 lnet_res_unlock(cpt);
5032
5033                 lnet_msg_free(msg);
5034                 lnet_rspt_free(rspt, cpt);
5035                 return -ENOENT;
5036         }
5037
5038         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5039
5040         lnet_msg_attach_md(msg, md, 0, 0);
5041
5042         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5043
5044         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5045         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5046         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5047         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5048
5049         /* NB handles only looked up by creator (no flips) */
5050         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5051                 the_lnet.ln_interface_cookie;
5052         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5053                 md->md_lh.lh_cookie;
5054
5055         lnet_res_unlock(cpt);
5056
5057         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5058
5059         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5060
5061         rc = lnet_send(self, msg, LNET_NID_ANY);
5062         if (rc < 0) {
5063                 CNETERR("Error sending GET to %s: %d\n",
5064                         libcfs_id2str(target), rc);
5065                 msg->msg_no_resend = true;
5066                 lnet_finalize(msg, rc);
5067         }
5068
5069         /* completion will be signalled by an event */
5070         return 0;
5071 }
5072 EXPORT_SYMBOL(LNetGet);
5073
5074 /**
5075  * Calculate distance to node at \a dstnid.
5076  *
5077  * \param dstnid Target NID.
5078  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5079  * is saved here.
5080  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5081  * here.
5082  *
5083  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5084  * local_nid_dist_zero is set, which is the default.
5085  * \retval positives Distance to target NID, i.e. number of hops plus one.
5086  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5087  */
5088 int
5089 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5090 {
5091         struct list_head *e;
5092         struct lnet_ni *ni = NULL;
5093         struct lnet_remotenet *rnet;
5094         __u32 dstnet = LNET_NIDNET(dstnid);
5095         int hops;
5096         int cpt;
5097         __u32 order = 2;
5098         struct list_head *rn_list;
5099
5100         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5101          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5102          * keep order 0 free for 0@lo and order 1 free for a local NID
5103          * match */
5104
5105         LASSERT(the_lnet.ln_refcount > 0);
5106
5107         cpt = lnet_net_lock_current();
5108
5109         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5110                 if (ni->ni_nid == dstnid) {
5111                         if (srcnidp != NULL)
5112                                 *srcnidp = dstnid;
5113                         if (orderp != NULL) {
5114                                 if (dstnid == LNET_NID_LO_0)
5115                                         *orderp = 0;
5116                                 else
5117                                         *orderp = 1;
5118                         }
5119                         lnet_net_unlock(cpt);
5120
5121                         return local_nid_dist_zero ? 0 : 1;
5122                 }
5123
5124                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5125                         /* Check if ni was originally created in
5126                          * current net namespace.
5127                          * If not, assign order above 0xffff0000,
5128                          * to make this ni not a priority. */
5129                         if (current->nsproxy &&
5130                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5131                                         order += 0xffff0000;
5132                         if (srcnidp != NULL)
5133                                 *srcnidp = ni->ni_nid;
5134                         if (orderp != NULL)
5135                                 *orderp = order;
5136                         lnet_net_unlock(cpt);
5137                         return 1;
5138                 }
5139
5140                 order++;
5141         }
5142
5143         rn_list = lnet_net2rnethash(dstnet);
5144         list_for_each(e, rn_list) {
5145                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5146
5147                 if (rnet->lrn_net == dstnet) {
5148                         struct lnet_route *route;
5149                         struct lnet_route *shortest = NULL;
5150                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5151                         __u32 route_hops;
5152
5153                         LASSERT(!list_empty(&rnet->lrn_routes));
5154
5155                         list_for_each_entry(route, &rnet->lrn_routes,
5156                                             lr_list) {
5157                                 route_hops = route->lr_hops;
5158                                 if (route_hops == LNET_UNDEFINED_HOPS)
5159                                         route_hops = 1;
5160                                 if (shortest == NULL ||
5161                                     route_hops < shortest_hops) {
5162                                         shortest = route;
5163                                         shortest_hops = route_hops;
5164                                 }
5165                         }
5166
5167                         LASSERT(shortest != NULL);
5168                         hops = shortest_hops;
5169                         if (srcnidp != NULL) {
5170                                 struct lnet_net *net;
5171                                 net = lnet_get_net_locked(shortest->lr_lnet);
5172                                 LASSERT(net);
5173                                 ni = lnet_get_next_ni_locked(net, NULL);
5174                                 *srcnidp = ni->ni_nid;
5175                         }
5176                         if (orderp != NULL)
5177                                 *orderp = order;
5178                         lnet_net_unlock(cpt);
5179                         return hops + 1;
5180                 }
5181                 order++;
5182         }
5183
5184         lnet_net_unlock(cpt);
5185         return -EHOSTUNREACH;
5186 }
5187 EXPORT_SYMBOL(LNetDist);