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         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
1985                 (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
1986                 LNET_NID_ANY;
1987
1988         CDEBUG(D_NET, "using src nid %s for route restriction\n",
1989                libcfs_nid2str(src_nid));
1990
1991         /* If a router nid was specified then we are replying to a GET or
1992          * sending an ACK. In this case we use the gateway associated with the
1993          * specified router nid.
1994          */
1995         if (sd->sd_rtr_nid != LNET_NID_ANY) {
1996                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
1997                 if (!gwni) {
1998                         CERROR("No peer NI for gateway %s\n",
1999                                libcfs_nid2str(sd->sd_rtr_nid));
2000                         return -EHOSTUNREACH;
2001                 }
2002                 gw = gwni->lpni_peer_net->lpn_peer;
2003                 lnet_peer_ni_decref_locked(gwni);
2004                 local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
2005         } else {
2006                 /* we've already looked up the initial lpni using dst_nid */
2007                 lpni = sd->sd_best_lpni;
2008                 /* the peer tree must be in existence */
2009                 LASSERT(lpni && lpni->lpni_peer_net &&
2010                         lpni->lpni_peer_net->lpn_peer);
2011                 lp = lpni->lpni_peer_net->lpn_peer;
2012
2013                 list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2014                         /* is this remote network reachable?  */
2015                         rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2016                         if (!rnet)
2017                                 continue;
2018
2019                         if (!best_lpn) {
2020                                 best_lpn = lpn;
2021                                 best_rnet = rnet;
2022                         }
2023
2024                         if (best_lpn->lpn_seq <= lpn->lpn_seq)
2025                                 continue;
2026
2027                         best_lpn = lpn;
2028                         best_rnet = rnet;
2029                 }
2030
2031                 if (!best_lpn) {
2032                         CERROR("peer %s has no available nets\n",
2033                                libcfs_nid2str(sd->sd_dst_nid));
2034                         return -EHOSTUNREACH;
2035                 }
2036
2037                 sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2038                                                        sd->sd_dst_nid,
2039                                                        lp,
2040                                                        best_lpn->lpn_net_id);
2041                 if (!sd->sd_best_lpni) {
2042                         CERROR("peer %s is unreachable\n",
2043                                libcfs_nid2str(sd->sd_dst_nid));
2044                         return -EHOSTUNREACH;
2045                 }
2046
2047                 /*
2048                  * We're attempting to round robin over the remote peer
2049                  * NI's so update the final destination we selected
2050                  */
2051                 sd->sd_final_dst_lpni = sd->sd_best_lpni;
2052
2053                 /*
2054                  * find the best route. Restrict the selection on the net of the
2055                  * local NI if we've already picked the local NI to send from.
2056                  * Otherwise, let's pick any route we can find and then find
2057                  * a local NI we can reach the route's gateway on. Any route we select
2058                  * will be reachable by virtue of the restriction we have when
2059                  * adding a route.
2060                  */
2061                 best_route = lnet_find_route_locked(best_rnet,
2062                                                     LNET_NIDNET(src_nid),
2063                                                     &last_route, &gwni);
2064
2065                 if (!best_route) {
2066                         CERROR("no route to %s from %s\n",
2067                                libcfs_nid2str(dst_nid),
2068                                libcfs_nid2str(src_nid));
2069                         return -EHOSTUNREACH;
2070                 }
2071
2072                 if (!gwni) {
2073                         CERROR("Internal Error. Route expected to %s from %s\n",
2074                                libcfs_nid2str(dst_nid),
2075                                libcfs_nid2str(src_nid));
2076                         return -EFAULT;
2077                 }
2078
2079                 gw = best_route->lr_gateway;
2080                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2081                 local_lnet = best_route->lr_lnet;
2082
2083                 /*
2084                  * Increment the sequence number of the remote lpni so we
2085                  * can round robin over the different interfaces of the
2086                  * remote lpni
2087                  */
2088                 sd->sd_best_lpni->lpni_seq++;
2089         }
2090
2091         /*
2092          * Discover this gateway if it hasn't already been discovered.
2093          * This means we might delay the message until discovery has
2094          * completed
2095          */
2096         sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
2097         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2098         if (rc)
2099                 return rc;
2100
2101         if (!sd->sd_best_ni)
2102                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
2103                                         lnet_peer_get_net_locked(gw,
2104                                                                  local_lnet),
2105                                         sd->sd_md_cpt,
2106                                         true);
2107
2108         if (!sd->sd_best_ni) {
2109                 CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
2110                        libcfs_net2str(local_lnet),
2111                        libcfs_nid2str(sd->sd_src_nid));
2112                 return -EFAULT;
2113         }
2114
2115         *gw_lpni = gwni;
2116         *gw_peer = gw;
2117
2118         /*
2119          * increment the sequence numbers since now we're sure we're
2120          * going to use this path
2121          */
2122         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2123                 LASSERT(best_route && last_route);
2124                 best_route->lr_seq = last_route->lr_seq + 1;
2125                 best_lpn->lpn_seq++;
2126         }
2127
2128         return 0;
2129 }
2130
2131 /*
2132  * Handle two cases:
2133  *
2134  * Case 1:
2135  *  Source specified
2136  *  Remote destination
2137  *  Non-MR destination
2138  *
2139  * Case 2:
2140  *  Source specified
2141  *  Remote destination
2142  *  MR destination
2143  *
2144  * The handling of these two cases is similar. Even though the destination
2145  * can be MR or non-MR, we'll deal directly with the router.
2146  */
2147 static int
2148 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2149 {
2150         int rc;
2151         struct lnet_peer_ni *gw_lpni = NULL;
2152         struct lnet_peer *gw_peer = NULL;
2153
2154         /* find local NI */
2155         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2156         if (!sd->sd_best_ni) {
2157                 CERROR("Can't send to %s: src %s is not a "
2158                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2159                                 libcfs_nid2str(sd->sd_src_nid));
2160                 return -EINVAL;
2161         }
2162
2163         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2164                                      &gw_peer);
2165         if (rc)
2166                 return rc;
2167
2168         if (sd->sd_send_case & NMR_DST)
2169                 /*
2170                  * since the final destination is non-MR let's set its preferred
2171                  * NID before we send
2172                  */
2173                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2174                                          sd->sd_msg);
2175
2176         /*
2177          * We're going to send to the gw found so let's set its
2178          * info
2179          */
2180         sd->sd_peer = gw_peer;
2181         sd->sd_best_lpni = gw_lpni;
2182
2183         return lnet_handle_send(sd);
2184 }
2185
2186 struct lnet_ni *
2187 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2188                                bool discovery)
2189 {
2190         struct lnet_peer_net *peer_net = NULL;
2191         struct lnet_ni *best_ni = NULL;
2192         int lpn_healthv = 0;
2193
2194         /*
2195          * The peer can have multiple interfaces, some of them can be on
2196          * the local network and others on a routed network. We should
2197          * prefer the local network. However if the local network is not
2198          * available then we need to try the routed network
2199          */
2200
2201         /* go through all the peer nets and find the best_ni */
2202         list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
2203                 /*
2204                  * The peer's list of nets can contain non-local nets. We
2205                  * want to only examine the local ones.
2206                  */
2207                 if (!lnet_get_net_locked(peer_net->lpn_net_id))
2208                         continue;
2209
2210                 /* always select the lpn with the best health */
2211                 if (lpn_healthv <= peer_net->lpn_healthv)
2212                         lpn_healthv = peer_net->lpn_healthv;
2213                 else
2214                         continue;
2215
2216                 best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer, peer_net,
2217                                                         md_cpt, false);
2218
2219                 /*
2220                  * if this is a discovery message and lp_disc_net_id is
2221                  * specified then use that net to send the discovery on.
2222                  */
2223                 if (peer->lp_disc_net_id == peer_net->lpn_net_id &&
2224                     discovery)
2225                         break;
2226         }
2227
2228         if (best_ni)
2229                 /* increment sequence number so we can round robin */
2230                 best_ni->ni_seq++;
2231
2232         return best_ni;
2233 }
2234
2235 static struct lnet_ni *
2236 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2237 {
2238         struct lnet_ni *best_ni = NULL;
2239         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2240         struct lnet_peer_ni *lpni_entry;
2241
2242         /*
2243          * We must use a consistent source address when sending to a
2244          * non-MR peer. However, a non-MR peer can have multiple NIDs
2245          * on multiple networks, and we may even need to talk to this
2246          * peer on multiple networks -- certain types of
2247          * load-balancing configuration do this.
2248          *
2249          * So we need to pick the NI the peer prefers for this
2250          * particular network.
2251          */
2252         LASSERT(peer_net);
2253         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2254                             lpni_peer_nis) {
2255                 if (lpni_entry->lpni_pref_nnids == 0)
2256                         continue;
2257                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2258                 best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
2259                 break;
2260         }
2261
2262         return best_ni;
2263 }
2264
2265 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2266 static int
2267 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2268 {
2269         struct lnet_ni *best_ni = NULL;
2270         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2271
2272         /*
2273          * We must use a consistent source address when sending to a
2274          * non-MR peer. However, a non-MR peer can have multiple NIDs
2275          * on multiple networks, and we may even need to talk to this
2276          * peer on multiple networks -- certain types of
2277          * load-balancing configuration do this.
2278          *
2279          * So we need to pick the NI the peer prefers for this
2280          * particular network.
2281          */
2282
2283         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2284                                                        sd->sd_cpt);
2285
2286         /* if best_ni is still not set just pick one */
2287         if (!best_ni) {
2288                 best_ni =
2289                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2290                                                 sd->sd_best_lpni->lpni_peer_net,
2291                                                 sd->sd_md_cpt, true);
2292                 /* If there is no best_ni we don't have a route */
2293                 if (!best_ni) {
2294                         CERROR("no path to %s from net %s\n",
2295                                 libcfs_nid2str(best_lpni->lpni_nid),
2296                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2297                         return -EHOSTUNREACH;
2298                 }
2299         }
2300
2301         sd->sd_best_ni = best_ni;
2302
2303         /* Set preferred NI if necessary. */
2304         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2305
2306         return 0;
2307 }
2308
2309
2310 /*
2311  * Source not specified
2312  * Local destination
2313  * Non-MR Peer
2314  *
2315  * always use the same source NID for NMR peers
2316  * If we've talked to that peer before then we already have a preferred
2317  * source NI associated with it. Otherwise, we select a preferred local NI
2318  * and store it in the peer
2319  */
2320 static int
2321 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2322 {
2323         int rc;
2324
2325         /* sd->sd_best_lpni is already set to the final destination */
2326
2327         /*
2328          * At this point we should've created the peer ni and peer. If we
2329          * can't find it, then something went wrong. Instead of assert
2330          * output a relevant message and fail the send
2331          */
2332         if (!sd->sd_best_lpni) {
2333                 CERROR("Internal fault. Unable to send msg %s to %s. "
2334                        "NID not known\n",
2335                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2336                        libcfs_nid2str(sd->sd_dst_nid));
2337                 return -EFAULT;
2338         }
2339
2340         rc = lnet_select_preferred_best_ni(sd);
2341         if (!rc)
2342                 rc = lnet_handle_send(sd);
2343
2344         return rc;
2345 }
2346
2347 static int
2348 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2349 {
2350         /*
2351          * NOTE we've already handled the remote peer case. So we only
2352          * need to worry about the local case here.
2353          *
2354          * if we're sending a response, ACK or reply, we need to send it
2355          * to the destination NID given to us. At this point we already
2356          * have the peer_ni we're suppose to send to, so just find the
2357          * best_ni on the peer net and use that. Since we're sending to an
2358          * MR peer then we can just run the selection algorithm on our
2359          * local NIs and pick the best one.
2360          */
2361         if (sd->sd_send_case & SND_RESP) {
2362                 sd->sd_best_ni =
2363                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2364                                                 sd->sd_best_lpni->lpni_peer_net,
2365                                                 sd->sd_md_cpt, true);
2366
2367                 if (!sd->sd_best_ni) {
2368                         /*
2369                          * We're not going to deal with not able to send
2370                          * a response to the provided final destination
2371                          */
2372                         CERROR("Can't send response to %s. "
2373                                "No local NI available\n",
2374                                 libcfs_nid2str(sd->sd_dst_nid));
2375                         return -EHOSTUNREACH;
2376                 }
2377
2378                 return lnet_handle_send(sd);
2379         }
2380
2381         /*
2382          * If we get here that means we're sending a fresh request, PUT or
2383          * GET, so we need to run our standard selection algorithm.
2384          * First find the best local interface that's on any of the peer's
2385          * networks.
