lnet/lnet/lib-move.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see [sun.com URL with a
  20  * copy of GPLv2].
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/lnet/lib-move.c
  37  *
  38  * Data movement routines
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include <lnet/lib-lnet.h>
  44
  45 static int local_nid_dist_zero = 1;
  46 CFS_MODULE_PARM(local_nid_dist_zero, "i", int, 0444,
  47                 "Reserved");
  48
  49 /* forward ref */
  50 static void lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg);
  51 static void lnet_drop_delayed_put(lnet_msg_t *msg, char *reason);
  52
  53 #define LNET_MATCHMD_NONE     0   /* Didn't match */
  54 #define LNET_MATCHMD_OK       1   /* Matched OK */
  55 #define LNET_MATCHMD_DROP     2   /* Must be discarded */
  56
  57 static int
  58 lnet_try_match_md (int index, int op_mask, lnet_process_id_t src,
  59                    unsigned int rlength, unsigned int roffset,
  60                    __u64 match_bits, lnet_libmd_t *md, lnet_msg_t *msg,
  61                    unsigned int *mlength_out, unsigned int *offset_out)
  62 {
  63         /* ALWAYS called holding the LNET_LOCK, and can't LNET_UNLOCK;
  64          * lnet_match_blocked_msg() relies on this to avoid races */
  65         unsigned int  offset;
  66         unsigned int  mlength;
  67         lnet_me_t    *me = md->md_me;
  68
  69         /* mismatched MD op */
  70         if ((md->md_options & op_mask) == 0)
  71                 return LNET_MATCHMD_NONE;
  72
  73         /* MD exhausted */
  74         if (lnet_md_exhausted(md))
  75                 return LNET_MATCHMD_NONE;
  76
  77         /* mismatched ME nid/pid? */
  78         if (me->me_match_id.nid != LNET_NID_ANY &&
  79             me->me_match_id.nid != src.nid)
  80                 return LNET_MATCHMD_NONE;
  81
  82         if (me->me_match_id.pid != LNET_PID_ANY &&
  83             me->me_match_id.pid != src.pid)
  84                 return LNET_MATCHMD_NONE;
  85
  86         /* mismatched ME matchbits? */
  87         if (((me->me_match_bits ^ match_bits) & ~me->me_ignore_bits) != 0)
  88                 return LNET_MATCHMD_NONE;
  89
  90         /* Hurrah! This _is_ a match; check it out... */
  91
  92         if ((md->md_options & LNET_MD_MANAGE_REMOTE) == 0)
  93                 offset = md->md_offset;
  94         else
  95                 offset = roffset;
  96
  97         if ((md->md_options & LNET_MD_MAX_SIZE) != 0) {
  98                 mlength = md->md_max_size;
  99                 LASSERT (md->md_offset + mlength <= md->md_length);
 100         } else {
 101                 mlength = md->md_length - offset;
 102         }
 103
 104         if (rlength <= mlength) {        /* fits in allowed space */
 105                 mlength = rlength;
 106         } else if ((md->md_options & LNET_MD_TRUNCATE) == 0) {
 107                 /* this packet _really_ is too big */
 108                 CERROR("Matching packet from %s, match "LPU64
 109                        " length %d too big: %d left, %d allowed\n",
 110                        libcfs_id2str(src), match_bits, rlength,
 111                        md->md_length - offset, mlength);
 112
 113                 return LNET_MATCHMD_DROP;
 114         }
 115
 116         /* Commit to this ME/MD */
 117         CDEBUG(D_NET, "Incoming %s index %x from %s of "
 118                "length %d/%d into md "LPX64" [%d] + %d\n",
 119                (op_mask == LNET_MD_OP_PUT) ? "put" : "get",
 120                index, libcfs_id2str(src), mlength, rlength,
 121                md->md_lh.lh_cookie, md->md_niov, offset);
 122
 123         lnet_commit_md(md, msg);
 124         md->md_offset = offset + mlength;
 125
 126         /* NB Caller will set ev.type and ev.hdr_data */
 127         msg->msg_ev.initiator = src;
 128         msg->msg_ev.pt_index = index;
 129         msg->msg_ev.match_bits = match_bits;
 130         msg->msg_ev.rlength = rlength;
 131         msg->msg_ev.mlength = mlength;
 132         msg->msg_ev.offset = offset;
 133
 134         lnet_md_deconstruct(md, &msg->msg_ev.md);
 135         lnet_md2handle(&msg->msg_ev.md_handle, md);
 136
 137         *offset_out = offset;
 138         *mlength_out = mlength;
 139
 140         /* Auto-unlink NOW, so the ME gets unlinked if required.
 141          * We bumped md->md_refcount above so the MD just gets flagged
 142          * for unlink when it is finalized. */
 143         if ((md->md_flags & LNET_MD_FLAG_AUTO_UNLINK) != 0 &&
 144             lnet_md_exhausted(md)) {
 145                 lnet_md_unlink(md);
 146         }
 147
 148         return LNET_MATCHMD_OK;
 149 }
 150
 151 static int
 152 lnet_match_md(int index, int op_mask, lnet_process_id_t src,
 153               unsigned int rlength, unsigned int roffset,
 154               __u64 match_bits, lnet_msg_t *msg,
 155               unsigned int *mlength_out, unsigned int *offset_out,
 156               lnet_libmd_t **md_out)
 157 {
 158         lnet_portal_t    *ptl = &the_lnet.ln_portals[index];
 159         lnet_me_t        *me;
 160         lnet_me_t        *tmp;
 161         lnet_libmd_t     *md;
 162         int               rc;
 163
 164         CDEBUG (D_NET, "Request from %s of length %d into portal %d "
 165                 "MB="LPX64"\n", libcfs_id2str(src), rlength, index, match_bits);
 166
 167         if (index < 0 || index >= the_lnet.ln_nportals) {
 168                 CERROR("Invalid portal %d not in [0-%d]\n",
 169                        index, the_lnet.ln_nportals);
 170                 return LNET_MATCHMD_DROP;
 171         }
 172
 173         list_for_each_entry_safe (me, tmp, &ptl->ptl_ml, me_list) {
 174                 md = me->me_md;
 175
 176                 /* ME attached but MD not attached yet */
 177                 if (md == NULL)
 178                         continue;
 179
 180                 LASSERT (me == md->md_me);
 181
 182                 rc = lnet_try_match_md(index, op_mask, src, rlength,
 183                                        roffset, match_bits, md, msg,
 184                                        mlength_out, offset_out);
 185                 switch (rc) {
 186                 default:
 187                         LBUG();
 188
 189                 case LNET_MATCHMD_NONE:
 190                         continue;
 191
 192                 case LNET_MATCHMD_OK:
 193                         *md_out = md;
 194                         return LNET_MATCHMD_OK;
 195
 196                 case LNET_MATCHMD_DROP:
 197                         return LNET_MATCHMD_DROP;
 198                 }
 199                 /* not reached */
 200         }
 201
 202         if (op_mask == LNET_MD_OP_GET ||
 203             (ptl->ptl_options & LNET_PTL_LAZY) == 0)
 204                 return LNET_MATCHMD_DROP;
 205
 206         return LNET_MATCHMD_NONE;
 207 }
 208
 209 int
 210 lnet_fail_nid (lnet_nid_t nid, unsigned int threshold)
 211 {
 212         lnet_test_peer_t   *tp;
 213         struct list_head  *el;
 214         struct list_head  *next;
 215         struct list_head   cull;
 216
 217         LASSERT (the_lnet.ln_init);
 218
 219         if (threshold != 0) {
 220                 /* Adding a new entry */
 221                 LIBCFS_ALLOC(tp, sizeof(*tp));
 222                 if (tp == NULL)
 223                         return -ENOMEM;
 224
 225                 tp->tp_nid = nid;
 226                 tp->tp_threshold = threshold;
 227
 228                 LNET_LOCK();
 229                 list_add_tail (&tp->tp_list, &the_lnet.ln_test_peers);
 230                 LNET_UNLOCK();
 231                 return 0;
 232         }
 233
 234         /* removing entries */
 235         CFS_INIT_LIST_HEAD (&cull);
 236
 237         LNET_LOCK();
 238
 239         list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
 240                 tp = list_entry (el, lnet_test_peer_t, tp_list);
 241
 242                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 243                     nid == LNET_NID_ANY ||       /* removing all entries */
 244                     tp->tp_nid == nid)          /* matched this one */
 245                 {
 246                         list_del (&tp->tp_list);
 247                         list_add (&tp->tp_list, &cull);
 248                 }
 249         }
 250
 251         LNET_UNLOCK();
 252
 253         while (!list_empty (&cull)) {
 254                 tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
 255
 256                 list_del (&tp->tp_list);
 257                 LIBCFS_FREE(tp, sizeof (*tp));
 258         }
 259         return 0;
 260 }
 261
 262 static int
 263 fail_peer (lnet_nid_t nid, int outgoing)
 264 {
 265         lnet_test_peer_t  *tp;
 266         struct list_head *el;
 267         struct list_head *next;
 268         struct list_head  cull;
 269         int               fail = 0;
 270
 271         CFS_INIT_LIST_HEAD (&cull);
 272
 273         LNET_LOCK();
 274
 275         list_for_each_safe (el, next, &the_lnet.ln_test_peers) {
 276                 tp = list_entry (el, lnet_test_peer_t, tp_list);
 277
 278                 if (tp->tp_threshold == 0) {
 279                         /* zombie entry */
 280                         if (outgoing) {
 281                                 /* only cull zombies on outgoing tests,
 282                                  * since we may be at interrupt priority on
 283                                  * incoming messages. */
 284                                 list_del (&tp->tp_list);
 285                                 list_add (&tp->tp_list, &cull);
 286                         }
 287                         continue;
 288                 }
 289