2386          */
2387         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2388                                         sd->sd_md_cpt,
2389                                         lnet_msg_discovery(sd->sd_msg));
2390         if (sd->sd_best_ni) {
2391                 sd->sd_best_lpni =
2392                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2393                                       sd->sd_peer,
2394                                       sd->sd_best_ni->ni_net->net_id);
2395
2396                 /*
2397                  * if we're successful in selecting a peer_ni on the local
2398                  * network, then send to it. Otherwise fall through and
2399                  * try and see if we can reach it over another routed
2400                  * network
2401                  */
2402                 if (sd->sd_best_lpni &&
2403                     sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid) {
2404                         /*
2405                          * in case we initially started with a routed
2406                          * destination, let's reset to local
2407                          */
2408                         sd->sd_send_case &= ~REMOTE_DST;
2409                         sd->sd_send_case |= LOCAL_DST;
2410                         return lnet_handle_lo_send(sd);
2411                 } else if (sd->sd_best_lpni) {
2412                         /*
2413                          * in case we initially started with a routed
2414                          * destination, let's reset to local
2415                          */
2416                         sd->sd_send_case &= ~REMOTE_DST;
2417                         sd->sd_send_case |= LOCAL_DST;
2418                         return lnet_handle_send(sd);
2419                 }
2420
2421                 CERROR("Internal Error. Expected to have a best_lpni: "
2422                        "%s -> %s\n",
2423                        libcfs_nid2str(sd->sd_src_nid),
2424                        libcfs_nid2str(sd->sd_dst_nid));
2425
2426                 return -EFAULT;
2427         }
2428
2429         /*
2430          * Peer doesn't have a local network. Let's see if there is
2431          * a remote network we can reach it on.
2432          */
2433         return PASS_THROUGH;
2434 }
2435
2436 /*
2437  * Case 1:
2438  *      Source NID not specified
2439  *      Local destination
2440  *      MR peer
2441  *
2442  * Case 2:
2443  *      Source NID not speified
2444  *      Remote destination
2445  *      MR peer
2446  *
2447  * In both of these cases if we're sending a response, ACK or REPLY, then
2448  * we need to send to the destination NID provided.
2449  *
2450  * In the remote case let's deal with MR routers.
2451  *
2452  */
2453
2454 static int
2455 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2456 {
2457         int rc = 0;
2458         struct lnet_peer *gw_peer = NULL;
2459         struct lnet_peer_ni *gw_lpni = NULL;
2460
2461         /*
2462          * handle sending a response to a remote peer here so we don't
2463          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2464          */
2465         if (sd->sd_send_case & REMOTE_DST &&
2466             sd->sd_send_case & SND_RESP) {
2467                 struct lnet_peer_ni *gw;
2468                 struct lnet_peer *gw_peer;
2469
2470                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2471                                                   &gw_peer);
2472                 if (rc < 0) {
2473                         CERROR("Can't send response to %s. "
2474                                "No route available\n",
2475                                 libcfs_nid2str(sd->sd_dst_nid));
2476                         return -EHOSTUNREACH;
2477                 } else if (rc > 0) {
2478                         return rc;
2479                 }
2480
2481                 sd->sd_best_lpni = gw;
2482                 sd->sd_peer = gw_peer;
2483
2484                 return lnet_handle_send(sd);
2485         }
2486
2487         /*
2488          * Even though the NID for the peer might not be on a local network,
2489          * since the peer is MR there could be other interfaces on the
2490          * local network. In that case we'd still like to prefer the local
2491          * network over the routed network. If we're unable to do that
2492          * then we select the best router among the different routed networks,
2493          * and if the router is MR then we can deal with it as such.
2494          */
2495         rc = lnet_handle_any_mr_dsta(sd);
2496         if (rc != PASS_THROUGH)
2497                 return rc;
2498
2499         /*
2500          * Now that we must route to the destination, we must consider the
2501          * MR case, where the destination has multiple interfaces, some of
2502          * which we can route to and others we do not. For this reason we
2503          * need to select the destination which we can route to and if
2504          * there are multiple, we need to round robin.
2505          */
2506         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2507                                           &gw_peer);
2508         if (rc)
2509                 return rc;
2510
2511         sd->sd_send_case &= ~LOCAL_DST;
2512         sd->sd_send_case |= REMOTE_DST;
2513
2514         sd->sd_peer = gw_peer;
2515         sd->sd_best_lpni = gw_lpni;
2516
2517         return lnet_handle_send(sd);
2518 }
2519
2520 /*
2521  * Source not specified
2522  * Remote destination
2523  * Non-MR peer
2524  *
2525  * Must send to the specified peer NID using the same source NID that
2526  * we've used before. If it's the first time to talk to that peer then
2527  * find the source NI and assign it as preferred to that peer
2528  */
2529 static int
2530 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2531 {
2532         int rc;
2533         struct lnet_peer_ni *gw_lpni = NULL;
2534         struct lnet_peer *gw_peer = NULL;
2535
2536         /*
2537          * Let's see if we have a preferred NI to talk to this NMR peer
2538          */
2539         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2540                                                               sd->sd_cpt);
2541
2542         /*
2543          * find the router and that'll find the best NI if we didn't find
2544          * it already.
2545          */
2546         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2547                                           &gw_peer);
2548         if (rc)
2549                 return rc;
2550
2551         /*
2552          * set the best_ni we've chosen as the preferred one for
2553          * this peer
2554          */
2555         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2556
2557         /* we'll be sending to the gw */
2558         sd->sd_best_lpni = gw_lpni;
2559         sd->sd_peer = gw_peer;
2560
2561         return lnet_handle_send(sd);
2562 }
2563
2564 static int
2565 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2566 {
2567         /*
2568          * turn off the SND_RESP bit.
2569          * It will be checked in the case handling
2570          */
2571         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2572
2573         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2574                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2575                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2576                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2577                 libcfs_nid2str(sd->sd_dst_nid),
2578                 (send_case & LOCAL_DST) ? "local" : "routed");
2579
2580         switch (send_case) {
2581         /*
2582          * For all cases where the source is specified, we should always
2583          * use the destination NID, whether it's an MR destination or not,
2584          * since we're continuing a series of related messages for the
2585          * same RPC
2586          */
2587         case SRC_SPEC_LOCAL_NMR_DST:
2588                 return lnet_handle_spec_local_nmr_dst(sd);
2589         case SRC_SPEC_LOCAL_MR_DST:
2590                 return lnet_handle_spec_local_mr_dst(sd);
2591         case SRC_SPEC_ROUTER_NMR_DST:
2592         case SRC_SPEC_ROUTER_MR_DST:
2593                 return lnet_handle_spec_router_dst(sd);
2594         case SRC_ANY_LOCAL_NMR_DST:
2595                 return lnet_handle_any_local_nmr_dst(sd);
2596         case SRC_ANY_LOCAL_MR_DST:
2597         case SRC_ANY_ROUTER_MR_DST:
2598                 return lnet_handle_any_mr_dst(sd);
2599         case SRC_ANY_ROUTER_NMR_DST:
2600                 return lnet_handle_any_router_nmr_dst(sd);
2601         default:
2602                 CERROR("Unknown send case\n");
2603                 return -1;
2604         }
2605 }
2606
2607 static int
2608 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2609                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2610 {
2611         struct lnet_peer_ni *lpni;
2612         struct lnet_peer *peer;
2613         struct lnet_send_data send_data;
2614         int cpt, rc;
2615         int md_cpt;
2616         __u32 send_case = 0;
2617         bool final_hop;
2618         bool mr_forwarding_allowed;
2619
2620         memset(&send_data, 0, sizeof(send_data));
2621
2622         /*
2623          * get an initial CPT to use for locking. The idea here is not to
2624          * serialize the calls to select_pathway, so that as many
2625          * operations can run concurrently as possible. To do that we use
2626          * the CPT where this call is being executed. Later on when we
2627          * determine the CPT to use in lnet_message_commit, we switch the
2628          * lock and check if there was any configuration change.  If none,
2629          * then we proceed, if there is, then we restart the operation.
2630          */
2631         cpt = lnet_net_lock_current();
2632
2633         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2634         if (md_cpt == CFS_CPT_ANY)
2635                 md_cpt = cpt;
2636
2637 again:
2638
2639         /*
2640          * If we're being asked to send to the loopback interface, there
2641          * is no need to go through any selection. We can just shortcut
2642          * the entire process and send over lolnd
2643          */
2644         send_data.sd_msg = msg;
2645         send_data.sd_cpt = cpt;
2646         if (dst_nid == LNET_NID_LO_0) {
2647                 rc = lnet_handle_lo_send(&send_data);
2648                 lnet_net_unlock(cpt);
2649                 return rc;
2650         }
2651
2652         /*
2653          * find an existing peer_ni, or create one and mark it as having been
2654          * created due to network traffic. This call will create the
2655          * peer->peer_net->peer_ni tree.
2656          */
2657         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2658         if (IS_ERR(lpni)) {
2659                 lnet_net_unlock(cpt);
2660                 return PTR_ERR(lpni);
2661         }
2662
2663         /*
2664          * Cache the original src_nid and rtr_nid. If we need to resend the
2665          * message then we'll need to know whether the src_nid was originally
2666          * specified for this message. If it was originally specified,
2667          * then we need to keep using the same src_nid since it's
2668          * continuing the same sequence of messages. Similarly, rtr_nid will
2669          * affect our choice of next hop.
2670          */
2671         msg->msg_src_nid_param = src_nid;
2672         msg->msg_rtr_nid_param = rtr_nid;
2673
2674         /*
2675          * If necessary, perform discovery on the peer that owns this peer_ni.
2676          * Note, this can result in the ownership of this peer_ni changing
2677          * to another peer object.
2678          */
2679         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2680         if (rc) {
2681                 lnet_peer_ni_decref_locked(lpni);
2682                 lnet_net_unlock(cpt);
2683                 return rc;
2684         }
2685         lnet_peer_ni_decref_locked(lpni);
2686
2687         peer = lpni->lpni_peer_net->lpn_peer;
2688
2689         /*
2690          * Identify the different send cases
2691          */
2692         if (src_nid == LNET_NID_ANY)
2693                 send_case |= SRC_ANY;
2694         else
2695                 send_case |= SRC_SPEC;
2696
2697         if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2698                 send_case |= LOCAL_DST;
2699         else
2700                 send_case |= REMOTE_DST;
2701
2702         final_hop = false;
2703         if (msg->msg_routing && (send_case & LOCAL_DST))
2704                 final_hop = true;
2705
2706         /* Determine whether to allow MR forwarding for this message.
2707          * NB: MR forwarding is allowed if the message originator and the
2708          * destination are both MR capable, and the destination lpni that was
2709          * originally chosen by the originator is unhealthy or down.
2710          * We check the MR capability of the destination further below
2711          */
2712         mr_forwarding_allowed = false;
2713         if (final_hop) {
2714                 struct lnet_peer *src_lp;
2715                 struct lnet_peer_ni *src_lpni;
2716
2717                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
2718                                                   LNET_NID_ANY, cpt);
2719                 /* We don't fail the send if we hit any errors here. We'll just
2720                  * try to send it via non-multi-rail criteria
2721                  */
2722                 if (!IS_ERR(src_lpni)) {
2723                         src_lp = lpni->lpni_peer_net->lpn_peer;
2724                         if (lnet_peer_is_multi_rail(src_lp) &&
2725                             !lnet_is_peer_ni_alive(lpni))
2726                                 mr_forwarding_allowed = true;
2727
2728                 }
2729                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
2730                        mr_forwarding_allowed ? "allowed" : "not allowed");
2731         }
2732
2733         /*
2734          * Deal with the peer as NMR in the following cases:
2735          * 1. the peer is NMR
2736          * 2. We're trying to recover a specific peer NI
2737          * 3. I'm a router sending to the final destination and MR forwarding is
2738          *    not allowed for this message (as determined above).
2739          *    In this case the source of the message would've
2740          *    already selected the final destination so my job
2741          *    is to honor the selection.
2742          */
2743         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
2744             (final_hop && !mr_forwarding_allowed))
2745                 send_case |= NMR_DST;
2746         else
2747                 send_case |= MR_DST;
2748
2749         if (lnet_msg_is_response(msg))
2750                 send_case |= SND_RESP;
2751
2752         /* assign parameters to the send_data */
2753         send_data.sd_rtr_nid = rtr_nid;
2754         send_data.sd_src_nid = src_nid;
2755         send_data.sd_dst_nid = dst_nid;
2756         send_data.sd_best_lpni = lpni;
2757         /*
2758          * keep a pointer to the final destination in case we're going to
2759          * route, so we'll need to access it later
2760          */
2761         send_data.sd_final_dst_lpni = lpni;
2762         send_data.sd_peer = peer;
2763         send_data.sd_md_cpt = md_cpt;
2764         send_data.sd_send_case = send_case;
2765
2766         rc = lnet_handle_send_case_locked(&send_data);
2767
2768         /*
2769          * Update the local cpt since send_data.sd_cpt might've been
2770          * updated as a result of calling lnet_handle_send_case_locked().
2771          */
2772         cpt = send_data.sd_cpt;
2773
2774         if (rc == REPEAT_SEND)
2775                 goto again;
2776
2777         lnet_net_unlock(cpt);
2778
2779         return rc;
2780 }
2781
2782 int
2783 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
2784 {
2785         lnet_nid_t              dst_nid = msg->msg_target.nid;
2786         int                     rc;
2787
2788         /*
2789          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
2790          * but we might want to use pre-determined router for ACK/REPLY
2791          * in the future
2792          */
2793         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
2794         LASSERT(msg->msg_txpeer == NULL);
2795         LASSERT(msg->msg_txni == NULL);
2796         LASSERT(!msg->msg_sending);
2797         LASSERT(!msg->msg_target_is_router);
2798         LASSERT(!msg->msg_receiving);
2799
2800         msg->msg_sending = 1;
2801
2802         LASSERT(!msg->msg_tx_committed);
2803
2804         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
2805         if (rc < 0) {
2806                 if (rc == -EHOSTUNREACH)
2807                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
2808                 else
2809                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
2810                 return rc;
2811         }
2812
2813         if (rc == LNET_CREDIT_OK)
2814                 lnet_ni_send(msg->msg_txni, msg);
2815
2816         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
2817         return 0;
2818 }
2819
2820 enum lnet_mt_event_type {
2821         MT_TYPE_LOCAL_NI = 0,
2822         MT_TYPE_PEER_NI
2823 };
2824
2825 struct lnet_mt_event_info {
2826         enum lnet_mt_event_type mt_type;
2827         lnet_nid_t mt_nid;
2828 };
2829
2830 /* called with res_lock held */
2831 void
2832 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
2833 {
2834         struct lnet_rsp_tracker *rspt;
2835
2836         /*
2837          * msg has a refcount on the MD so the MD is not going away.
2838          * The rspt queue for the cpt is protected by
2839          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
2840          */
2841         if (!md->md_rspt_ptr)
2842                 return;
2843
2844         rspt = md->md_rspt_ptr;
2845
2846         /* debug code */
2847         LASSERT(rspt->rspt_cpt == cpt);
2848
2849         md->md_rspt_ptr = NULL;
2850
2851         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2852                 /*
2853                  * The monitor thread has invalidated this handle because the
2854                  * response timed out, but it failed to lookup the MD. That
2855                  * means this response tracker is on the zombie list. We can
2856                  * safely remove it under the resource lock (held by caller) and
2857                  * free the response tracker block.
2858                  */
2859                 list_del(&rspt->rspt_on_list);
2860                 lnet_rspt_free(rspt, cpt);
2861         } else {
2862                 /*
2863                  * invalidate the handle to indicate that a response has been
2864                  * received, which will then lead the monitor thread to clean up
2865                  * the rspt block.
2866                  */
2867                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
2868         }
2869 }
2870
2871 void
2872 lnet_clean_zombie_rstqs(void)
2873 {
2874         struct lnet_rsp_tracker *rspt, *tmp;
2875         int i;
2876
2877         cfs_cpt_for_each(i, lnet_cpt_table()) {
2878                 list_for_each_entry_safe(rspt, tmp,
2879                                          the_lnet.ln_mt_zombie_rstqs[i],
2880                                          rspt_on_list) {
2881                         list_del(&rspt->rspt_on_list);
2882                         lnet_rspt_free(rspt, i);
2883                 }
2884         }
2885
2886         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
2887 }
2888
2889 static void
2890 lnet_finalize_expired_responses(void)
2891 {
2892         struct lnet_libmd *md;
2893         struct lnet_rsp_tracker *rspt, *tmp;
2894         ktime_t now;
2895         int i;
2896
2897         if (the_lnet.ln_mt_rstq == NULL)
2898                 return;
2899
2900         cfs_cpt_for_each(i, lnet_cpt_table()) {
2901                 LIST_HEAD(local_queue);
2902
2903                 lnet_net_lock(i);
2904                 if (!the_lnet.ln_mt_rstq[i]) {
2905                         lnet_net_unlock(i);
2906                         continue;
2907                 }
2908                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
2909                 lnet_net_unlock(i);
2910
2911                 now = ktime_get();
2912
2913                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
2914                         /*
2915                          * The rspt mdh will be invalidated when a response
2916                          * is received or whenever we want to discard the
2917                          * block the monitor thread will walk the queue
2918                          * and clean up any rsts with an invalid mdh.
2919                          * The monitor thread will walk the queue until
2920                          * the first unexpired rspt block. This means that
2921                          * some rspt blocks which received their
2922                          * corresponding responses will linger in the
2923                          * queue until they are cleaned up eventually.
2924                          */
2925                         lnet_res_lock(i);
2926                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
2927                                 lnet_res_unlock(i);
2928                                 list_del(&rspt->rspt_on_list);
2929                                 lnet_rspt_free(rspt, i);
2930                                 continue;
2931                         }
2932
2933                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
2934                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
2935                                 struct lnet_peer_ni *lpni;
2936                                 lnet_nid_t nid;
2937
2938                                 md = lnet_handle2md(&rspt->rspt_mdh);
2939                                 if (!md) {
2940                                         /* MD has been queued for unlink, but
2941                                          * rspt hasn't been detached (Note we've
2942                                          * checked above that the rspt_mdh is
2943                                          * valid). Since we cannot lookup the MD
2944                                          * we're unable to detach the rspt
2945                                          * ourselves. Thus, move the rspt to the
2946                                          * zombie list where we'll wait for
2947                                          * either:
2948                                          *   1. The remaining operations on the
2949                                          *   MD to complete. In this case the
2950                                          *   final operation will result in
2951                                          *   lnet_msg_detach_md()->
2952                                          *   lnet_detach_rsp_tracker() where
2953                                          *   we will clean up this response
2954                                          *   tracker.
2955                                          *   2. LNet to shutdown. In this case
2956                                          *   we'll wait until after all LND Nets
2957                                          *   have shutdown and then we can
2958                                          *   safely free any remaining response
2959                                          *   tracker blocks on the zombie list.
2960                                          * Note: We need to hold the resource
2961                                          * lock when adding to the zombie list
2962                                          * because we may have concurrent access
2963                                          * with lnet_detach_rsp_tracker().
2964                                          */
2965                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
2966                                         list_move(&rspt->rspt_on_list,
2967                                                   the_lnet.ln_mt_zombie_rstqs[i]);
2968                                         lnet_res_unlock(i);
2969                                         continue;
2970                                 }
2971                                 LASSERT(md->md_rspt_ptr == rspt);
2972                                 md->md_rspt_ptr = NULL;
2973                                 lnet_res_unlock(i);
2974
2975                                 LNetMDUnlink(rspt->rspt_mdh);
2976
2977                                 nid = rspt->rspt_next_hop_nid;
2978
2979                                 list_del(&rspt->rspt_on_list);
2980                                 lnet_rspt_free(rspt, i);
2981
2982                                 /* If we're shutting down we just want to clean
2983                                  * up the rspt blocks
2984                                  */
2985                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
2986                                         continue;
2987
2988                                 lnet_net_lock(i);
2989                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
2990                                 lnet_net_unlock(i);
2991
2992                                 CDEBUG(D_NET,
2993                                        "Response timeout: md = %p: nid = %s\n",
2994                                        md, libcfs_nid2str(nid));
2995
2996                                 /*
2997                                  * If there is a timeout on the response
2998                                  * from the next hop decrement its health
2999                                  * value so that we don't use it
3000                                  */
3001                                 lnet_net_lock(0);
3002                                 lpni = lnet_find_peer_ni_locked(nid);
3003                                 if (lpni) {
3004                                         lnet_handle_remote_failure_locked(lpni);
3005                                         lnet_peer_ni_decref_locked(lpni);
3006                                 }
3007                                 lnet_net_unlock(0);
3008                         } else {
3009                                 lnet_res_unlock(i);
3010                                 break;
3011                         }
3012                 }
3013
3014                 if (!list_empty(&local_queue)) {
3015                         lnet_net_lock(i);
3016                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3017                         lnet_net_unlock(i);
3018                 }
3019         }
3020 }
3021
3022 static void
3023 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3024 {
3025         struct lnet_msg *msg;
3026
3027         while (!list_empty(resendq)) {
3028                 struct lnet_peer_ni *lpni;
3029
3030                 msg = list_entry(resendq->next, struct lnet_msg,
3031                                  msg_list);
3032
3033                 list_del_init(&msg->msg_list);
3034
3035                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3036                 if (!lpni) {
3037                         lnet_net_unlock(cpt);
3038                         CERROR("Expected that a peer is already created for %s\n",
3039                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3040                         msg->msg_no_resend = true;
3041                         lnet_finalize(msg, -EFAULT);
3042                         lnet_net_lock(cpt);
3043                 } else {
3044                         int rc;
3045
3046                         lnet_peer_ni_decref_locked(lpni);
3047
3048                         lnet_net_unlock(cpt);
3049                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3050                                libcfs_nid2str(msg->msg_src_nid_param),
3051                                libcfs_id2str(msg->msg_target),
3052                                lnet_msgtyp2str(msg->msg_type),
3053                                msg->msg_recovery,
3054                                msg->msg_retry_count);
3055                         rc = lnet_send(msg->msg_src_nid_param, msg,
3056                                        msg->msg_rtr_nid_param);
3057                         if (rc) {
3058                                 CERROR("Error sending %s to %s: %d\n",
3059                                        lnet_msgtyp2str(msg->msg_type),
3060                                        libcfs_id2str(msg->msg_target), rc);
3061                                 msg->msg_no_resend = true;
3062                                 lnet_finalize(msg, rc);
3063                         }
3064                         lnet_net_lock(cpt);
3065                         if (!rc)
3066                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3067                 }
3068         }
3069 }
3070
3071 static void
3072 lnet_resend_pending_msgs(void)
3073 {
3074         int i;
3075
3076         cfs_cpt_for_each(i, lnet_cpt_table()) {
3077                 lnet_net_lock(i);
3078                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3079                 lnet_net_unlock(i);
3080         }
3081 }
3082
3083 /* called with cpt and ni_lock held */
3084 static void
3085 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3086 {
3087         struct lnet_handle_md recovery_mdh;
3088
3089         LNetInvalidateMDHandle(&recovery_mdh);
3090
3091         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3092             force) {
3093                 recovery_mdh = ni->ni_ping_mdh;
3094                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3095         }
3096         lnet_ni_unlock(ni);
3097         lnet_net_unlock(cpt);
3098         if (!LNetMDHandleIsInvalid(recovery_mdh))
3099                 LNetMDUnlink(recovery_mdh);
3100         lnet_net_lock(cpt);
3101         lnet_ni_lock(ni);
3102 }
3103
3104 static void
3105 lnet_recover_local_nis(void)
3106 {
3107         struct lnet_mt_event_info *ev_info;
3108         LIST_HEAD(processed_list);
3109         LIST_HEAD(local_queue);
3110         struct lnet_handle_md mdh;
3111         struct lnet_ni *tmp;
3112         struct lnet_ni *ni;
3113         lnet_nid_t nid;
3114         int healthv;
3115         int rc;
3116
3117         /*
3118          * splice the recovery queue on a local queue. We will iterate
3119          * through the local queue and update it as needed. Once we're
3120          * done with the traversal, we'll splice the local queue back on
3121          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3122          * will be traversed in the next iteration.
3123          */
3124         lnet_net_lock(0);
3125         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3126                          &local_queue);
3127         lnet_net_unlock(0);
3128
3129         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3130                 /*
3131                  * if an NI is being deleted or it is now healthy, there
3132                  * is no need to keep it around in the recovery queue.
3133                  * The monitor thread is the only thread responsible for
3134                  * removing the NI from the recovery queue.
3135                  * Multiple threads can be adding NIs to the recovery
3136                  * queue.
3137                  */
3138                 healthv = atomic_read(&ni->ni_healthv);
3139
3140                 lnet_net_lock(0);
3141                 lnet_ni_lock(ni);
3142                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3143                     healthv == LNET_MAX_HEALTH_VALUE) {
3144                         list_del_init(&ni->ni_recovery);
3145                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3146                         lnet_ni_unlock(ni);
3147                         lnet_ni_decref_locked(ni, 0);
3148                         lnet_net_unlock(0);
3149                         continue;
3150                 }
3151
3152                 /*
3153                  * if the local NI failed recovery we must unlink the md.
3154                  * But we want to keep the local_ni on the recovery queue
3155                  * so we can continue the attempts to recover it.