 290                 if (tp->tp_nid == LNET_NID_ANY || /* fail every peer */
 291                     nid == tp->tp_nid) {        /* fail this peer */
 292                         fail = 1;
 293
 294                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 295                                 tp->tp_threshold--;
 296                                 if (outgoing &&
 297                                     tp->tp_threshold == 0) {
 298                                         /* see above */
 299                                         list_del (&tp->tp_list);
 300                                         list_add (&tp->tp_list, &cull);
 301                                 }
 302                         }
 303                         break;
 304                 }
 305         }
 306
 307         LNET_UNLOCK ();
 308
 309         while (!list_empty (&cull)) {
 310                 tp = list_entry (cull.next, lnet_test_peer_t, tp_list);
 311                 list_del (&tp->tp_list);
 312
 313                 LIBCFS_FREE(tp, sizeof (*tp));
 314         }
 315
 316         return (fail);
 317 }
 318
 319 unsigned int
 320 lnet_iov_nob (unsigned int niov, struct iovec *iov)
 321 {
 322         unsigned int nob = 0;
 323
 324         while (niov-- > 0)
 325                 nob += (iov++)->iov_len;
 326
 327         return (nob);
 328 }
 329
 330 void
 331 lnet_copy_iov2iov (unsigned int ndiov, struct iovec *diov, unsigned int doffset,
 332                    unsigned int nsiov, struct iovec *siov, unsigned int soffset,
 333                    unsigned int nob)
 334 {
 335         /* NB diov, siov are READ-ONLY */
 336         unsigned int  this_nob;
 337
 338         if (nob == 0)
 339                 return;
 340
 341         /* skip complete frags before 'doffset' */
 342         LASSERT (ndiov > 0);
 343         while (doffset >= diov->iov_len) {
 344                 doffset -= diov->iov_len;
 345                 diov++;
 346                 ndiov--;
 347                 LASSERT (ndiov > 0);
 348         }
 349
 350         /* skip complete frags before 'soffset' */
 351         LASSERT (nsiov > 0);
 352         while (soffset >= siov->iov_len) {
 353                 soffset -= siov->iov_len;
 354                 siov++;
 355                 nsiov--;
 356                 LASSERT (nsiov > 0);
 357         }
 358
 359         do {
 360                 LASSERT (ndiov > 0);
 361                 LASSERT (nsiov > 0);
 362                 this_nob = MIN(diov->iov_len - doffset,
 363                                siov->iov_len - soffset);
 364                 this_nob = MIN(this_nob, nob);
 365
 366                 memcpy ((char *)diov->iov_base + doffset,
 367                         (char *)siov->iov_base + soffset, this_nob);
 368                 nob -= this_nob;
 369
 370                 if (diov->iov_len > doffset + this_nob) {
 371                         doffset += this_nob;
 372                 } else {
 373                         diov++;
 374                         ndiov--;
 375                         doffset = 0;
 376                 }
 377
 378                 if (siov->iov_len > soffset + this_nob) {
 379                         soffset += this_nob;
 380                 } else {
 381                         siov++;
 382                         nsiov--;
 383                         soffset = 0;
 384                 }
 385         } while (nob > 0);
 386 }
 387
 388 int
 389 lnet_extract_iov (int dst_niov, struct iovec *dst,
 390                   int src_niov, struct iovec *src,
 391                   unsigned int offset, unsigned int len)
 392 {
 393         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 394          * for exactly 'len' bytes, and return the number of entries.
 395          * NB not destructive to 'src' */
 396         unsigned int    frag_len;
 397         unsigned int    niov;
 398
 399         if (len == 0)                           /* no data => */
 400                 return (0);                     /* no frags */
 401
 402         LASSERT (src_niov > 0);
 403         while (offset >= src->iov_len) {      /* skip initial frags */
 404                 offset -= src->iov_len;
 405                 src_niov--;
 406                 src++;
 407                 LASSERT (src_niov > 0);
 408         }
 409
 410         niov = 1;
 411         for (;;) {
 412                 LASSERT (src_niov > 0);
 413                 LASSERT (niov <= dst_niov);
 414
 415                 frag_len = src->iov_len - offset;
 416                 dst->iov_base = ((char *)src->iov_base) + offset;
 417
 418                 if (len <= frag_len) {
 419                         dst->iov_len = len;
 420                         return (niov);
 421                 }
 422
 423                 dst->iov_len = frag_len;
 424
 425                 len -= frag_len;
 426                 dst++;
 427                 src++;
 428                 niov++;
 429                 src_niov--;
 430                 offset = 0;
 431         }
 432 }
 433
 434 #ifndef __KERNEL__
 435 unsigned int
 436 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
 437 {
 438         LASSERT (0);
 439         return (0);
 440 }
 441
 442 void
 443 lnet_copy_kiov2kiov (unsigned int ndkiov, lnet_kiov_t *dkiov, unsigned int doffset,
 444                      unsigned int nskiov, lnet_kiov_t *skiov, unsigned int soffset,
 445                      unsigned int nob)
 446 {
 447         LASSERT (0);
 448 }
 449
 450 void
 451 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
 452                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 453                     unsigned int nob)
 454 {
 455         LASSERT (0);
 456 }
 457
 458 void
 459 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 460                     unsigned int niov, struct iovec *iov, unsigned int iovoffset,
 461                     unsigned int nob)
 462 {
 463         LASSERT (0);
 464 }
 465
 466 int
 467 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
 468                    int src_niov, lnet_kiov_t *src,
 469                    unsigned int offset, unsigned int len)
 470 {
 471         LASSERT (0);
 472 }
 473
 474 #else /* __KERNEL__ */
 475
 476 unsigned int
 477 lnet_kiov_nob (unsigned int niov, lnet_kiov_t *kiov)
 478 {
 479         unsigned int  nob = 0;
 480
 481         while (niov-- > 0)
 482                 nob += (kiov++)->kiov_len;
 483
 484         return (nob);
 485 }
 486
 487 void
 488 lnet_copy_kiov2kiov (unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
 489                      unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
 490                      unsigned int nob)
 491 {
 492         /* NB diov, siov are READ-ONLY */
 493         unsigned int    this_nob;
 494         char           *daddr = NULL;
 495         char           *saddr = NULL;
 496
 497         if (nob == 0)
 498                 return;
 499
 500         LASSERT (!in_interrupt ());
 501
 502         LASSERT (ndiov > 0);
 503         while (doffset >= diov->kiov_len) {
 504                 doffset -= diov->kiov_len;
 505                 diov++;
 506                 ndiov--;
 507                 LASSERT (ndiov > 0);
 508         }
 509
 510         LASSERT (nsiov > 0);
 511         while (soffset >= siov->kiov_len) {
 512                 soffset -= siov->kiov_len;
 513                 siov++;
 514                 nsiov--;
 515                 LASSERT (nsiov > 0);
 516         }
 517
 518         do {
 519                 LASSERT (ndiov > 0);
 520                 LASSERT (nsiov > 0);
 521                 this_nob = MIN(diov->kiov_len - doffset,
 522                                siov->kiov_len - soffset);
 523                 this_nob = MIN(this_nob, nob);
 524
 525                 if (daddr == NULL)
 526                         daddr = ((char *)cfs_kmap(diov->kiov_page)) +
 527                                 diov->kiov_offset + doffset;
 528                 if (saddr == NULL)
 529                         saddr = ((char *)cfs_kmap(siov->kiov_page)) +
 530                                 siov->kiov_offset + soffset;
 531
 532                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 533                  * However in practice at least one of the kiovs will be mapped
 534                  * kernel pages and the map/unmap will be NOOPs */
 535
 536                 memcpy (daddr, saddr, this_nob);
 537                 nob -= this_nob;
 538
 539                 if (diov->kiov_len > doffset + this_nob) {
 540                         daddr += this_nob;
 541                         doffset += this_nob;
 542                 } else {
 543                         cfs_kunmap(diov->kiov_page);
 544                         daddr = NULL;
 545                         diov++;
 546                         ndiov--;
 547                         doffset = 0;
 548                 }
 549
 550                 if (siov->kiov_len > soffset + this_nob) {
 551                         saddr += this_nob;
 552                         soffset += this_nob;
 553                 } else {
 554                         cfs_kunmap(siov->kiov_page);
 555                         saddr = NULL;
 556                         siov++;
 557                         nsiov--;
 558                         soffset = 0;
 559                 }
 560         } while (nob > 0);
 561
 562         if (daddr != NULL)
 563                 cfs_kunmap(diov->kiov_page);
 564         if (saddr != NULL)
 565                 cfs_kunmap(siov->kiov_page);
 566 }
 567
 568 void
 569 lnet_copy_kiov2iov (unsigned int niov, struct iovec *iov, unsigned int iovoffset,
 570                     unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 571                     unsigned int nob)