3156                  */
3157                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3158                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3159                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3160                 }
3161
3162                 lnet_ni_unlock(ni);
3163                 lnet_net_unlock(0);
3164
3165
3166                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3167                        libcfs_nid2str(ni->ni_nid));
3168
3169                 lnet_ni_lock(ni);
3170                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3171                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3172                         lnet_ni_unlock(ni);
3173
3174                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3175                         if (!ev_info) {
3176                                 CERROR("out of memory. Can't recover %s\n",
3177                                        libcfs_nid2str(ni->ni_nid));
3178                                 lnet_ni_lock(ni);
3179                                 ni->ni_recovery_state &=
3180                                   ~LNET_NI_RECOVERY_PENDING;
3181                                 lnet_ni_unlock(ni);
3182                                 continue;
3183                         }
3184
3185                         mdh = ni->ni_ping_mdh;
3186                         /*
3187                          * Invalidate the ni mdh in case it's deleted.
3188                          * We'll unlink the mdh in this case below.
3189                          */
3190                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3191                         nid = ni->ni_nid;
3192
3193                         /*
3194                          * remove the NI from the local queue and drop the
3195                          * reference count to it while we're recovering
3196                          * it. The reason for that, is that the NI could
3197                          * be deleted, and the way the code is structured
3198                          * is if we don't drop the NI, then the deletion
3199                          * code will enter a loop waiting for the
3200                          * reference count to be removed while holding the
3201                          * ln_mutex_lock(). When we look up the peer to
3202                          * send to in lnet_select_pathway() we will try to
3203                          * lock the ln_mutex_lock() as well, leading to
3204                          * a deadlock. By dropping the refcount and
3205                          * removing it from the list, we allow for the NI
3206                          * to be removed, then we use the cached NID to
3207                          * look it up again. If it's gone, then we just
3208                          * continue examining the rest of the queue.
3209                          */
3210                         lnet_net_lock(0);
3211                         list_del_init(&ni->ni_recovery);
3212                         lnet_ni_decref_locked(ni, 0);
3213                         lnet_net_unlock(0);
3214
3215                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3216                         ev_info->mt_nid = nid;
3217                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3218                                             ev_info, the_lnet.ln_mt_handler,
3219                                             true);
3220                         /* lookup the nid again */
3221                         lnet_net_lock(0);
3222                         ni = lnet_nid2ni_locked(nid, 0);
3223                         if (!ni) {
3224                                 /*
3225                                  * the NI has been deleted when we dropped
3226                                  * the ref count
3227                                  */
3228                                 lnet_net_unlock(0);
3229                                 LNetMDUnlink(mdh);
3230                                 continue;
3231                         }
3232                         /*
3233                          * Same note as in lnet_recover_peer_nis(). When
3234                          * we're sending the ping, the NI is free to be
3235                          * deleted or manipulated. By this point it
3236                          * could've been added back on the recovery queue,
3237                          * and a refcount taken on it.
3238                          * So we can't just add it blindly again or we'll
3239                          * corrupt the queue. We must check under lock if
3240                          * it's not on any list and if not then add it
3241                          * to the processed list, which will eventually be
3242                          * spliced back on to the recovery queue.
3243                          */
3244                         ni->ni_ping_mdh = mdh;
3245                         if (list_empty(&ni->ni_recovery)) {
3246                                 list_add_tail(&ni->ni_recovery, &processed_list);
3247                                 lnet_ni_addref_locked(ni, 0);
3248                         }
3249                         lnet_net_unlock(0);
3250
3251                         lnet_ni_lock(ni);
3252                         if (rc)
3253                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3254                 }
3255                 lnet_ni_unlock(ni);
3256         }
3257
3258         /*
3259          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3260          * reexamined in the next iteration.
3261          */
3262         list_splice_init(&processed_list, &local_queue);
3263         lnet_net_lock(0);
3264         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3265         lnet_net_unlock(0);
3266 }
3267
3268 static int
3269 lnet_resendqs_create(void)
3270 {
3271         struct list_head **resendqs;
3272         resendqs = lnet_create_array_of_queues();
3273
3274         if (!resendqs)
3275                 return -ENOMEM;
3276
3277         lnet_net_lock(LNET_LOCK_EX);
3278         the_lnet.ln_mt_resendqs = resendqs;
3279         lnet_net_unlock(LNET_LOCK_EX);
3280
3281         return 0;
3282 }
3283
3284 static void
3285 lnet_clean_local_ni_recoveryq(void)
3286 {
3287         struct lnet_ni *ni;
3288
3289         /* This is only called when the monitor thread has stopped */
3290         lnet_net_lock(0);
3291
3292         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3293                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3294                                 struct lnet_ni, ni_recovery);
3295                 list_del_init(&ni->ni_recovery);
3296                 lnet_ni_lock(ni);
3297                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3298                 lnet_ni_unlock(ni);
3299                 lnet_ni_decref_locked(ni, 0);
3300         }
3301
3302         lnet_net_unlock(0);
3303 }
3304
3305 static void
3306 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3307                                      bool force)
3308 {
3309         struct lnet_handle_md recovery_mdh;
3310
3311         LNetInvalidateMDHandle(&recovery_mdh);
3312
3313         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3314                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3315                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3316         }
3317         spin_unlock(&lpni->lpni_lock);
3318         lnet_net_unlock(cpt);
3319         if (!LNetMDHandleIsInvalid(recovery_mdh))
3320                 LNetMDUnlink(recovery_mdh);
3321         lnet_net_lock(cpt);
3322         spin_lock(&lpni->lpni_lock);
3323 }
3324
3325 static void
3326 lnet_clean_peer_ni_recoveryq(void)
3327 {
3328         struct lnet_peer_ni *lpni, *tmp;
3329
3330         lnet_net_lock(LNET_LOCK_EX);
3331
3332         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3333                                  lpni_recovery) {
3334                 list_del_init(&lpni->lpni_recovery);
3335                 spin_lock(&lpni->lpni_lock);
3336                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3337                 spin_unlock(&lpni->lpni_lock);
3338                 lnet_peer_ni_decref_locked(lpni);
3339         }
3340
3341         lnet_net_unlock(LNET_LOCK_EX);
3342 }
3343
3344 static void
3345 lnet_clean_resendqs(void)
3346 {
3347         struct lnet_msg *msg, *tmp;
3348         LIST_HEAD(msgs);
3349         int i;
3350
3351         cfs_cpt_for_each(i, lnet_cpt_table()) {
3352                 lnet_net_lock(i);
3353                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3354                 lnet_net_unlock(i);
3355                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3356                         list_del_init(&msg->msg_list);
3357                         msg->msg_no_resend = true;
3358                         lnet_finalize(msg, -ESHUTDOWN);
3359                 }
3360         }
3361
3362         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3363 }
3364
3365 static void
3366 lnet_recover_peer_nis(void)
3367 {
3368         struct lnet_mt_event_info *ev_info;
3369         LIST_HEAD(processed_list);
3370         LIST_HEAD(local_queue);
3371         struct lnet_handle_md mdh;
3372         struct lnet_peer_ni *lpni;
3373         struct lnet_peer_ni *tmp;
3374         lnet_nid_t nid;
3375         int healthv;
3376         int rc;
3377
3378         /*
3379          * Always use cpt 0 for locking across all interactions with
3380          * ln_mt_peerNIRecovq
3381          */
3382         lnet_net_lock(0);
3383         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3384                          &local_queue);
3385         lnet_net_unlock(0);
3386
3387         list_for_each_entry_safe(lpni, tmp, &local_queue,
3388                                  lpni_recovery) {
3389                 /*
3390                  * The same protection strategy is used here as is in the
3391                  * local recovery case.
3392                  */
3393                 lnet_net_lock(0);
3394                 healthv = atomic_read(&lpni->lpni_healthv);
3395                 spin_lock(&lpni->lpni_lock);
3396                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3397                     healthv == LNET_MAX_HEALTH_VALUE) {
3398                         list_del_init(&lpni->lpni_recovery);
3399                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3400                         spin_unlock(&lpni->lpni_lock);
3401                         lnet_peer_ni_decref_locked(lpni);
3402                         lnet_net_unlock(0);
3403                         continue;
3404                 }
3405
3406                 /*
3407                  * If the peer NI has failed recovery we must unlink the
3408                  * md. But we want to keep the peer ni on the recovery
3409                  * queue so we can try to continue recovering it
3410                  */
3411                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3412                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3413                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3414                 }
3415
3416                 spin_unlock(&lpni->lpni_lock);
3417                 lnet_net_unlock(0);
3418
3419                 /*
3420                  * NOTE: we're racing with peer deletion from user space.
3421                  * It's possible that a peer is deleted after we check its
3422                  * state. In this case the recovery can create a new peer
3423                  */
3424                 spin_lock(&lpni->lpni_lock);
3425                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3426                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3427                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3428                         spin_unlock(&lpni->lpni_lock);
3429
3430                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3431                         if (!ev_info) {
3432                                 CERROR("out of memory. Can't recover %s\n",
3433                                        libcfs_nid2str(lpni->lpni_nid));
3434                                 spin_lock(&lpni->lpni_lock);
3435                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3436                                 spin_unlock(&lpni->lpni_lock);
3437                                 continue;
3438                         }
3439
3440                         /* look at the comments in lnet_recover_local_nis() */
3441                         mdh = lpni->lpni_recovery_ping_mdh;
3442                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3443                         nid = lpni->lpni_nid;
3444                         lnet_net_lock(0);
3445                         list_del_init(&lpni->lpni_recovery);
3446                         lnet_peer_ni_decref_locked(lpni);
3447                         lnet_net_unlock(0);
3448
3449                         ev_info->mt_type = MT_TYPE_PEER_NI;
3450                         ev_info->mt_nid = nid;
3451                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3452                                             ev_info, the_lnet.ln_mt_handler,
3453                                             true);
3454                         lnet_net_lock(0);
3455                         /*
3456                          * lnet_find_peer_ni_locked() grabs a refcount for
3457                          * us. No need to take it explicitly.
3458                          */
3459                         lpni = lnet_find_peer_ni_locked(nid);
3460                         if (!lpni) {
3461                                 lnet_net_unlock(0);
3462                                 LNetMDUnlink(mdh);
3463                                 continue;
3464                         }
3465
3466                         lpni->lpni_recovery_ping_mdh = mdh;
3467                         /*
3468                          * While we're unlocked the lpni could've been
3469                          * readded on the recovery queue. In this case we
3470                          * don't need to add it to the local queue, since
3471                          * it's already on there and the thread that added
3472                          * it would've incremented the refcount on the
3473                          * peer, which means we need to decref the refcount
3474                          * that was implicitly grabbed by find_peer_ni_locked.
3475                          * Otherwise, if the lpni is still not on
3476                          * the recovery queue, then we'll add it to the
3477                          * processed list.
3478                          */
3479                         if (list_empty(&lpni->lpni_recovery))
3480                                 list_add_tail(&lpni->lpni_recovery, &processed_list);
3481                         else
3482                                 lnet_peer_ni_decref_locked(lpni);
3483                         lnet_net_unlock(0);
3484
3485                         spin_lock(&lpni->lpni_lock);
3486                         if (rc)
3487                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3488                 }
3489                 spin_unlock(&lpni->lpni_lock);
3490         }
3491
3492         list_splice_init(&processed_list, &local_queue);
3493         lnet_net_lock(0);
3494         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3495         lnet_net_unlock(0);
3496 }
3497
3498 static int
3499 lnet_monitor_thread(void *arg)
3500 {
3501         time64_t recovery_timeout = 0;
3502         time64_t rsp_timeout = 0;
3503         int interval;
3504         time64_t now;
3505
3506         wait_for_completion(&the_lnet.ln_started);
3507         /*
3508          * The monitor thread takes care of the following:
3509          *  1. Checks the aliveness of routers
3510          *  2. Checks if there are messages on the resend queue to resend
3511          *     them.
3512          *  3. Check if there are any NIs on the local recovery queue and
3513          *     pings them
3514          *  4. Checks if there are any NIs on the remote recovery queue
3515          *     and pings them.
3516          */
3517         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3518                 now = ktime_get_real_seconds();
3519
3520                 if (lnet_router_checker_active())
3521                         lnet_check_routers();
3522
3523                 lnet_resend_pending_msgs();
3524
3525                 if (now >= rsp_timeout) {
3526                         lnet_finalize_expired_responses();
3527                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3528                 }
3529
3530                 if (now >= recovery_timeout) {
3531                         lnet_recover_local_nis();
3532                         lnet_recover_peer_nis();
3533                         recovery_timeout = now + lnet_recovery_interval;
3534                 }
3535
3536                 /*
3537                  * TODO do we need to check if we should sleep without
3538                  * timeout?  Technically, an active system will always
3539                  * have messages in flight so this check will always
3540                  * evaluate to false. And on an idle system do we care
3541                  * if we wake up every 1 second? Although, we've seen
3542                  * cases where we get a complaint that an idle thread
3543                  * is waking up unnecessarily.
3544                  *
3545                  * Take into account the current net_count when you wake
3546                  * up for alive router checking, since we need to check
3547                  * possibly as many networks as we have configured.
3548                  */
3549                 interval = min(lnet_recovery_interval,
3550                                min((unsigned int) alive_router_check_interval /
3551                                         lnet_current_net_count,
3552                                    lnet_transaction_timeout / 2));
3553                 wait_for_completion_interruptible_timeout(
3554                         &the_lnet.ln_mt_wait_complete,
3555                         cfs_time_seconds(interval));
3556                 /* Must re-init the completion before testing anything,
3557                  * including ln_mt_state.
3558                  */
3559                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3560         }
3561
3562         /* Shutting down */
3563         lnet_net_lock(LNET_LOCK_EX);
3564         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3565         lnet_net_unlock(LNET_LOCK_EX);
3566
3567         /* signal that the monitor thread is exiting */
3568         up(&the_lnet.ln_mt_signal);
3569
3570         return 0;
3571 }
3572
3573 /*
3574  * lnet_send_ping
3575  * Sends a ping.
3576  * Returns == 0 if success
3577  * Returns > 0 if LNetMDBind or prior fails
3578  * Returns < 0 if LNetGet fails
3579  */
3580 int
3581 lnet_send_ping(lnet_nid_t dest_nid,
3582                struct lnet_handle_md *mdh, int nnis,
3583                void *user_data, lnet_handler_t handler, bool recovery)
3584 {
3585         struct lnet_md md = { NULL };
3586         struct lnet_process_id id;
3587         struct lnet_ping_buffer *pbuf;
3588         int rc;
3589
3590         if (dest_nid == LNET_NID_ANY) {
3591                 rc = -EHOSTUNREACH;
3592                 goto fail_error;
3593         }
3594
3595         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3596         if (!pbuf) {
3597                 rc = ENOMEM;
3598                 goto fail_error;
3599         }
3600
3601         /* initialize md content */
3602         md.start     = &pbuf->pb_info;
3603         md.length    = LNET_PING_INFO_SIZE(nnis);
3604         md.threshold = 2; /* GET/REPLY */
3605         md.max_size  = 0;
3606         md.options   = LNET_MD_TRUNCATE;
3607         md.user_ptr  = user_data;
3608         md.handler   = handler;
3609
3610         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3611         if (rc) {
3612                 lnet_ping_buffer_decref(pbuf);
3613                 CERROR("Can't bind MD: %d\n", rc);
3614                 rc = -rc; /* change the rc to positive */
3615                 goto fail_error;
3616         }
3617         id.pid = LNET_PID_LUSTRE;
3618         id.nid = dest_nid;
3619
3620         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3621                      LNET_RESERVED_PORTAL,
3622                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3623
3624         if (rc)
3625                 goto fail_unlink_md;
3626
3627         return 0;
3628
3629 fail_unlink_md:
3630         LNetMDUnlink(*mdh);
3631         LNetInvalidateMDHandle(mdh);
3632 fail_error:
3633         return rc;
3634 }
3635
3636 static void
3637 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3638                            int status, bool send, bool unlink_event)
3639 {
3640         lnet_nid_t nid = ev_info->mt_nid;
3641
3642         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3643                 struct lnet_ni *ni;
3644
3645                 lnet_net_lock(0);
3646                 ni = lnet_nid2ni_locked(nid, 0);
3647                 if (!ni) {
3648                         lnet_net_unlock(0);
3649                         return;
3650                 }
3651                 lnet_ni_lock(ni);
3652                 if (!send || (send && status != 0))
3653                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3654                 if (status)
3655                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3656                 lnet_ni_unlock(ni);
3657                 lnet_net_unlock(0);
3658
3659                 if (status != 0) {
3660                         CERROR("local NI (%s) recovery failed with %d\n",
3661                                libcfs_nid2str(nid), status);
3662                         return;
3663                 }
3664                 /*
3665                  * need to increment healthv for the ni here, because in
3666                  * the lnet_finalize() path we don't have access to this
3667                  * NI. And in order to get access to it, we'll need to
3668                  * carry forward too much information.
3669                  * In the peer case, it'll naturally be incremented
3670                  */
3671                 if (!unlink_event)
3672                         lnet_inc_healthv(&ni->ni_healthv,
3673                                          lnet_health_sensitivity);
3674         } else {
3675                 struct lnet_peer_ni *lpni;
3676                 int cpt;
3677
3678                 cpt = lnet_net_lock_current();
3679                 lpni = lnet_find_peer_ni_locked(nid);
3680                 if (!lpni) {
3681                         lnet_net_unlock(cpt);
3682                         return;
3683                 }
3684                 spin_lock(&lpni->lpni_lock);
3685                 if (!send || (send && status != 0))
3686                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3687                 if (status)
3688                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3689                 spin_unlock(&lpni->lpni_lock);
3690                 lnet_peer_ni_decref_locked(lpni);
3691                 lnet_net_unlock(cpt);
3692
3693                 if (status != 0)
3694                         CERROR("peer NI (%s) recovery failed with %d\n",
3695                                libcfs_nid2str(nid), status);
3696         }
3697 }
3698
3699 void
3700 lnet_mt_event_handler(struct lnet_event *event)
3701 {
3702         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3703         struct lnet_ping_buffer *pbuf;
3704
3705         /* TODO: remove assert */
3706         LASSERT(event->type == LNET_EVENT_REPLY ||
3707                 event->type == LNET_EVENT_SEND ||
3708                 event->type == LNET_EVENT_UNLINK);
3709
3710         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
3711                event->status);
3712
3713         switch (event->type) {
3714         case LNET_EVENT_UNLINK:
3715                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
3716                        libcfs_nid2str(ev_info->mt_nid));
3717                 /* fallthrough */
3718         case LNET_EVENT_REPLY:
3719                 lnet_handle_recovery_reply(ev_info, event->status, false,
3720                                            event->type == LNET_EVENT_UNLINK);
3721                 break;
3722         case LNET_EVENT_SEND:
3723                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
3724                                libcfs_nid2str(ev_info->mt_nid),
3725                                (event->status) ? "unsuccessfully" :
3726                                "successfully", event->status);
3727                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
3728                 break;
3729         default:
3730                 CERROR("Unexpected event: %d\n", event->type);
3731                 break;
3732         }
3733         if (event->unlinked) {
3734                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
3735                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
3736                 lnet_ping_buffer_decref(pbuf);
3737         }
3738 }
3739
3740 static int
3741 lnet_rsp_tracker_create(void)
3742 {
3743         struct list_head **rstqs;
3744         rstqs = lnet_create_array_of_queues();
3745
3746         if (!rstqs)
3747                 return -ENOMEM;
3748
3749         the_lnet.ln_mt_rstq = rstqs;
3750
3751         return 0;
3752 }
3753
3754 static void
3755 lnet_rsp_tracker_clean(void)
3756 {
3757         lnet_finalize_expired_responses();
3758
3759         cfs_percpt_free(the_lnet.ln_mt_rstq);
3760         the_lnet.ln_mt_rstq = NULL;
3761 }
3762
3763 int lnet_monitor_thr_start(void)
3764 {
3765         int rc = 0;
3766         struct task_struct *task;
3767
3768         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
3769                 return -EALREADY;
3770
3771         rc = lnet_resendqs_create();
3772         if (rc)
3773                 return rc;
3774
3775         rc = lnet_rsp_tracker_create();
3776         if (rc)
3777                 goto clean_queues;
3778
3779         sema_init(&the_lnet.ln_mt_signal, 0);
3780
3781         lnet_net_lock(LNET_LOCK_EX);
3782         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
3783         lnet_net_unlock(LNET_LOCK_EX);
3784         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
3785         if (IS_ERR(task)) {
3786                 rc = PTR_ERR(task);
3787                 CERROR("Can't start monitor thread: %d\n", rc);
3788                 goto clean_thread;
3789         }
3790
3791         return 0;
3792
3793 clean_thread:
3794         lnet_net_lock(LNET_LOCK_EX);
3795         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3796         lnet_net_unlock(LNET_LOCK_EX);
3797         /* block until event callback signals exit */
3798         down(&the_lnet.ln_mt_signal);
3799         /* clean up */
3800         lnet_net_lock(LNET_LOCK_EX);
3801         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3802         lnet_net_unlock(LNET_LOCK_EX);
3803         lnet_rsp_tracker_clean();
3804         lnet_clean_local_ni_recoveryq();
3805         lnet_clean_peer_ni_recoveryq();
3806         lnet_clean_resendqs();
3807         the_lnet.ln_mt_handler = NULL;
3808         return rc;
3809 clean_queues:
3810         lnet_rsp_tracker_clean();
3811         lnet_clean_local_ni_recoveryq();
3812         lnet_clean_peer_ni_recoveryq();
3813         lnet_clean_resendqs();
3814         return rc;
3815 }
3816
3817 void lnet_monitor_thr_stop(void)
3818 {
3819         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3820                 return;
3821
3822         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
3823         lnet_net_lock(LNET_LOCK_EX);
3824         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
3825         lnet_net_unlock(LNET_LOCK_EX);
3826
3827         /* tell the monitor thread that we're shutting down */
3828         complete(&the_lnet.ln_mt_wait_complete);
3829
3830         /* block until monitor thread signals that it's done */
3831         down(&the_lnet.ln_mt_signal);
3832         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
3833
3834         /* perform cleanup tasks */
3835         lnet_rsp_tracker_clean();
3836         lnet_clean_local_ni_recoveryq();
3837         lnet_clean_peer_ni_recoveryq();
3838         lnet_clean_resendqs();
3839 }
3840
3841 void
3842 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
3843                   __u32 msg_type)
3844 {
3845         lnet_net_lock(cpt);
3846         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
3847         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
3848         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
3849         lnet_net_unlock(cpt);
3850
3851         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
3852 }
3853
3854 static void
3855 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
3856 {
3857         struct lnet_hdr *hdr = &msg->msg_hdr;
3858
3859         if (msg->msg_wanted != 0)
3860                 lnet_setpayloadbuffer(msg);
3861
3862         lnet_build_msg_event(msg, LNET_EVENT_PUT);
3863
3864         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
3865          * it back into the ACK during lnet_finalize() */
3866         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
3867                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
3868
3869         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
3870                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
3871 }
3872
3873 static int
3874 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
3875 {
3876         struct lnet_hdr         *hdr = &msg->msg_hdr;
3877         struct lnet_match_info  info;
3878         int                     rc;
3879         bool                    ready_delay;
3880
3881         /* Convert put fields to host byte order */
3882         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
3883         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
3884         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
3885
3886         /* Primary peer NID. */
3887         info.mi_id.nid  = msg->msg_initiator;
3888         info.mi_id.pid  = hdr->src_pid;
3889         info.mi_opc     = LNET_MD_OP_PUT;
3890         info.mi_portal  = hdr->msg.put.ptl_index;
3891         info.mi_rlength = hdr->payload_length;
3892         info.mi_roffset = hdr->msg.put.offset;
3893         info.mi_mbits   = hdr->msg.put.match_bits;
3894         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3895
3896         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
3897         ready_delay = msg->msg_rx_ready_delay;
3898
3899  again:
3900         rc = lnet_ptl_match_md(&info, msg);
3901         switch (rc) {
3902         default:
3903                 LBUG();
3904
3905         case LNET_MATCHMD_OK:
3906                 lnet_recv_put(ni, msg);
3907                 return 0;
3908
3909         case LNET_MATCHMD_NONE:
3910                 if (ready_delay)
3911                         /* no eager_recv or has already called it, should
3912                          * have been attached on delayed list */
3913                         return 0;
3914
3915                 rc = lnet_ni_eager_recv(ni, msg);
3916                 if (rc == 0) {
3917                         ready_delay = true;
3918                         goto again;
3919                 }
3920                 /* fall through */
3921
3922         case LNET_MATCHMD_DROP:
3923                 CNETERR("Dropping PUT from %s portal %d match %llu"
3924                         " offset %d length %d: %d\n",
3925                         libcfs_id2str(info.mi_id), info.mi_portal,
3926                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
3927
3928                 return -ENOENT; /* -ve: OK but no match */
3929         }
3930 }
3931
3932 static int
3933 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
3934 {
3935         struct lnet_match_info info;
3936         struct lnet_hdr *hdr = &msg->msg_hdr;
3937         struct lnet_process_id source_id;
3938         struct lnet_handle_wire reply_wmd;
3939         int rc;
3940
3941         /* Convert get fields to host byte order */
3942         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
3943         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
3944         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
3945         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
3946
3947         source_id.nid = hdr->src_nid;
3948         source_id.pid = hdr->src_pid;
3949         /* Primary peer NID */
3950         info.mi_id.nid  = msg->msg_initiator;
3951         info.mi_id.pid  = hdr->src_pid;
3952         info.mi_opc     = LNET_MD_OP_GET;
3953         info.mi_portal  = hdr->msg.get.ptl_index;
3954         info.mi_rlength = hdr->msg.get.sink_length;
3955         info.mi_roffset = hdr->msg.get.src_offset;
3956         info.mi_mbits   = hdr->msg.get.match_bits;
3957         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
3958
3959         rc = lnet_ptl_match_md(&info, msg);
3960         if (rc == LNET_MATCHMD_DROP) {
3961                 CNETERR("Dropping GET from %s portal %d match %llu"
3962                         " offset %d length %d\n",
3963                         libcfs_id2str(info.mi_id), info.mi_portal,
3964                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