 572 {
 573         /* NB iov, kiov are READ-ONLY */
 574         unsigned int    this_nob;
 575         char           *addr = NULL;
 576
 577         if (nob == 0)
 578                 return;
 579
 580         LASSERT (!in_interrupt ());
 581
 582         LASSERT (niov > 0);
 583         while (iovoffset >= iov->iov_len) {
 584                 iovoffset -= iov->iov_len;
 585                 iov++;
 586                 niov--;
 587                 LASSERT (niov > 0);
 588         }
 589
 590         LASSERT (nkiov > 0);
 591         while (kiovoffset >= kiov->kiov_len) {
 592                 kiovoffset -= kiov->kiov_len;
 593                 kiov++;
 594                 nkiov--;
 595                 LASSERT (nkiov > 0);
 596         }
 597
 598         do {
 599                 LASSERT (niov > 0);
 600                 LASSERT (nkiov > 0);
 601                 this_nob = MIN(iov->iov_len - iovoffset,
 602                                kiov->kiov_len - kiovoffset);
 603                 this_nob = MIN(this_nob, nob);
 604
 605                 if (addr == NULL)
 606                         addr = ((char *)cfs_kmap(kiov->kiov_page)) +
 607                                 kiov->kiov_offset + kiovoffset;
 608
 609                 memcpy ((char *)iov->iov_base + iovoffset, addr, this_nob);
 610                 nob -= this_nob;
 611
 612                 if (iov->iov_len > iovoffset + this_nob) {
 613                         iovoffset += this_nob;
 614                 } else {
 615                         iov++;
 616                         niov--;
 617                         iovoffset = 0;
 618                 }
 619
 620                 if (kiov->kiov_len > kiovoffset + this_nob) {
 621                         addr += this_nob;
 622                         kiovoffset += this_nob;
 623                 } else {
 624                         cfs_kunmap(kiov->kiov_page);
 625                         addr = NULL;
 626                         kiov++;
 627                         nkiov--;
 628                         kiovoffset = 0;
 629                 }
 630
 631         } while (nob > 0);
 632
 633         if (addr != NULL)
 634                 cfs_kunmap(kiov->kiov_page);
 635 }
 636
 637 void
 638 lnet_copy_iov2kiov (unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
 639                     unsigned int niov, struct iovec *iov, unsigned int iovoffset,
 640                     unsigned int nob)
 641 {
 642         /* NB kiov, iov are READ-ONLY */
 643         unsigned int    this_nob;
 644         char           *addr = NULL;
 645
 646         if (nob == 0)
 647                 return;
 648
 649         LASSERT (!in_interrupt ());
 650
 651         LASSERT (nkiov > 0);
 652         while (kiovoffset >= kiov->kiov_len) {
 653                 kiovoffset -= kiov->kiov_len;
 654                 kiov++;
 655                 nkiov--;
 656                 LASSERT (nkiov > 0);
 657         }
 658
 659         LASSERT (niov > 0);
 660         while (iovoffset >= iov->iov_len) {
 661                 iovoffset -= iov->iov_len;
 662                 iov++;
 663                 niov--;
 664                 LASSERT (niov > 0);
 665         }
 666
 667         do {
 668                 LASSERT (nkiov > 0);
 669                 LASSERT (niov > 0);
 670                 this_nob = MIN(kiov->kiov_len - kiovoffset,
 671                                iov->iov_len - iovoffset);
 672                 this_nob = MIN(this_nob, nob);
 673
 674                 if (addr == NULL)
 675                         addr = ((char *)cfs_kmap(kiov->kiov_page)) +
 676                                 kiov->kiov_offset + kiovoffset;
 677
 678                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 679                 nob -= this_nob;
 680
 681                 if (kiov->kiov_len > kiovoffset + this_nob) {
 682                         addr += this_nob;
 683                         kiovoffset += this_nob;
 684                 } else {
 685                         cfs_kunmap(kiov->kiov_page);
 686                         addr = NULL;
 687                         kiov++;
 688                         nkiov--;
 689                         kiovoffset = 0;
 690                 }
 691
 692                 if (iov->iov_len > iovoffset + this_nob) {
 693                         iovoffset += this_nob;
 694                 } else {
 695                         iov++;
 696                         niov--;
 697                         iovoffset = 0;
 698                 }
 699         } while (nob > 0);
 700
 701         if (addr != NULL)
 702                 cfs_kunmap(kiov->kiov_page);
 703 }
 704
 705 int
 706 lnet_extract_kiov (int dst_niov, lnet_kiov_t *dst,
 707                    int src_niov, lnet_kiov_t *src,
 708                    unsigned int offset, unsigned int len)
 709 {
 710         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 711          * for exactly 'len' bytes, and return the number of entries.
 712          * NB not destructive to 'src' */
 713         unsigned int    frag_len;
 714         unsigned int    niov;
 715
 716         if (len == 0)                           /* no data => */
 717                 return (0);                     /* no frags */
 718
 719         LASSERT (src_niov > 0);
 720         while (offset >= src->kiov_len) {      /* skip initial frags */
 721                 offset -= src->kiov_len;
 722                 src_niov--;
 723                 src++;
 724                 LASSERT (src_niov > 0);
 725         }
 726
 727         niov = 1;
 728         for (;;) {
 729                 LASSERT (src_niov > 0);
 730                 LASSERT (niov <= dst_niov);
 731
 732                 frag_len = src->kiov_len - offset;
 733                 dst->kiov_page = src->kiov_page;
 734                 dst->kiov_offset = src->kiov_offset + offset;
 735
 736                 if (len <= frag_len) {
 737                         dst->kiov_len = len;
 738                         LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
 739                         return (niov);
 740                 }
 741
 742                 dst->kiov_len = frag_len;
 743                 LASSERT (dst->kiov_offset + dst->kiov_len <= CFS_PAGE_SIZE);
 744
 745                 len -= frag_len;
 746                 dst++;
 747                 src++;
 748                 niov++;
 749                 src_niov--;
 750                 offset = 0;
 751         }
 752 }
 753 #endif
 754
 755 void
 756 lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
 757              unsigned int offset, unsigned int mlen, unsigned int rlen)
 758 {
 759         unsigned int  niov = 0;
 760         struct iovec *iov = NULL;
 761         lnet_kiov_t  *kiov = NULL;
 762         int           rc;
 763
 764         LASSERT (!in_interrupt ());
 765         LASSERT (mlen == 0 || msg != NULL);
 766
 767         if (msg != NULL) {
 768                 LASSERT(msg->msg_receiving);
 769                 LASSERT(!msg->msg_sending);
 770                 LASSERT(rlen == msg->msg_len);
 771                 LASSERT(mlen <= msg->msg_len);
 772
 773                 msg->msg_wanted = mlen;
 774                 msg->msg_offset = offset;
 775                 msg->msg_receiving = 0;
 776
 777                 if (mlen != 0) {
 778                         niov = msg->msg_niov;
 779                         iov  = msg->msg_iov;
 780                         kiov = msg->msg_kiov;
 781
 782                         LASSERT (niov > 0);
 783                         LASSERT ((iov == NULL) != (kiov == NULL));
 784                 }
 785         }
 786
 787         rc = (ni->ni_lnd->lnd_recv)(ni, private, msg, delayed,
 788                                     niov, iov, kiov, offset, mlen, rlen);
 789         if (rc < 0)
 790                 lnet_finalize(ni, msg, rc);
 791 }
 792
 793 int
 794 lnet_compare_routers(lnet_peer_t *p1, lnet_peer_t *p2)
 795 {
 796         if (p1->lp_txqnob < p2->lp_txqnob)
 797                 return 1;
 798
 799         if (p1->lp_txqnob > p2->lp_txqnob)
 800                 return -1;
 801
 802         if (p1->lp_txcredits > p2->lp_txcredits)
 803                 return 1;
 804
 805         if (p1->lp_txcredits < p2->lp_txcredits)
 806                 return -1;
 807
 808         return 0;
 809 }
 810
 811
 812 void
 813 lnet_setpayloadbuffer(lnet_msg_t *msg)
 814 {
 815         lnet_libmd_t *md = msg->msg_md;
 816
 817         LASSERT (msg->msg_len > 0);
 818         LASSERT (!msg->msg_routing);
 819         LASSERT (md != NULL);
 820         LASSERT (msg->msg_niov == 0);
 821         LASSERT (msg->msg_iov == NULL);
 822         LASSERT (msg->msg_kiov == NULL);
 823
 824         msg->msg_niov = md->md_niov;
 825         if ((md->md_options & LNET_MD_KIOV) != 0)
 826                 msg->msg_kiov = md->md_iov.kiov;
 827         else
 828                 msg->msg_iov = md->md_iov.iov;
 829 }
 830
 831 void
 832 lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
 833                unsigned int offset, unsigned int len)
 834 {
 835         msg->msg_type = type;
 836         msg->msg_target = target;
 837         msg->msg_len = len;
 838         msg->msg_offset = offset;
 839
 840         if (len != 0)
 841                 lnet_setpayloadbuffer(msg);
 842
 843         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 844         msg->msg_hdr.type           = cpu_to_le32(type);
 845         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 846         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 847         /* src_nid will be set later */
 848         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 849         msg->msg_hdr.payload_length = cpu_to_le32(len);
 850 }
 851
 852 void