3965                 return -ENOENT; /* -ve: OK but no match */
3966         }
3967
3968         LASSERT(rc == LNET_MATCHMD_OK);
3969
3970         lnet_build_msg_event(msg, LNET_EVENT_GET);
3971
3972         reply_wmd = hdr->msg.get.return_wmd;
3973
3974         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
3975                        msg->msg_offset, msg->msg_wanted);
3976
3977         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
3978
3979         if (rdma_get) {
3980                 /* The LND completes the REPLY from her recv procedure */
3981                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
3982                              msg->msg_offset, msg->msg_len, msg->msg_len);
3983                 return 0;
3984         }
3985
3986         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
3987         msg->msg_receiving = 0;
3988
3989         rc = lnet_send(ni->ni_nid, msg, msg->msg_from);
3990         if (rc < 0) {
3991                 /* didn't get as far as lnet_ni_send() */
3992                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
3993                        libcfs_nid2str(ni->ni_nid),
3994                        libcfs_id2str(info.mi_id), rc);
3995
3996                 lnet_finalize(msg, rc);
3997         }
3998
3999         return 0;
4000 }
4001
4002 static int
4003 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4004 {
4005         void *private = msg->msg_private;
4006         struct lnet_hdr *hdr = &msg->msg_hdr;
4007         struct lnet_process_id src = {0};
4008         struct lnet_libmd *md;
4009         unsigned int rlength;
4010         unsigned int mlength;
4011         int cpt;
4012
4013         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4014         lnet_res_lock(cpt);
4015
4016         src.nid = hdr->src_nid;
4017         src.pid = hdr->src_pid;
4018
4019         /* NB handles only looked up by creator (no flips) */
4020         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4021         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4022                 CNETERR("%s: Dropping REPLY from %s for %s "
4023                         "MD %#llx.%#llx\n",
4024                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4025                         (md == NULL) ? "invalid" : "inactive",
4026                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4027                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4028                 if (md != NULL && md->md_me != NULL)
4029                         CERROR("REPLY MD also attached to portal %d\n",
4030                                md->md_me->me_portal);
4031
4032                 lnet_res_unlock(cpt);
4033                 return -ENOENT; /* -ve: OK but no match */
4034         }
4035
4036         LASSERT(md->md_offset == 0);
4037
4038         rlength = hdr->payload_length;
4039         mlength = min(rlength, md->md_length);
4040
4041         if (mlength < rlength &&
4042             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4043                 CNETERR("%s: Dropping REPLY from %s length %d "
4044                         "for MD %#llx would overflow (%d)\n",
4045                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4046                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4047                         mlength);
4048                 lnet_res_unlock(cpt);
4049                 return -ENOENT; /* -ve: OK but no match */
4050         }
4051
4052         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4053                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4054                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4055
4056         lnet_msg_attach_md(msg, md, 0, mlength);
4057
4058         if (mlength != 0)
4059                 lnet_setpayloadbuffer(msg);
4060
4061         lnet_res_unlock(cpt);
4062
4063         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4064
4065         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4066         return 0;
4067 }
4068
4069 static int
4070 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4071 {
4072         struct lnet_hdr *hdr = &msg->msg_hdr;
4073         struct lnet_process_id src = {0};
4074         struct lnet_libmd *md;
4075         int cpt;
4076
4077         src.nid = hdr->src_nid;
4078         src.pid = hdr->src_pid;
4079
4080         /* Convert ack fields to host byte order */
4081         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4082         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4083
4084         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4085         lnet_res_lock(cpt);
4086
4087         /* NB handles only looked up by creator (no flips) */
4088         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4089         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4090                 /* Don't moan; this is expected */
4091                 CDEBUG(D_NET,
4092                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4093                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4094                        (md == NULL) ? "invalid" : "inactive",
4095                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4096                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4097                 if (md != NULL && md->md_me != NULL)
4098                         CERROR("Source MD also attached to portal %d\n",
4099                                md->md_me->me_portal);
4100
4101                 lnet_res_unlock(cpt);
4102                 return -ENOENT;                  /* -ve! */
4103         }
4104
4105         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4106                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
4107                hdr->msg.ack.dst_wmd.wh_object_cookie);
4108
4109         lnet_msg_attach_md(msg, md, 0, 0);
4110
4111         lnet_res_unlock(cpt);
4112
4113         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4114
4115         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4116         return 0;
4117 }
4118
4119 /**
4120  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4121  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4122  * \retval -ve                  error code
4123  */
4124 int
4125 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4126 {
4127         int     rc = 0;
4128
4129         if (!the_lnet.ln_routing)
4130                 return -ECANCELED;
4131
4132         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4133             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4134                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4135                         msg->msg_rx_ready_delay = 1;
4136                 } else {
4137                         lnet_net_unlock(msg->msg_rx_cpt);
4138                         rc = lnet_ni_eager_recv(ni, msg);
4139                         lnet_net_lock(msg->msg_rx_cpt);
4140                 }
4141         }
4142
4143         if (rc == 0)
4144                 rc = lnet_post_routed_recv_locked(msg, 0);
4145         return rc;
4146 }
4147
4148 int
4149 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4150 {
4151         int     rc;
4152
4153         switch (msg->msg_type) {
4154         case LNET_MSG_ACK:
4155                 rc = lnet_parse_ack(ni, msg);
4156                 break;
4157         case LNET_MSG_PUT:
4158                 rc = lnet_parse_put(ni, msg);
4159                 break;
4160         case LNET_MSG_GET:
4161                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4162                 break;
4163         case LNET_MSG_REPLY:
4164                 rc = lnet_parse_reply(ni, msg);
4165                 break;
4166         default: /* prevent an unused label if !kernel */
4167                 LASSERT(0);
4168                 return -EPROTO;
4169         }
4170
4171         LASSERT(rc == 0 || rc == -ENOENT);
4172         return rc;
4173 }
4174
4175 char *
4176 lnet_msgtyp2str (int type)
4177 {
4178         switch (type) {
4179         case LNET_MSG_ACK:
4180                 return ("ACK");
4181         case LNET_MSG_PUT:
4182                 return ("PUT");
4183         case LNET_MSG_GET:
4184                 return ("GET");
4185         case LNET_MSG_REPLY:
4186                 return ("REPLY");
4187         case LNET_MSG_HELLO:
4188                 return ("HELLO");
4189         default:
4190                 return ("<UNKNOWN>");
4191         }
4192 }
4193
4194 void
4195 lnet_print_hdr(struct lnet_hdr *hdr)
4196 {
4197         struct lnet_process_id src = {
4198                 .nid = hdr->src_nid,
4199                 .pid = hdr->src_pid,
4200         };
4201         struct lnet_process_id dst = {
4202                 .nid = hdr->dest_nid,
4203                 .pid = hdr->dest_pid,
4204         };
4205         char *type_str = lnet_msgtyp2str(hdr->type);
4206
4207         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
4208         CWARN("    From %s\n", libcfs_id2str(src));
4209         CWARN("    To   %s\n", libcfs_id2str(dst));
4210
4211         switch (hdr->type) {
4212         default:
4213                 break;
4214
4215         case LNET_MSG_PUT:
4216                 CWARN("    Ptl index %d, ack md %#llx.%#llx, "
4217                       "match bits %llu\n",
4218                       hdr->msg.put.ptl_index,
4219                       hdr->msg.put.ack_wmd.wh_interface_cookie,
4220                       hdr->msg.put.ack_wmd.wh_object_cookie,
4221                       hdr->msg.put.match_bits);
4222                 CWARN("    Length %d, offset %d, hdr data %#llx\n",
4223                       hdr->payload_length, hdr->msg.put.offset,
4224                       hdr->msg.put.hdr_data);
4225                 break;
4226
4227         case LNET_MSG_GET:
4228                 CWARN("    Ptl index %d, return md %#llx.%#llx, "
4229                       "match bits %llu\n", hdr->msg.get.ptl_index,
4230                       hdr->msg.get.return_wmd.wh_interface_cookie,
4231                       hdr->msg.get.return_wmd.wh_object_cookie,
4232                       hdr->msg.get.match_bits);
4233                 CWARN("    Length %d, src offset %d\n",
4234                       hdr->msg.get.sink_length,
4235                       hdr->msg.get.src_offset);
4236                 break;
4237
4238         case LNET_MSG_ACK:
4239                 CWARN("    dst md %#llx.%#llx, "
4240                       "manipulated length %d\n",
4241                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
4242                       hdr->msg.ack.dst_wmd.wh_object_cookie,
4243                       hdr->msg.ack.mlength);
4244                 break;
4245
4246         case LNET_MSG_REPLY:
4247                 CWARN("    dst md %#llx.%#llx, "
4248                       "length %d\n",
4249                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
4250                       hdr->msg.reply.dst_wmd.wh_object_cookie,
4251                       hdr->payload_length);
4252         }
4253
4254 }
4255
4256 int
4257 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
4258            void *private, int rdma_req)
4259 {
4260         struct lnet_peer_ni *lpni;
4261         struct lnet_msg *msg;
4262         __u32 payload_length;
4263         lnet_pid_t dest_pid;
4264         lnet_nid_t dest_nid;
4265         lnet_nid_t src_nid;
4266         bool push = false;
4267         int for_me;
4268         __u32 type;
4269         int rc = 0;
4270         int cpt;
4271
4272         LASSERT (!in_interrupt ());
4273
4274         type = le32_to_cpu(hdr->type);
4275         src_nid = le64_to_cpu(hdr->src_nid);
4276         dest_nid = le64_to_cpu(hdr->dest_nid);
4277         dest_pid = le32_to_cpu(hdr->dest_pid);
4278         payload_length = le32_to_cpu(hdr->payload_length);
4279
4280         for_me = (ni->ni_nid == dest_nid);
4281         cpt = lnet_cpt_of_nid(from_nid, ni);
4282
4283         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4284                 libcfs_nid2str(dest_nid),
4285                 libcfs_nid2str(ni->ni_nid),
4286                 libcfs_nid2str(src_nid),
4287                 lnet_msgtyp2str(type),
4288                 (for_me) ? "for me" : "routed");
4289
4290         switch (type) {
4291         case LNET_MSG_ACK:
4292         case LNET_MSG_GET:
4293                 if (payload_length > 0) {
4294                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4295                                libcfs_nid2str(from_nid),
4296                                libcfs_nid2str(src_nid),
4297                                lnet_msgtyp2str(type), payload_length);
4298                         return -EPROTO;
4299                 }
4300                 break;
4301
4302         case LNET_MSG_PUT:
4303         case LNET_MSG_REPLY:
4304                 if (payload_length >
4305                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4306                         CERROR("%s, src %s: bad %s payload %d "
4307                                "(%d max expected)\n",
4308                                libcfs_nid2str(from_nid),
4309                                libcfs_nid2str(src_nid),
4310                                lnet_msgtyp2str(type),
4311                                payload_length,
4312                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4313                         return -EPROTO;
4314                 }
4315                 break;
4316
4317         default:
4318                 CERROR("%s, src %s: Bad message type 0x%x\n",
4319                        libcfs_nid2str(from_nid),
4320                        libcfs_nid2str(src_nid), type);
4321                 return -EPROTO;
4322         }
4323
4324         if (the_lnet.ln_routing &&
4325             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4326                 lnet_ni_lock(ni);
4327                 spin_lock(&ni->ni_net->net_lock);
4328                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4329                 spin_unlock(&ni->ni_net->net_lock);
4330                 if (ni->ni_status != NULL &&
4331                     ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
4332                         ni->ni_status->ns_status = LNET_NI_STATUS_UP;
4333                         push = true;
4334                 }
4335                 lnet_ni_unlock(ni);
4336         }
4337
4338         if (push)
4339                 lnet_push_update_to_peers(1);
4340
4341         /* Regard a bad destination NID as a protocol error.  Senders should
4342          * know what they're doing; if they don't they're misconfigured, buggy
4343          * or malicious so we chop them off at the knees :) */
4344
4345         if (!for_me) {
4346                 if (LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
4347                         /* should have gone direct */
4348                         CERROR("%s, src %s: Bad dest nid %s "
4349                                "(should have been sent direct)\n",
4350                                 libcfs_nid2str(from_nid),
4351                                 libcfs_nid2str(src_nid),
4352                                 libcfs_nid2str(dest_nid));
4353                         return -EPROTO;
4354                 }
4355
4356                 if (lnet_islocalnid(dest_nid)) {
4357                         /* dest is another local NI; sender should have used
4358                          * this node's NID on its own network */
4359                         CERROR("%s, src %s: Bad dest nid %s "
4360                                "(it's my nid but on a different network)\n",
4361                                 libcfs_nid2str(from_nid),
4362                                 libcfs_nid2str(src_nid),
4363                                 libcfs_nid2str(dest_nid));
4364                         return -EPROTO;
4365                 }
4366
4367                 if (rdma_req && type == LNET_MSG_GET) {
4368                         CERROR("%s, src %s: Bad optimized GET for %s "
4369                                "(final destination must be me)\n",
4370                                 libcfs_nid2str(from_nid),
4371                                 libcfs_nid2str(src_nid),
4372                                 libcfs_nid2str(dest_nid));
4373                         return -EPROTO;
4374                 }
4375
4376                 if (!the_lnet.ln_routing) {
4377                         CERROR("%s, src %s: Dropping message for %s "
4378                                "(routing not enabled)\n",
4379                                 libcfs_nid2str(from_nid),
4380                                 libcfs_nid2str(src_nid),
4381                                 libcfs_nid2str(dest_nid));
4382                         goto drop;
4383                 }
4384         }
4385
4386         /* Message looks OK; we're not going to return an error, so we MUST
4387          * call back lnd_recv() come what may... */
4388
4389         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4390             fail_peer(src_nid, 0)) {                    /* shall we now? */
4391                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4392                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4393                        lnet_msgtyp2str(type));
4394                 goto drop;
4395         }
4396
4397         if (!list_empty(&the_lnet.ln_drop_rules) &&
4398             lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
4399                 CDEBUG(D_NET,
4400                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4401                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4402                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4403                 goto drop;
4404         }
4405
4406         if (lnet_drop_asym_route && for_me &&
4407             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
4408                 struct lnet_net *net;
4409                 struct lnet_remotenet *rnet;
4410                 bool found = true;
4411
4412                 /* we are dealing with a routed message,
4413                  * so see if route to reach src_nid goes through from_nid
4414                  */
4415                 lnet_net_lock(cpt);
4416                 net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
4417                 if (!net) {
4418                         lnet_net_unlock(cpt);
4419                         CERROR("net %s not found\n",
4420                                libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
4421                         return -EPROTO;
4422                 }
4423
4424                 rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
4425                 if (rnet) {
4426                         struct lnet_peer *gw = NULL;
4427                         struct lnet_peer_ni *lpni = NULL;
4428                         struct lnet_route *route;
4429
4430                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
4431                                 found = false;
4432                                 gw = route->lr_gateway;
4433                                 if (route->lr_lnet != net->net_id)
4434                                         continue;
4435                                 /*
4436                                  * if the nid is one of the gateway's NIDs
4437                                  * then this is a valid gateway
4438                                  */
4439                                 while ((lpni = lnet_get_next_peer_ni_locked(gw,
4440                                                 NULL, lpni)) != NULL) {
4441                                         if (lpni->lpni_nid == from_nid) {
4442                                                 found = true;
4443                                                 break;
4444                                         }
4445                                 }
4446                         }
4447                 }
4448                 lnet_net_unlock(cpt);
4449                 if (!found) {
4450                         /* we would not use from_nid to route a message to
4451                          * src_nid
4452                          * => asymmetric routing detected but forbidden
4453                          */
4454                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4455                                libcfs_nid2str(from_nid),
4456                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4457                         goto drop;
4458                 }
4459         }
4460
4461         msg = lnet_msg_alloc();
4462         if (msg == NULL) {
4463                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4464                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4465                        lnet_msgtyp2str(type));
4466                 goto drop;
4467         }
4468
4469         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4470          * pointers NULL etc */
4471
4472         msg->msg_type = type;
4473         msg->msg_private = private;
4474         msg->msg_receiving = 1;
4475         msg->msg_rdma_get = rdma_req;
4476         msg->msg_len = msg->msg_wanted = payload_length;
4477         msg->msg_offset = 0;
4478         msg->msg_hdr = *hdr;
4479         /* for building message event */
4480         msg->msg_from = from_nid;
4481         if (!for_me) {
4482                 msg->msg_target.pid     = dest_pid;
4483                 msg->msg_target.nid     = dest_nid;
4484                 msg->msg_routing        = 1;
4485
4486         } else {
4487                 /* convert common msg->hdr fields to host byteorder */
4488                 msg->msg_hdr.type       = type;
4489                 msg->msg_hdr.src_nid    = src_nid;
4490                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4491                 msg->msg_hdr.dest_nid   = dest_nid;
4492                 msg->msg_hdr.dest_pid   = dest_pid;
4493                 msg->msg_hdr.payload_length = payload_length;
4494         }
4495
4496         lnet_net_lock(cpt);
4497         lpni = lnet_nid2peerni_locked(from_nid, ni->ni_nid, cpt);
4498         if (IS_ERR(lpni)) {
4499                 lnet_net_unlock(cpt);
4500                 CERROR("%s, src %s: Dropping %s "
4501                        "(error %ld looking up sender)\n",
4502                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
4503                        lnet_msgtyp2str(type), PTR_ERR(lpni));
4504                 lnet_msg_free(msg);
4505                 if (rc == -ESHUTDOWN)
4506                         /* We are shutting down.  Don't do anything more */
4507                         return 0;
4508                 goto drop;
4509         }
4510
4511         if (the_lnet.ln_routing)
4512                 lpni->lpni_last_alive = ktime_get_seconds();
4513
4514         msg->msg_rxpeer = lpni;
4515         msg->msg_rxni = ni;
4516         lnet_ni_addref_locked(ni, cpt);
4517         /* Multi-Rail: Primary NID of source. */
4518         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4519
4520         /*
4521          * mark the status of this lpni as UP since we received a message
4522          * from it. The ping response reports back the ns_status which is
4523          * marked on the remote as up or down and we cache it here.
4524          */
4525         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4526
4527         lnet_msg_commit(msg, cpt);
4528
4529         /* message delay simulation */
4530         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4531                      lnet_delay_rule_match_locked(hdr, msg))) {
4532                 lnet_net_unlock(cpt);
4533                 return 0;
4534         }
4535
4536         if (!for_me) {
4537                 rc = lnet_parse_forward_locked(ni, msg);
4538                 lnet_net_unlock(cpt);
4539
4540                 if (rc < 0)
4541                         goto free_drop;
4542
4543                 if (rc == LNET_CREDIT_OK) {
4544                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4545                                      0, payload_length, payload_length);
4546                 }
4547                 return 0;
4548         }
4549
4550         lnet_net_unlock(cpt);
4551
4552         rc = lnet_parse_local(ni, msg);
4553         if (rc != 0)
4554                 goto free_drop;
4555         return 0;
4556
4557  free_drop:
4558         LASSERT(msg->msg_md == NULL);
4559         lnet_finalize(msg, rc);
4560
4561  drop:
4562         lnet_drop_message(ni, cpt, private, payload_length, type);
4563         return 0;
4564 }
4565 EXPORT_SYMBOL(lnet_parse);
4566
4567 void
4568 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4569 {
4570         while (!list_empty(head)) {
4571                 struct lnet_process_id id = {0};
4572                 struct lnet_msg *msg;
4573
4574                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4575                 list_del(&msg->msg_list);
4576
4577                 id.nid = msg->msg_hdr.src_nid;
4578                 id.pid = msg->msg_hdr.src_pid;
4579
4580                 LASSERT(msg->msg_md == NULL);
4581                 LASSERT(msg->msg_rx_delayed);
4582                 LASSERT(msg->msg_rxpeer != NULL);
4583                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4584
4585                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4586                       " offset %d length %d: %s\n",
4587                       libcfs_id2str(id),
4588                       msg->msg_hdr.msg.put.ptl_index,
4589                       msg->msg_hdr.msg.put.match_bits,
4590                       msg->msg_hdr.msg.put.offset,
4591                       msg->msg_hdr.payload_length, reason);
4592
4593                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4594                  * called lnet_drop_message(), so I just hang onto msg as well
4595                  * until that's done */
4596
4597                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4598                                   msg->msg_private, msg->msg_len,
4599                                   msg->msg_type);
4600
4601                 msg->msg_no_resend = true;
4602                 /*
4603                  * NB: message will not generate event because w/o attached MD,
4604                  * but we still should give error code so lnet_msg_decommit()
4605                  * can skip counters operations and other checks.
4606                  */
4607                 lnet_finalize(msg, -ENOENT);
4608         }
4609 }
4610
4611 void
4612 lnet_recv_delayed_msg_list(struct list_head *head)
4613 {
4614         while (!list_empty(head)) {
4615                 struct lnet_msg *msg;
4616                 struct lnet_process_id id;
4617
4618                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4619                 list_del(&msg->msg_list);
4620
4621                 /* md won't disappear under me, since each msg
4622                  * holds a ref on it */
4623
4624                 id.nid = msg->msg_hdr.src_nid;
4625                 id.pid = msg->msg_hdr.src_pid;
4626
4627                 LASSERT(msg->msg_rx_delayed);
4628                 LASSERT(msg->msg_md != NULL);
4629                 LASSERT(msg->msg_rxpeer != NULL);
4630                 LASSERT(msg->msg_rxni != NULL);
4631                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4632
4633                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4634                        "match %llu offset %d length %d.\n",
4635                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4636                         msg->msg_hdr.msg.put.match_bits,
4637                         msg->msg_hdr.msg.put.offset,
4638                         msg->msg_hdr.payload_length);
4639
4640                 lnet_recv_put(msg->msg_rxni, msg);
4641         }
4642 }
4643
4644 static void
4645 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4646                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4647 {
4648         s64 timeout_ns;
4649         struct lnet_rsp_tracker *local_rspt;
4650
4651         /*
4652          * MD has a refcount taken by message so it's not going away.
4653          * The MD however can be looked up. We need to secure the access
4654          * to the md_rspt_ptr by taking the res_lock.
4655          * The rspt can be accessed without protection up to when it gets
4656          * added to the list.
4657          */
4658
4659         lnet_res_lock(cpt);
4660         local_rspt = md->md_rspt_ptr;
4661         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4662         if (local_rspt != NULL) {
4663                 /*
4664                  * we already have an rspt attached to the md, so we'll
4665                  * update the deadline on that one.
4666                  */
4667                 lnet_rspt_free(rspt, cpt);
4668         } else {
4669                 /* new md */
4670                 rspt->rspt_mdh = mdh;
4671                 rspt->rspt_cpt = cpt;
4672                 /* store the rspt so we can access it when we get the REPLY */
4673                 md->md_rspt_ptr = rspt;
4674                 local_rspt = rspt;
4675         }
4676         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4677
4678         /*
4679          * add to the list of tracked responses. It's added to tail of the
4680          * list in order to expire all the older entries first.
4681          */
4682         lnet_net_lock(cpt);
4683         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4684         lnet_net_unlock(cpt);
4685         lnet_res_unlock(cpt);
4686 }
4687
4688 /**
4689  * Initiate an asynchronous PUT operation.
4690  *
4691  * There are several events associated with a PUT: completion of the send on
4692  * the initiator node (LNET_EVENT_SEND), and when the send completes
4693  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4694  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4695  * used at the target node to indicate the completion of incoming data
4696  * delivery.
4697  *
4698  * The local events will be logged in the EQ associated with the MD pointed to
4699  * by \a mdh handle. Using a MD without an associated EQ results in these
4700  * events being discarded. In this case, the caller must have another
4701  * mechanism (e.g., a higher level protocol) for determining when it is safe
4702  * to modify the memory region associated with the MD.
4703  *
4704  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4705  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4706  *
4707  * \param self Indicates the NID of a local interface through which to send
4708  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4709  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4710  * must be "free floating" (See LNetMDBind()).
4711  * \param ack Controls whether an acknowledgment is requested.
4712  * Acknowledgments are only sent when they are requested by the initiating
4713  * process and the target MD enables them.
4714  * \param target A process identifier for the target process.
4715  * \param portal The index in the \a target's portal table.
4716  * \param match_bits The match bits to use for MD selection at the target
4717  * process.
4718  * \param offset The offset into the target MD (only used when the target
4719  * MD has the LNET_MD_MANAGE_REMOTE option set).
4720  * \param hdr_data 64 bits of user data that can be included in the message
4721  * header. This data is written to an event queue entry at the target if an
4722  * EQ is present on the matching MD.
4723  *
4724  * \retval  0      Success, and only in this case events will be generated
4725  * and logged to EQ (if it exists).
4726  * \retval -EIO    Simulated failure.
4727  * \retval -ENOMEM Memory allocation failure.
4728  * \retval -ENOENT Invalid MD object.