 853 lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
 854 {
 855         void   *priv = msg->msg_private;
 856         int     rc;
 857
 858         LASSERT (!in_interrupt ());
 859         LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
 860                  (msg->msg_txcredit && msg->msg_peertxcredit));
 861
 862         rc = (ni->ni_lnd->lnd_send)(ni, priv, msg);
 863         if (rc < 0)
 864                 lnet_finalize(ni, msg, rc);
 865 }
 866
 867 int
 868 lnet_eager_recv_locked(lnet_msg_t *msg)
 869 {
 870         lnet_peer_t *peer;
 871         lnet_ni_t   *ni;
 872         int          rc = 0;
 873
 874         LASSERT (!msg->msg_delayed);
 875         msg->msg_delayed = 1;
 876
 877         LASSERT (msg->msg_receiving);
 878         LASSERT (!msg->msg_sending);
 879
 880         peer = msg->msg_rxpeer;
 881         ni   = peer->lp_ni;
 882
 883         if (ni->ni_lnd->lnd_eager_recv != NULL) {
 884                 LNET_UNLOCK();
 885
 886                 rc = (ni->ni_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 887                                                   &msg->msg_private);
 888                 if (rc != 0) {
 889                         CERROR("recv from %s / send to %s aborted: "
 890                                "eager_recv failed %d\n",
 891                                libcfs_nid2str(peer->lp_nid),
 892                                libcfs_id2str(msg->msg_target), rc);
 893                         LASSERT (rc < 0); /* required by my callers */
 894                 }
 895
 896                 LNET_LOCK();
 897         }
 898
 899         return rc;
 900 }
 901
 902 int
 903 lnet_post_send_locked (lnet_msg_t *msg, int do_send)
 904 {
 905         /* lnet_send is going to LNET_UNLOCK immediately after this, so it sets
 906          * do_send FALSE and I don't do the unlock/send/lock bit.  I return
 907          * EAGAIN if msg blocked and 0 if sent or OK to send */
 908         lnet_peer_t *lp = msg->msg_txpeer;
 909         lnet_ni_t   *ni = lp->lp_ni;
 910
 911         /* non-lnet_send() callers have checked before */
 912         LASSERT (!do_send || msg->msg_delayed);
 913         LASSERT (!msg->msg_receiving);
 914
 915         if (!msg->msg_peertxcredit) {
 916                 LASSERT ((lp->lp_txcredits < 0) == !list_empty(&lp->lp_txq));
 917
 918                 msg->msg_peertxcredit = 1;
 919                 lp->lp_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
 920                 lp->lp_txcredits--;
 921
 922                 if (lp->lp_txcredits < lp->lp_mintxcredits)
 923                         lp->lp_mintxcredits = lp->lp_txcredits;
 924
 925                 if (lp->lp_txcredits < 0) {
 926                         msg->msg_delayed = 1;
 927                         list_add_tail (&msg->msg_list, &lp->lp_txq);
 928                         return EAGAIN;
 929                 }
 930         }
 931
 932         if (!msg->msg_txcredit) {
 933                 LASSERT ((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
 934
 935                 msg->msg_txcredit = 1;
 936                 ni->ni_txcredits--;
 937
 938                 if (ni->ni_txcredits < ni->ni_mintxcredits)
 939                         ni->ni_mintxcredits = ni->ni_txcredits;
 940
 941                 if (ni->ni_txcredits < 0) {
 942                         msg->msg_delayed = 1;
 943                         list_add_tail (&msg->msg_list, &ni->ni_txq);
 944                         return EAGAIN;
 945                 }
 946         }
 947
 948         if (do_send) {
 949                 LNET_UNLOCK();
 950                 lnet_ni_send(ni, msg);
 951                 LNET_LOCK();
 952         }
 953         return 0;
 954 }
 955
 956 #ifdef __KERNEL__
 957 static void
 958 lnet_commit_routedmsg (lnet_msg_t *msg)
 959 {
 960         /* ALWAYS called holding the LNET_LOCK */
 961         LASSERT (msg->msg_routing);
 962
 963         the_lnet.ln_counters.msgs_alloc++;
 964         if (the_lnet.ln_counters.msgs_alloc >
 965             the_lnet.ln_counters.msgs_max)
 966                 the_lnet.ln_counters.msgs_max =
 967                         the_lnet.ln_counters.msgs_alloc;
 968
 969         the_lnet.ln_counters.route_count++;
 970         the_lnet.ln_counters.route_length += msg->msg_len;
 971
 972         LASSERT (!msg->msg_onactivelist);
 973         msg->msg_onactivelist = 1;
 974         list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
 975 }
 976
 977 lnet_rtrbufpool_t *
 978 lnet_msg2bufpool(lnet_msg_t *msg)
 979 {
 980         lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[0];
 981
 982         LASSERT (msg->msg_len <= LNET_MTU);
 983         while (msg->msg_len > rbp->rbp_npages * CFS_PAGE_SIZE) {
 984                 rbp++;
 985                 LASSERT (rbp < &the_lnet.ln_rtrpools[LNET_NRBPOOLS]);
 986         }
 987
 988         return rbp;
 989 }
 990
 991 int
 992 lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
 993 {
 994         /* lnet_parse is going to LNET_UNLOCK immediately after this, so it
 995          * sets do_recv FALSE and I don't do the unlock/send/lock bit.  I
 996          * return EAGAIN if msg blocked and 0 if sent or OK to send */
 997         lnet_peer_t         *lp = msg->msg_rxpeer;
 998         lnet_rtrbufpool_t   *rbp;
 999         lnet_rtrbuf_t       *rb;
1000
1001         LASSERT (msg->msg_iov == NULL);
1002         LASSERT (msg->msg_kiov == NULL);
1003         LASSERT (msg->msg_niov == 0);
1004         LASSERT (msg->msg_routing);
1005         LASSERT (msg->msg_receiving);
1006         LASSERT (!msg->msg_sending);
1007
1008         /* non-lnet_parse callers only send delayed messages */
1009         LASSERT (!do_recv || msg->msg_delayed);
1010
1011         if (!msg->msg_peerrtrcredit) {
1012                 LASSERT ((lp->lp_rtrcredits < 0) == !list_empty(&lp->lp_rtrq));
1013
1014                 msg->msg_peerrtrcredit = 1;
1015                 lp->lp_rtrcredits--;
1016                 if (lp->lp_rtrcredits < lp->lp_minrtrcredits)
1017                         lp->lp_minrtrcredits = lp->lp_rtrcredits;
1018
1019                 if (lp->lp_rtrcredits < 0) {
1020                         /* must have checked eager_recv before here */
1021                         LASSERT (msg->msg_delayed);
1022                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1023                         return EAGAIN;
1024                 }
1025         }
1026
1027         rbp = lnet_msg2bufpool(msg);
1028
1029         if (!msg->msg_rtrcredit) {
1030                 LASSERT ((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
1031
1032                 msg->msg_rtrcredit = 1;
1033                 rbp->rbp_credits--;
1034                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1035                         rbp->rbp_mincredits = rbp->rbp_credits;
1036
1037                 if (rbp->rbp_credits < 0) {
1038                         /* must have checked eager_recv before here */
1039                         LASSERT (msg->msg_delayed);
1040                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1041                         return EAGAIN;
1042                 }
1043         }
1044
1045         LASSERT (!list_empty(&rbp->rbp_bufs));
1046         rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
1047         list_del(&rb->rb_list);
1048
1049         msg->msg_niov = rbp->rbp_npages;
1050         msg->msg_kiov = &rb->rb_kiov[0];
1051
1052         if (do_recv) {
1053                 LNET_UNLOCK();
1054                 lnet_ni_recv(lp->lp_ni, msg->msg_private, msg, 1,
1055                              0, msg->msg_len, msg->msg_len);
1056                 LNET_LOCK();
1057         }
1058         return 0;
1059 }
1060 #endif
1061
1062 void
1063 lnet_return_credits_locked (lnet_msg_t *msg)
1064 {
1065         lnet_peer_t       *txpeer = msg->msg_txpeer;
1066         lnet_peer_t       *rxpeer = msg->msg_rxpeer;
1067         lnet_msg_t        *msg2;
1068         lnet_ni_t         *ni;
1069
1070         if (msg->msg_txcredit) {
1071                 /* give back NI txcredits */
1072                 msg->msg_txcredit = 0;
1073                 ni = txpeer->lp_ni;
1074
1075                 LASSERT((ni->ni_txcredits < 0) == !list_empty(&ni->ni_txq));
1076
1077                 ni->ni_txcredits++;
1078                 if (ni->ni_txcredits <= 0) {
1079                         msg2 = list_entry(ni->ni_txq.next, lnet_msg_t, msg_list);
1080                         list_del(&msg2->msg_list);
1081
1082                         LASSERT(msg2->msg_txpeer->lp_ni == ni);
1083                         LASSERT(msg2->msg_delayed);
1084
1085                         (void) lnet_post_send_locked(msg2, 1);
1086                 }
1087         }
1088
1089         if (msg->msg_peertxcredit) {
1090                 /* give back peer txcredits */
1091                 msg->msg_peertxcredit = 0;
1092
1093                 LASSERT((txpeer->lp_txcredits < 0) == !list_empty(&txpeer->lp_txq));
1094
1095                 txpeer->lp_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
1096                 LASSERT (txpeer->lp_txqnob >= 0);
1097
1098                 txpeer->lp_txcredits++;
1099                 if (txpeer->lp_txcredits <= 0) {
1100                         msg2 = list_entry(txpeer->lp_txq.next,
1101                                           lnet_msg_t, msg_list);
1102                         list_del(&msg2->msg_list);
1103
1104                         LASSERT (msg2->msg_txpeer == txpeer);
1105                         LASSERT (msg2->msg_delayed);
1106
1107                         (void) lnet_post_send_locked(msg2, 1);
1108                 }
1109         }
1110
1111         if (txpeer != NULL) {
1112                 msg->msg_txpeer = NULL;
1113                 lnet_peer_decref_locked(txpeer);
1114         }
1115
1116 #ifdef __KERNEL__
1117         if (msg->msg_rtrcredit) {
1118                 /* give back global router credits */
1119                 lnet_rtrbuf_t     *rb;
1120                 lnet_rtrbufpool_t *rbp;
1121
1122                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1123                  * there until it gets one allocated, or aborts the wait
1124                  * itself */
1125                 LASSERT (msg->msg_kiov != NULL);
1126
1127                 rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
1128                 rbp = rb->rb_pool;
1129                 LASSERT (rbp == lnet_msg2bufpool(msg));
1130
1131                 msg->msg_kiov = NULL;
1132                 msg->msg_rtrcredit = 0;
1133
1134                 LASSERT((rbp->rbp_credits < 0) == !list_empty(&rbp->rbp_msgs));
1135                 LASSERT((rbp->rbp_credits > 0) == !list_empty(&rbp->rbp_bufs));
1136
1137                 list_add(&rb->rb_list, &rbp->rbp_bufs);
1138                 rbp->rbp_credits++;
1139                 if (rbp->rbp_credits <= 0) {
1140                         msg2 = list_entry(rbp->rbp_msgs.next,
1141                                           lnet_msg_t, msg_list);
1142                         list_del(&msg2->msg_list);
1143
1144                         (void) lnet_post_routed_recv_locked(msg2, 1);
1145                 }
1146         }
1147
1148         if (msg->msg_peerrtrcredit) {
1149                 /* give back peer router credits */
1150                 msg->msg_peerrtrcredit = 0;
1151
1152                 LASSERT((rxpeer->lp_rtrcredits < 0) == !list_empty(&rxpeer->lp_rtrq));
1153
1154                 rxpeer->lp_rtrcredits++;
1155                 if (rxpeer->lp_rtrcredits <= 0) {
1156                         msg2 = list_entry(rxpeer->lp_rtrq.next,
1157                                           lnet_msg_t, msg_list);
1158                         list_del(&msg2->msg_list);
1159
1160                         (void) lnet_post_routed_recv_locked(msg2, 1);
1161                 }
1162         }
1163 #else
1164         LASSERT (!msg->msg_rtrcredit);
1165         LASSERT (!msg->msg_peerrtrcredit);
1166 #endif
1167         if (rxpeer != NULL) {
1168                 msg->msg_rxpeer = NULL;
1169                 lnet_peer_decref_locked(rxpeer);
1170         }
1171 }
1172
1173 int
1174 lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg)
1175 {
1176         lnet_nid_t        dst_nid = msg->msg_target.nid;
1177         lnet_ni_t        *src_ni;
1178         lnet_ni_t        *local_ni;
1179         lnet_remotenet_t *rnet;
1180         lnet_route_t     *route;
1181         lnet_route_t     *best_route;
1182         struct list_head *tmp;
1183         lnet_peer_t      *lp;
1184         lnet_peer_t      *lp2;
1185         int               rc;
1186
1187         LASSERT (msg->msg_txpeer == NULL);
1188         LASSERT (!msg->msg_sending);
1189         LASSERT (!msg->msg_target_is_router);
1190         LASSERT (!msg->msg_receiving);
1191
1192         msg->msg_sending = 1;
1193
1194         /* NB! ni != NULL == interface pre-determined (ACK/REPLY) */
1195
1196         LNET_LOCK();
1197
1198         if (the_lnet.ln_shutdown) {
1199                 LNET_UNLOCK();
1200                 return -ESHUTDOWN;
1201         }
1202
1203         if (src_nid == LNET_NID_ANY) {
1204                 src_ni = NULL;
1205         } else {
1206                 src_ni = lnet_nid2ni_locked(src_nid);
1207                 if (src_ni == NULL) {
1208                         LNET_UNLOCK();
1209                         CERROR("Can't send to %s: src %s is not a local nid\n",
1210                                libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1211                         return -EINVAL;
1212                 }
1213                 LASSERT (!msg->msg_routing);
1214         }
1215
1216         /* Is this for someone on a local network? */
1217         local_ni = lnet_net2ni_locked(LNET_NIDNET(dst_nid));
1218
1219         if (local_ni != NULL) {
1220                 if (src_ni == NULL) {
1221                         src_ni = local_ni;
1222                         src_nid = src_ni->ni_nid;
1223                 } else if (src_ni == local_ni) {
1224                         lnet_ni_decref_locked(local_ni);
1225                 } else {
1226                         lnet_ni_decref_locked(local_ni);
1227                         lnet_ni_decref_locked(src_ni);
1228                         LNET_UNLOCK();
1229                         CERROR("no route to %s via from %s\n",
1230                                libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
1231                         return -EINVAL;
1232                 }
1233
1234                 LASSERT (src_nid != LNET_NID_ANY);
1235
1236                 if (!msg->msg_routing) {
1237                         src_nid = lnet_ptlcompat_srcnid(src_nid, dst_nid);
1238                         msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1239                 }
1240
1241                 if (src_ni == the_lnet.ln_loni) {
1242                         /* No send credit hassles with LOLND */
1243                         LNET_UNLOCK();
1244                         lnet_ni_send(src_ni, msg);
1245                         lnet_ni_decref(src_ni);
1246                         return 0;
1247                 }
1248
1249                 rc = lnet_nid2peer_locked(&lp, dst_nid);
1250                 lnet_ni_decref_locked(src_ni);  /* lp has ref on src_ni; lose mine */
1251                 if (rc != 0) {
1252                         LNET_UNLOCK();
1253                         CERROR("Error %d finding peer %s\n", rc,
1254                                libcfs_nid2str(dst_nid));
1255                         /* ENOMEM or shutting down */
1256                         return rc;
1257                 }
1258                 LASSERT (lp->lp_ni == src_ni);
1259         } else {
1260                 /* sending to a remote network */
1261                 rnet = lnet_find_net_locked(LNET_NIDNET(dst_nid));
1262                 if (rnet == NULL) {
1263                         if (src_ni != NULL)
1264                                 lnet_ni_decref_locked(src_ni);
1265                         LNET_UNLOCK();
1266                         CERROR("No route to %s\n", libcfs_id2str(msg->msg_target));
1267                         return -EHOSTUNREACH;
1268                 }
1269
1270                 /* Find the best gateway I can use */
1271                 lp = NULL;
1272                 best_route = NULL;
1273                 list_for_each(tmp, &rnet->lrn_routes) {
1274                         route = list_entry(tmp, lnet_route_t, lr_list);
1275                         lp2 = route->lr_gateway;
1276
1277                         if (lp2->lp_alive &&
1278                             (src_ni == NULL || lp2->lp_ni == src_ni) &&
1279                             (lp == NULL || lnet_compare_routers(lp2, lp) > 0)) {
1280                                 best_route = route;
1281                                 lp = lp2;
1282                         }
1283                 }
1284
1285                 if (lp == NULL) {
1286                         if (src_ni != NULL)
1287                                 lnet_ni_decref_locked(src_ni);
1288                         LNET_UNLOCK();
1289                         CERROR("No route to %s (all routers down)\n",
1290                                libcfs_id2str(msg->msg_target));
1291                         return -EHOSTUNREACH;
1292                 }
1293
1294                 /* Place selected route at the end of the route list to ensure
1295                  * fairness; everything else being equal... */
1296                 list_del(&best_route->lr_list);
1297                 list_add_tail(&best_route->lr_list, &rnet->lrn_routes);
1298
1299                 if (src_ni == NULL) {
1300                         src_ni = lp->lp_ni;
1301                         src_nid = src_ni->ni_nid;
1302                 } else {
1303                         LASSERT (src_ni == lp->lp_ni);
1304                         lnet_ni_decref_locked(src_ni);
1305                 }
1306
1307                 lnet_peer_addref_locked(lp);
1308
1309                 LASSERT (src_nid != LNET_NID_ANY);
1310
1311                 if (!msg->msg_routing) {
1312                         /* I'm the source and now I know which NI to send on */
1313                         src_nid = lnet_ptlcompat_srcnid(src_nid, dst_nid);
1314                         msg->msg_hdr.src_nid = cpu_to_le64(src_nid);
1315                 }
1316
1317                 msg->msg_target_is_router = 1;
1318                 msg->msg_target.nid = lp->lp_nid;
1319                 msg->msg_target.pid = LUSTRE_SRV_LNET_PID;
1320         }
1321
1322         /* 'lp' is our best choice of peer */
1323
1324         LASSERT (!msg->msg_peertxcredit);
1325         LASSERT (!msg->msg_txcredit);
1326         LASSERT (msg->msg_txpeer == NULL);
1327
1328         msg->msg_txpeer = lp;                   /* msg takes my ref on lp */
1329
1330         rc = lnet_post_send_locked(msg, 0);
1331         LNET_UNLOCK();
1332
1333         if (rc == 0)
1334                 lnet_ni_send(src_ni, msg);
1335
1336         return 0;
1337 }
1338
1339 static void
1340 lnet_commit_md (lnet_libmd_t *md, lnet_msg_t *msg)
1341 {
1342         /* ALWAYS called holding the LNET_LOCK */
1343         /* Here, we commit the MD to a network OP by marking it busy and
1344          * decrementing its threshold.  Come what may, the network "owns"
1345          * the MD until a call to lnet_finalize() signals completion. */
1346         LASSERT (!msg->msg_routing);