4729  *
4730  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4731  */
4732 int
4733 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4734         struct lnet_process_id target, unsigned int portal,
4735         __u64 match_bits, unsigned int offset,
4736         __u64 hdr_data)
4737 {
4738         struct lnet_msg *msg;
4739         struct lnet_libmd *md;
4740         int cpt;
4741         int rc;
4742         struct lnet_rsp_tracker *rspt = NULL;
4743
4744         LASSERT(the_lnet.ln_refcount > 0);
4745
4746         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4747             fail_peer(target.nid, 1)) {                 /* shall we now? */
4748                 CERROR("Dropping PUT to %s: simulated failure\n",
4749                        libcfs_id2str(target));
4750                 return -EIO;
4751         }
4752
4753         msg = lnet_msg_alloc();
4754         if (msg == NULL) {
4755                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4756                        libcfs_id2str(target));
4757                 return -ENOMEM;
4758         }
4759         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4760
4761         cpt = lnet_cpt_of_cookie(mdh.cookie);
4762
4763         if (ack == LNET_ACK_REQ) {
4764                 rspt = lnet_rspt_alloc(cpt);
4765                 if (!rspt) {
4766                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
4767                                 libcfs_id2str(target));
4768                         return -ENOMEM;
4769                 }
4770                 INIT_LIST_HEAD(&rspt->rspt_on_list);
4771         }
4772
4773         lnet_res_lock(cpt);
4774
4775         md = lnet_handle2md(&mdh);
4776         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4777                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
4778                        match_bits, portal, libcfs_id2str(target),
4779                        md == NULL ? -1 : md->md_threshold);
4780                 if (md != NULL && md->md_me != NULL)
4781                         CERROR("Source MD also attached to portal %d\n",
4782                                md->md_me->me_portal);
4783                 lnet_res_unlock(cpt);
4784
4785                 lnet_rspt_free(rspt, cpt);
4786                 lnet_msg_free(msg);
4787                 return -ENOENT;
4788         }
4789
4790         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
4791
4792         lnet_msg_attach_md(msg, md, 0, 0);
4793
4794         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
4795
4796         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
4797         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
4798         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
4799         msg->msg_hdr.msg.put.hdr_data = hdr_data;
4800
4801         /* NB handles only looked up by creator (no flips) */
4802         if (ack == LNET_ACK_REQ) {
4803                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4804                         the_lnet.ln_interface_cookie;
4805                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4806                         md->md_lh.lh_cookie;
4807         } else {
4808                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
4809                         LNET_WIRE_HANDLE_COOKIE_NONE;
4810                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
4811                         LNET_WIRE_HANDLE_COOKIE_NONE;
4812         }
4813
4814         lnet_res_unlock(cpt);
4815
4816         lnet_build_msg_event(msg, LNET_EVENT_SEND);
4817
4818         if (ack == LNET_ACK_REQ)
4819                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
4820
4821         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
4822                                  CFS_FAIL_ONCE))
4823                 rc = -EIO;
4824         else
4825                 rc = lnet_send(self, msg, LNET_NID_ANY);
4826
4827         if (rc != 0) {
4828                 CNETERR("Error sending PUT to %s: %d\n",
4829                         libcfs_id2str(target), rc);
4830                 msg->msg_no_resend = true;
4831                 lnet_finalize(msg, rc);
4832         }
4833
4834         /* completion will be signalled by an event */
4835         return 0;
4836 }
4837 EXPORT_SYMBOL(LNetPut);
4838
4839 /*
4840  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
4841  * returns a msg for the LND to pass to lnet_finalize() when the sink
4842  * data has been received.
4843  *
4844  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
4845  * lnet_finalize() is called on it, so the LND must call this first
4846  */
4847 struct lnet_msg *
4848 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
4849 {
4850         struct lnet_msg *msg = lnet_msg_alloc();
4851         struct lnet_libmd *getmd = getmsg->msg_md;
4852         struct lnet_process_id peer_id = getmsg->msg_target;
4853         int cpt;
4854
4855         LASSERT(!getmsg->msg_target_is_router);
4856         LASSERT(!getmsg->msg_routing);
4857
4858         if (msg == NULL) {
4859                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
4860                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
4861                 goto drop;
4862         }
4863
4864         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
4865         lnet_res_lock(cpt);
4866
4867         LASSERT(getmd->md_refcount > 0);
4868
4869         if (getmd->md_threshold == 0) {
4870                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
4871                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
4872                         getmd);
4873                 lnet_res_unlock(cpt);
4874                 goto drop;
4875         }
4876
4877         LASSERT(getmd->md_offset == 0);
4878
4879         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
4880                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
4881
4882         /* setup information for lnet_build_msg_event */
4883         msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
4884         msg->msg_from = peer_id.nid;
4885         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
4886         msg->msg_hdr.src_nid = peer_id.nid;
4887         msg->msg_hdr.payload_length = getmd->md_length;
4888         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
4889
4890         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
4891         lnet_res_unlock(cpt);
4892
4893         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4894
4895         lnet_net_lock(cpt);
4896         lnet_msg_commit(msg, cpt);
4897         lnet_net_unlock(cpt);
4898
4899         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4900
4901         return msg;
4902
4903  drop:
4904         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
4905
4906         lnet_net_lock(cpt);
4907         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
4908         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4909         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
4910                 getmd->md_length;
4911         lnet_net_unlock(cpt);
4912
4913         if (msg != NULL)
4914                 lnet_msg_free(msg);
4915
4916         return NULL;
4917 }
4918 EXPORT_SYMBOL(lnet_create_reply_msg);
4919
4920 void
4921 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
4922                        unsigned int len)
4923 {
4924         /* Set the REPLY length, now the RDMA that elides the REPLY message has
4925          * completed and I know it. */
4926         LASSERT(reply != NULL);
4927         LASSERT(reply->msg_type == LNET_MSG_GET);
4928         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
4929
4930         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
4931          * the end of my buffer, I might as well be dead. */
4932         LASSERT(len <= reply->msg_ev.mlength);
4933
4934         reply->msg_ev.mlength = len;
4935 }
4936 EXPORT_SYMBOL(lnet_set_reply_msg_len);
4937
4938 /**
4939  * Initiate an asynchronous GET operation.
4940  *
4941  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
4942  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
4943  * the target node in the REPLY has been written to local MD.
4944  *
4945  * On the target node, an LNET_EVENT_GET is logged when the GET request
4946  * arrives and is accepted into a MD.
4947  *
4948  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
4949  * \param mdh A handle for the MD that describes the memory into which the
4950  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
4951  *
4952  * \retval  0      Success, and only in this case events will be generated
4953  * and logged to EQ (if it exists) of the MD.
4954  * \retval -EIO    Simulated failure.
4955  * \retval -ENOMEM Memory allocation failure.
4956  * \retval -ENOENT Invalid MD object.
4957  */
4958 int
4959 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
4960         struct lnet_process_id target, unsigned int portal,
4961         __u64 match_bits, unsigned int offset, bool recovery)
4962 {
4963         struct lnet_msg *msg;
4964         struct lnet_libmd *md;
4965         struct lnet_rsp_tracker *rspt;
4966         int cpt;
4967         int rc;
4968
4969         LASSERT(the_lnet.ln_refcount > 0);
4970
4971         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4972             fail_peer(target.nid, 1))                   /* shall we now? */
4973         {
4974                 CERROR("Dropping GET to %s: simulated failure\n",
4975                        libcfs_id2str(target));
4976                 return -EIO;
4977         }
4978
4979         msg = lnet_msg_alloc();
4980         if (!msg) {
4981                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
4982                        libcfs_id2str(target));
4983                 return -ENOMEM;
4984         }
4985
4986         cpt = lnet_cpt_of_cookie(mdh.cookie);
4987
4988         rspt = lnet_rspt_alloc(cpt);
4989         if (!rspt) {
4990                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
4991                        libcfs_id2str(target));
4992                 return -ENOMEM;
4993         }
4994         INIT_LIST_HEAD(&rspt->rspt_on_list);
4995
4996         msg->msg_recovery = recovery;
4997
4998         lnet_res_lock(cpt);
4999
5000         md = lnet_handle2md(&mdh);
5001         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5002                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5003                        match_bits, portal, libcfs_id2str(target),
5004                        md == NULL ? -1 : md->md_threshold);
5005                 if (md != NULL && md->md_me != NULL)
5006                         CERROR("REPLY MD also attached to portal %d\n",
5007                                md->md_me->me_portal);
5008
5009                 lnet_res_unlock(cpt);
5010
5011                 lnet_msg_free(msg);
5012                 lnet_rspt_free(rspt, cpt);
5013                 return -ENOENT;
5014         }
5015
5016         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5017
5018         lnet_msg_attach_md(msg, md, 0, 0);
5019
5020         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5021
5022         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5023         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5024         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5025         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5026
5027         /* NB handles only looked up by creator (no flips) */
5028         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5029                 the_lnet.ln_interface_cookie;
5030         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5031                 md->md_lh.lh_cookie;
5032
5033         lnet_res_unlock(cpt);
5034
5035         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5036
5037         lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5038
5039         rc = lnet_send(self, msg, LNET_NID_ANY);
5040         if (rc < 0) {
5041                 CNETERR("Error sending GET to %s: %d\n",
5042                         libcfs_id2str(target), rc);
5043                 msg->msg_no_resend = true;
5044                 lnet_finalize(msg, rc);
5045         }
5046
5047         /* completion will be signalled by an event */
5048         return 0;
5049 }
5050 EXPORT_SYMBOL(LNetGet);
5051
5052 /**
5053  * Calculate distance to node at \a dstnid.
5054  *
5055  * \param dstnid Target NID.
5056  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5057  * is saved here.
5058  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5059  * here.
5060  *
5061  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5062  * local_nid_dist_zero is set, which is the default.
5063  * \retval positives Distance to target NID, i.e. number of hops plus one.
5064  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5065  */
5066 int
5067 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5068 {
5069         struct list_head *e;
5070         struct lnet_ni *ni = NULL;
5071         struct lnet_remotenet *rnet;
5072         __u32 dstnet = LNET_NIDNET(dstnid);
5073         int hops;
5074         int cpt;
5075         __u32 order = 2;
5076         struct list_head *rn_list;
5077
5078         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5079          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5080          * keep order 0 free for 0@lo and order 1 free for a local NID
5081          * match */
5082
5083         LASSERT(the_lnet.ln_refcount > 0);
5084
5085         cpt = lnet_net_lock_current();
5086
5087         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5088                 if (ni->ni_nid == dstnid) {
5089                         if (srcnidp != NULL)
5090                                 *srcnidp = dstnid;
5091                         if (orderp != NULL) {
5092                                 if (dstnid == LNET_NID_LO_0)
5093                                         *orderp = 0;
5094                                 else
5095                                         *orderp = 1;
5096                         }
5097                         lnet_net_unlock(cpt);
5098
5099                         return local_nid_dist_zero ? 0 : 1;
5100                 }
5101
5102                 if (LNET_NIDNET(ni->ni_nid) == dstnet) {
5103                         /* Check if ni was originally created in
5104                          * current net namespace.
5105                          * If not, assign order above 0xffff0000,
5106                          * to make this ni not a priority. */
5107                         if (current->nsproxy &&
5108                             !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
5109                                         order += 0xffff0000;
5110                         if (srcnidp != NULL)
5111                                 *srcnidp = ni->ni_nid;
5112                         if (orderp != NULL)
5113                                 *orderp = order;
5114                         lnet_net_unlock(cpt);
5115                         return 1;
5116                 }
5117
5118                 order++;
5119         }
5120
5121         rn_list = lnet_net2rnethash(dstnet);
5122         list_for_each(e, rn_list) {
5123                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5124
5125                 if (rnet->lrn_net == dstnet) {
5126                         struct lnet_route *route;
5127                         struct lnet_route *shortest = NULL;
5128                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5129                         __u32 route_hops;
5130
5131                         LASSERT(!list_empty(&rnet->lrn_routes));
5132
5133                         list_for_each_entry(route, &rnet->lrn_routes,
5134                                             lr_list) {
5135                                 route_hops = route->lr_hops;
5136                                 if (route_hops == LNET_UNDEFINED_HOPS)
5137                                         route_hops = 1;
5138                                 if (shortest == NULL ||
5139                                     route_hops < shortest_hops) {
5140                                         shortest = route;
5141                                         shortest_hops = route_hops;
5142                                 }
5143                         }
5144
5145                         LASSERT(shortest != NULL);
5146                         hops = shortest_hops;
5147                         if (srcnidp != NULL) {
5148                                 struct lnet_net *net;
5149                                 net = lnet_get_net_locked(shortest->lr_lnet);
5150                                 LASSERT(net);
5151                                 ni = lnet_get_next_ni_locked(net, NULL);
5152                                 *srcnidp = ni->ni_nid;
5153                         }
5154                         if (orderp != NULL)
5155                                 *orderp = order;
5156                         lnet_net_unlock(cpt);
5157                         return hops + 1;
5158                 }
5159                 order++;
5160         }
5161
5162         lnet_net_unlock(cpt);
5163         return -EHOSTUNREACH;
5164 }
5165 EXPORT_SYMBOL(LNetDist);