1347
1348         msg->msg_md = md;
1349
1350         md->md_refcount++;
1351         if (md->md_threshold != LNET_MD_THRESH_INF) {
1352                 LASSERT (md->md_threshold > 0);
1353                 md->md_threshold--;
1354         }
1355
1356         the_lnet.ln_counters.msgs_alloc++;
1357         if (the_lnet.ln_counters.msgs_alloc >
1358             the_lnet.ln_counters.msgs_max)
1359                 the_lnet.ln_counters.msgs_max =
1360                         the_lnet.ln_counters.msgs_alloc;
1361
1362         LASSERT (!msg->msg_onactivelist);
1363         msg->msg_onactivelist = 1;
1364         list_add (&msg->msg_activelist, &the_lnet.ln_active_msgs);
1365 }
1366
1367 static void
1368 lnet_drop_message (lnet_ni_t *ni, void *private, unsigned int nob)
1369 {
1370         LNET_LOCK();
1371         the_lnet.ln_counters.drop_count++;
1372         the_lnet.ln_counters.drop_length += nob;
1373         LNET_UNLOCK();
1374
1375         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
1376 }
1377
1378 static void
1379 lnet_drop_delayed_put(lnet_msg_t *msg, char *reason)
1380 {
1381         LASSERT (msg->msg_md == NULL);
1382         LASSERT (msg->msg_delayed);
1383         LASSERT (msg->msg_rxpeer != NULL);
1384         LASSERT (msg->msg_hdr.type == LNET_MSG_PUT);
1385
1386         CWARN("Dropping delayed PUT from %s portal %d match "LPU64
1387               " offset %d length %d: %s\n",
1388               libcfs_id2str((lnet_process_id_t){
1389                       .nid = msg->msg_hdr.src_nid,
1390                       .pid = msg->msg_hdr.src_pid}),
1391               msg->msg_hdr.msg.put.ptl_index,
1392               msg->msg_hdr.msg.put.match_bits,
1393               msg->msg_hdr.msg.put.offset,
1394               msg->msg_hdr.payload_length,
1395               reason);
1396
1397         /* NB I can't drop msg's ref on msg_rxpeer until after I've
1398          * called lnet_drop_message(), so I just hang onto msg as well
1399          * until that's done */
1400
1401         lnet_drop_message(msg->msg_rxpeer->lp_ni,
1402                           msg->msg_private, msg->msg_len);
1403
1404         LNET_LOCK();
1405
1406         lnet_peer_decref_locked(msg->msg_rxpeer);
1407         msg->msg_rxpeer = NULL;
1408
1409         lnet_msg_free(msg);
1410
1411         LNET_UNLOCK();
1412 }
1413
1414 int
1415 LNetSetLazyPortal(int portal)
1416 {
1417         lnet_portal_t *ptl = &the_lnet.ln_portals[portal];
1418
1419         if (portal < 0 || portal >= the_lnet.ln_nportals)
1420                 return -EINVAL;
1421
1422         CDEBUG(D_NET, "Setting portal %d lazy\n", portal);
1423
1424         LNET_LOCK();
1425
1426         ptl->ptl_options |= LNET_PTL_LAZY;
1427
1428         LNET_UNLOCK();
1429
1430         return 0;
1431 }
1432
1433 int
1434 LNetClearLazyPortal(int portal)
1435 {
1436         struct list_head  zombies;
1437         lnet_portal_t    *ptl = &the_lnet.ln_portals[portal];
1438         lnet_msg_t       *msg;
1439
1440         if (portal < 0 || portal >= the_lnet.ln_nportals)
1441                 return -EINVAL;
1442
1443         LNET_LOCK();
1444
1445         if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1446                 LNET_UNLOCK();
1447                 return 0;
1448         }
1449
1450         if (the_lnet.ln_shutdown)
1451                 CWARN ("Active lazy portal %d on exit\n", portal);
1452         else
1453                 CDEBUG (D_NET, "clearing portal %d lazy\n", portal);
1454
1455         /* grab all the blocked messages atomically */
1456         list_add(&zombies, &ptl->ptl_msgq);
1457         list_del_init(&ptl->ptl_msgq);
1458
1459         ptl->ptl_msgq_version++;
1460         ptl->ptl_options &= ~LNET_PTL_LAZY;
1461
1462         LNET_UNLOCK();
1463
1464         while (!list_empty(&zombies)) {
1465                 msg = list_entry(zombies.next, lnet_msg_t, msg_list);
1466                 list_del(&msg->msg_list);
1467
1468                 lnet_drop_delayed_put(msg, "Clearing lazy portal attr");
1469         }
1470
1471         return 0;
1472 }
1473
1474 static void
1475 lnet_recv_put(lnet_libmd_t *md, lnet_msg_t *msg, int delayed,
1476               unsigned int offset, unsigned int mlength)
1477 {
1478         lnet_hdr_t       *hdr = &msg->msg_hdr;
1479
1480         LNET_LOCK();
1481
1482         the_lnet.ln_counters.recv_count++;
1483         the_lnet.ln_counters.recv_length += mlength;
1484
1485         LNET_UNLOCK();
1486
1487         if (mlength != 0)
1488                 lnet_setpayloadbuffer(msg);
1489
1490         msg->msg_ev.type       = LNET_EVENT_PUT;
1491         msg->msg_ev.target.pid = hdr->dest_pid;
1492         msg->msg_ev.target.nid = hdr->dest_nid;
1493         msg->msg_ev.hdr_data   = hdr->msg.put.hdr_data;
1494
1495         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
1496          * it back into the ACK during lnet_finalize() */
1497         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
1498                         (md->md_options & LNET_MD_ACK_DISABLE) == 0);
1499
1500         lnet_ni_recv(msg->msg_rxpeer->lp_ni,
1501                      msg->msg_private,
1502                      msg, delayed, offset, mlength,
1503                      hdr->payload_length);
1504 }
1505
1506 /* called with LNET_LOCK held */
1507 void
1508 lnet_match_blocked_msg(lnet_libmd_t *md)
1509 {
1510         CFS_LIST_HEAD    (drops);
1511         CFS_LIST_HEAD    (matches);
1512         struct list_head *tmp;
1513         struct list_head *entry;
1514         lnet_msg_t       *msg;
1515         lnet_me_t        *me  = md->md_me;
1516         lnet_portal_t    *ptl = &the_lnet.ln_portals[me->me_portal];
1517
1518         LASSERT (me->me_portal < the_lnet.ln_nportals);
1519
1520         if ((ptl->ptl_options & LNET_PTL_LAZY) == 0) {
1521                 LASSERT (list_empty(&ptl->ptl_msgq));
1522                 return;
1523         }
1524
1525         LASSERT (md->md_refcount == 0); /* a brand new MD */
1526
1527         list_for_each_safe (entry, tmp, &ptl->ptl_msgq) {
1528                 int               rc;
1529                 int               index;
1530                 unsigned int      mlength;
1531                 unsigned int      offset;
1532                 lnet_hdr_t       *hdr;
1533                 lnet_process_id_t src;
1534
1535                 msg = list_entry(entry, lnet_msg_t, msg_list);
1536
1537                 LASSERT (msg->msg_delayed);
1538
1539                 hdr   = &msg->msg_hdr;
1540                 index = hdr->msg.put.ptl_index;
1541
1542                 src.nid = hdr->src_nid;
1543                 src.pid = hdr->src_pid;
1544
1545                 rc = lnet_try_match_md(index, LNET_MD_OP_PUT, src,
1546                                        hdr->payload_length,
1547                                        hdr->msg.put.offset,
1548                                        hdr->msg.put.match_bits,
1549                                        md, msg, &mlength, &offset);
1550
1551                 if (rc == LNET_MATCHMD_NONE)
1552                         continue;
1553
1554                 /* Hurrah! This _is_ a match */
1555                 list_del(&msg->msg_list);
1556                 ptl->ptl_msgq_version++;
1557
1558                 if (rc == LNET_MATCHMD_OK) {
1559                         list_add_tail(&msg->msg_list, &matches);
1560
1561                         CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
1562                                "match "LPU64" offset %d length %d.\n",
1563                                libcfs_id2str(src),
1564                                hdr->msg.put.ptl_index,
1565                                hdr->msg.put.match_bits,
1566                                hdr->msg.put.offset,
1567                                hdr->payload_length);
1568                 } else {
1569                         LASSERT (rc == LNET_MATCHMD_DROP);
1570
1571                         list_add_tail(&msg->msg_list, &drops);
1572                 }
1573
1574                 if (lnet_md_exhausted(md))
1575                         break;
1576         }
1577
1578         LNET_UNLOCK();
1579
1580         list_for_each_safe (entry, tmp, &drops) {
1581                 msg = list_entry(entry, lnet_msg_t, msg_list);
1582
1583                 list_del(&msg->msg_list);
1584
1585                 lnet_drop_delayed_put(msg, "Bad match");
1586         }
1587
1588         list_for_each_safe (entry, tmp, &matches) {
1589                 msg = list_entry(entry, lnet_msg_t, msg_list);
1590
1591                 list_del(&msg->msg_list);
1592
1593                 /* md won't disappear under me, since each msg
1594                  * holds a ref on it */
1595                 lnet_recv_put(md, msg, 1,
1596                               msg->msg_ev.offset,
1597                               msg->msg_ev.mlength);
1598         }
1599
1600         LNET_LOCK();
1601 }
1602
1603 static int
1604 lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
1605 {
1606         int               rc;
1607         int               index;
1608         lnet_hdr_t       *hdr = &msg->msg_hdr;
1609         unsigned int      rlength = hdr->payload_length;
1610         unsigned int      mlength = 0;
1611         unsigned int      offset = 0;
1612         lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
1613                                  /* .pid = */ hdr->src_pid};
1614         lnet_libmd_t     *md;
1615
1616         /* Convert put fields to host byte order */
1617         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
1618         hdr->msg.put.ptl_index = le32_to_cpu(hdr->msg.put.ptl_index);
1619         hdr->msg.put.offset = le32_to_cpu(hdr->msg.put.offset);
1620
1621         index = hdr->msg.put.ptl_index;
1622
1623         LNET_LOCK();
1624
1625         rc = lnet_match_md(index, LNET_MD_OP_PUT, src,
1626                            rlength, hdr->msg.put.offset,
1627                            hdr->msg.put.match_bits, msg,
1628                            &mlength, &offset, &md);
1629         switch (rc) {
1630         default:
1631                 LBUG();
1632
1633         case LNET_MATCHMD_OK:
1634                 LNET_UNLOCK();
1635                 lnet_recv_put(md, msg, 0, offset, mlength);
1636                 return 0;
1637
1638         case LNET_MATCHMD_NONE:
1639                 rc = lnet_eager_recv_locked(msg);
1640                 if (rc == 0 && !the_lnet.ln_shutdown) {
1641                         list_add_tail(&msg->msg_list,
1642                                       &the_lnet.ln_portals[index].ptl_msgq);
1643
1644                         the_lnet.ln_portals[index].ptl_msgq_version++;
1645
1646                         CDEBUG(D_NET, "Delaying PUT from %s portal %d match "
1647                                LPU64" offset %d length %d: no match \n",
1648                                libcfs_id2str(src), index,
1649                                hdr->msg.put.match_bits,
1650                                hdr->msg.put.offset, rlength);
1651
1652                         LNET_UNLOCK();
1653                         return 0;
1654                 }
1655                 /* fall through */
1656
1657         case LNET_MATCHMD_DROP:
1658                 CDEBUG(D_NETERROR,
1659                        "Dropping PUT from %s portal %d match "LPU64
1660                        " offset %d length %d: %d\n",
1661                        libcfs_id2str(src), index,
1662                        hdr->msg.put.match_bits,
1663                        hdr->msg.put.offset, rlength, rc);
1664                 LNET_UNLOCK();
1665
1666                 return ENOENT;          /* +ve: OK but no match */
1667         }
1668 }
1669
1670 static int
1671 lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
1672 {
1673         lnet_hdr_t        *hdr = &msg->msg_hdr;
1674         unsigned int       mlength = 0;
1675         unsigned int       offset = 0;
1676         lnet_process_id_t  src = {/* .nid = */ hdr->src_nid,
1677                                   /* .pid = */ hdr->src_pid};
1678         lnet_handle_wire_t reply_wmd;
1679         lnet_libmd_t      *md;
1680         int                rc;
1681
1682         /* Convert get fields to host byte order */
1683         hdr->msg.get.match_bits = le64_to_cpu(hdr->msg.get.match_bits);
1684         hdr->msg.get.ptl_index = le32_to_cpu(hdr->msg.get.ptl_index);
1685         hdr->msg.get.sink_length = le32_to_cpu(hdr->msg.get.sink_length);
1686         hdr->msg.get.src_offset = le32_to_cpu(hdr->msg.get.src_offset);
1687
1688         LNET_LOCK();
1689
1690         rc = lnet_match_md(hdr->msg.get.ptl_index, LNET_MD_OP_GET, src,
1691                            hdr->msg.get.sink_length, hdr->msg.get.src_offset,
1692                            hdr->msg.get.match_bits, msg,
1693                            &mlength, &offset, &md);
1694         if (rc == LNET_MATCHMD_DROP) {
1695                 CDEBUG(D_NETERROR,
1696                        "Dropping GET from %s portal %d match "LPU64
1697                        " offset %d length %d\n",
1698                        libcfs_id2str(src),
1699                        hdr->msg.get.ptl_index,
1700                        hdr->msg.get.match_bits,
1701                        hdr->msg.get.src_offset,
1702                        hdr->msg.get.sink_length);
1703                 LNET_UNLOCK();
1704                 return ENOENT;                  /* +ve: OK but no match */
1705         }
1706
1707         LASSERT (rc == LNET_MATCHMD_OK);
1708
1709         the_lnet.ln_counters.send_count++;
1710         the_lnet.ln_counters.send_length += mlength;
1711
1712         LNET_UNLOCK();
1713
1714         reply_wmd = hdr->msg.get.return_wmd;
1715
1716         lnet_prep_send(msg, LNET_MSG_REPLY, src, offset, mlength);
1717
1718         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
1719
1720         msg->msg_ev.type = LNET_EVENT_GET;
1721         msg->msg_ev.target.pid = hdr->dest_pid;
1722         msg->msg_ev.target.nid = hdr->dest_nid;
1723         msg->msg_ev.hdr_data = 0;
1724
1725         if (rdma_get) {
1726                 /* The LND completes the REPLY from her recv procedure */
1727                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
1728                              msg->msg_offset, msg->msg_len, msg->msg_len);
1729                 return 0;
1730         }
1731
1732         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
1733         msg->msg_receiving = 0;
1734
1735         rc = lnet_send(ni->ni_nid, msg);
1736         if (rc < 0) {
1737                 /* didn't get as far as lnet_ni_send() */
1738                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
1739                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src), rc);
1740
1741                 lnet_finalize(ni, msg, rc);
1742         }
1743
1744         return 0;
1745 }
1746
1747 static int
1748 lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
1749 {
1750         void             *private = msg->msg_private;
1751         lnet_hdr_t       *hdr = &msg->msg_hdr;
1752         lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
1753                                  /* .pid = */ hdr->src_pid};
1754         lnet_libmd_t     *md;
1755         int               rlength;
1756         int               mlength;
1757
1758         LNET_LOCK();
1759
1760         /* NB handles only looked up by creator (no flips) */
1761         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
1762         if (md == NULL || md->md_threshold == 0) {
1763                 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s for %s "
1764                        "MD "LPX64"."LPX64"\n",
1765                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1766                        (md == NULL) ? "invalid" : "inactive",
1767                        hdr->msg.reply.dst_wmd.wh_interface_cookie,
1768                        hdr->msg.reply.dst_wmd.wh_object_cookie);
1769
1770                 LNET_UNLOCK();
1771                 return ENOENT;                  /* +ve: OK but no match */
1772         }
1773
1774         LASSERT (md->md_offset == 0);
1775
1776         rlength = hdr->payload_length;
1777         mlength = MIN(rlength, md->md_length);
1778
1779         if (mlength < rlength &&
1780             (md->md_options & LNET_MD_TRUNCATE) == 0) {
1781                 CDEBUG(D_NETERROR, "%s: Dropping REPLY from %s length %d "
1782                        "for MD "LPX64" would overflow (%d)\n",
1783                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1784                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
1785                         mlength);
1786                 LNET_UNLOCK();
1787                 return ENOENT;          /* +ve: OK but no match */
1788         }
1789
1790         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md "LPX64"\n",
1791                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1792                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
1793
1794         lnet_commit_md(md, msg);
1795
1796         if (mlength != 0)
1797                 lnet_setpayloadbuffer(msg);
1798
1799         msg->msg_ev.type = LNET_EVENT_REPLY;
1800         msg->msg_ev.target.pid = hdr->dest_pid;
1801         msg->msg_ev.target.nid = hdr->dest_nid;
1802         msg->msg_ev.initiator = src;
1803         msg->msg_ev.rlength = rlength;
1804         msg->msg_ev.mlength = mlength;
1805         msg->msg_ev.offset = 0;
1806
1807         lnet_md_deconstruct(md, &msg->msg_ev.md);
1808         lnet_md2handle(&msg->msg_ev.md_handle, md);
1809
1810         the_lnet.ln_counters.recv_count++;
1811         the_lnet.ln_counters.recv_length += mlength;
1812
1813         LNET_UNLOCK();
1814
1815         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
1816         return 0;
1817 }
1818
1819 static int
1820 lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
1821 {
1822         lnet_hdr_t       *hdr = &msg->msg_hdr;
1823         lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
1824                                  /* .pid = */ hdr->src_pid};
1825         lnet_libmd_t    *md;
1826
1827         /* Convert ack fields to host byte order */
1828         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
1829         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
1830
1831         LNET_LOCK();
1832
1833         /* NB handles only looked up by creator (no flips) */
1834         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
1835         if (md == NULL || md->md_threshold == 0) {
1836                 /* Don't moan; this is expected */
1837                 CDEBUG(D_NET,
1838                        "%s: Dropping ACK from %s to %s MD "LPX64"."LPX64"\n",
1839                        libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1840                        (md == NULL) ? "invalid" : "inactive",
1841                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
1842                        hdr->msg.ack.dst_wmd.wh_object_cookie);
1843                 LNET_UNLOCK();
1844                 return ENOENT;                  /* +ve! */
1845         }
1846
1847         CDEBUG(D_NET, "%s: ACK from %s into md "LPX64"\n",
1848                libcfs_nid2str(ni->ni_nid), libcfs_id2str(src),
1849                hdr->msg.ack.dst_wmd.wh_object_cookie);
1850
1851         lnet_commit_md(md, msg);
1852
1853         msg->msg_ev.type = LNET_EVENT_ACK;
1854         msg->msg_ev.target.pid = hdr->dest_pid;
1855         msg->msg_ev.target.nid = hdr->dest_nid;
1856         msg->msg_ev.initiator = src;
1857         msg->msg_ev.mlength = hdr->msg.ack.mlength;
1858         msg->msg_ev.match_bits = hdr->msg.ack.match_bits;
1859
1860         lnet_md_deconstruct(md, &msg->msg_ev.md);
1861         lnet_md2handle(&msg->msg_ev.md_handle, md);
1862
1863         the_lnet.ln_counters.recv_count++;
1864
1865         LNET_UNLOCK();
1866
1867         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
1868         return 0;
1869 }
1870
1871 char *
1872 lnet_msgtyp2str (int type)
1873 {
1874         switch (type) {
1875         case LNET_MSG_ACK:
1876                 return ("ACK");
1877         case LNET_MSG_PUT:
1878                 return ("PUT");
1879         case LNET_MSG_GET:
1880                 return ("GET");
1881         case LNET_MSG_REPLY:
1882                 return ("REPLY");
1883         case LNET_MSG_HELLO:
1884                 return ("HELLO");
1885         default:
1886                 return ("<UNKNOWN>");
1887         }
1888 }
1889
1890 void
1891 lnet_print_hdr(lnet_hdr_t * hdr)
1892 {
1893         lnet_process_id_t src = {/* .nid = */ hdr->src_nid,
1894                                  /* .pid = */ hdr->src_pid};
1895         lnet_process_id_t dst = {/* .nid = */ hdr->dest_nid,
1896                                  /* .pid = */ hdr->dest_pid};
1897         char *type_str = lnet_msgtyp2str (hdr->type);
1898
1899         CWARN("P3 Header at %p of type %s\n", hdr, type_str);
1900         CWARN("    From %s\n", libcfs_id2str(src));
1901         CWARN("    To   %s\n", libcfs_id2str(dst));
1902
1903         switch (hdr->type) {
1904         default:
1905                 break;
1906
1907         case LNET_MSG_PUT:
1908                 CWARN("    Ptl index %d, ack md "LPX64"."LPX64", "
1909                       "match bits "LPU64"\n",
1910                       hdr->msg.put.ptl_index,
1911                       hdr->msg.put.ack_wmd.wh_interface_cookie,
1912                       hdr->msg.put.ack_wmd.wh_object_cookie,
1913                       hdr->msg.put.match_bits);
1914                 CWARN("    Length %d, offset %d, hdr data "LPX64"\n",
1915                       hdr->payload_length, hdr->msg.put.offset,
1916                       hdr->msg.put.hdr_data);
1917                 break;
1918
1919         case LNET_MSG_GET:
1920                 CWARN("    Ptl index %d, return md "LPX64"."LPX64", "
1921                       "match bits "LPU64"\n", hdr->msg.get.ptl_index,
1922                       hdr->msg.get.return_wmd.wh_interface_cookie,
1923                       hdr->msg.get.return_wmd.wh_object_cookie,
1924                       hdr->msg.get.match_bits);
1925                 CWARN("    Length %d, src offset %d\n",
1926                       hdr->msg.get.sink_length,
1927                       hdr->msg.get.src_offset);
1928                 break;
1929
1930         case LNET_MSG_ACK:
1931                 CWARN("    dst md "LPX64"."LPX64", "
1932                       "manipulated length %d\n",
1933                       hdr->msg.ack.dst_wmd.wh_interface_cookie,
1934                       hdr->msg.ack.dst_wmd.wh_object_cookie,
1935                       hdr->msg.ack.mlength);
1936                 break;
1937
1938         case LNET_MSG_REPLY:
1939                 CWARN("    dst md "LPX64"."LPX64", "
1940                       "length %d\n",
1941                       hdr->msg.reply.dst_wmd.wh_interface_cookie,
1942                       hdr->msg.reply.dst_wmd.wh_object_cookie,
1943                       hdr->payload_length);
1944         }
1945
1946 }
1947
1948
1949 int
1950 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
1951            void *private, int rdma_req)
1952 {
1953         int            rc = 0;
1954         int            for_me;
1955         lnet_msg_t    *msg;
1956         lnet_nid_t     dest_nid;
1957         lnet_nid_t     src_nid;
1958         __u32          payload_length;
1959         __u32          type;
1960
1961         LASSERT (!in_interrupt ());
1962
1963         type = le32_to_cpu(hdr->type);
1964         src_nid = le64_to_cpu(hdr->src_nid);
1965         dest_nid = le64_to_cpu(hdr->dest_nid);
1966         payload_length = le32_to_cpu(hdr->payload_length);
1967
1968         for_me = lnet_ptlcompat_matchnid(ni->ni_nid, dest_nid);
1969
1970         switch (type) {
1971         case LNET_MSG_ACK:
1972         case LNET_MSG_GET:
1973                 if (payload_length > 0) {
1974                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
1975                                libcfs_nid2str(from_nid),
1976                                libcfs_nid2str(src_nid),
1977                                lnet_msgtyp2str(type), payload_length);
1978                         return -EPROTO;
1979                 }
1980                 break;
1981
1982         case LNET_MSG_PUT:
1983         case LNET_MSG_REPLY:
1984                 if (payload_length > (for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
1985                         CERROR("%s, src %s: bad %s payload %d "
1986                                "(%d max expected)\n",
1987                                libcfs_nid2str(from_nid),
1988                                libcfs_nid2str(src_nid),
1989                                lnet_msgtyp2str(type),
1990                                payload_length,
1991                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
1992                         return -EPROTO;
1993                 }
1994                 break;
1995
1996         default:
1997                 CERROR("%s, src %s: Bad message type 0x%x\n",
1998                        libcfs_nid2str(from_nid),
1999                        libcfs_nid2str(src_nid), type);
2000                 return -EPROTO;
2001         }
2002
2003         /* Regard a bad destination NID as a protocol error.  Senders should
2004          * know what they're doing; if they don't they're misconfigured, buggy
2005          * or malicious so we chop them off at the knees :) */
2006
2007         if (!for_me) {
2008                 if (the_lnet.ln_ptlcompat > 0) {
2009                         /* portals compatibility is single-network */
2010                         CERROR ("%s, src %s: Bad dest nid %s "
2011                                 "(routing not supported)\n",
2012                                 libcfs_nid2str(from_nid),
2013                                 libcfs_nid2str(src_nid),
2014                                 libcfs_nid2str(dest_nid));
2015                         return -EPROTO;
2016                 }
2017
2018                 if (the_lnet.ln_ptlcompat == 0 &&
2019                     LNET_NIDNET(dest_nid) == LNET_NIDNET(ni->ni_nid)) {
2020                         /* should have gone direct */
2021                         CERROR ("%s, src %s: Bad dest nid %s "
2022                                 "(should have been sent direct)\n",
2023                                 libcfs_nid2str(from_nid),
2024                                 libcfs_nid2str(src_nid),
2025                                 libcfs_nid2str(dest_nid));
2026                         return -EPROTO;
2027                 }
2028
2029                 if (the_lnet.ln_ptlcompat == 0 &&
2030                     lnet_islocalnid(dest_nid)) {
2031                         /* dest is another local NI; sender should have used
2032                          * this node's NID on its own network */
2033                         CERROR ("%s, src %s: Bad dest nid %s "
2034                                 "(it's my nid but on a different network)\n",
2035                                 libcfs_nid2str(from_nid),
2036                                 libcfs_nid2str(src_nid),
2037                                 libcfs_nid2str(dest_nid));
2038                         return -EPROTO;
2039                 }
2040
2041                 if (rdma_req && type == LNET_MSG_GET) {
2042                         CERROR ("%s, src %s: Bad optimized GET for %s "
2043                                 "(final destination must be me)\n",
2044                                 libcfs_nid2str(from_nid),
2045                                 libcfs_nid2str(src_nid),
2046                                 libcfs_nid2str(dest_nid));
2047                         return -EPROTO;
2048                 }
2049
2050                 if (!the_lnet.ln_routing) {
2051                         CERROR ("%s, src %s: Dropping message for %s "
2052                                 "(routing not enabled)\n",
2053                                 libcfs_nid2str(from_nid),
2054                                 libcfs_nid2str(src_nid),
2055                                 libcfs_nid2str(dest_nid));
2056                         goto drop;
2057                 }
2058         }
2059
2060         /* Message looks OK; we're not going to return an error, so we MUST
2061          * call back lnd_recv() come what may... */
2062
2063         if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2064             fail_peer (src_nid, 0))             /* shall we now? */
2065         {
2066                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
2067                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2068                        lnet_msgtyp2str(type));
2069                 goto drop;
2070         }
2071
2072         msg = lnet_msg_alloc();
2073         if (msg == NULL) {
2074                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
2075                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2076                        lnet_msgtyp2str(type));
2077                 goto drop;
2078         }
2079
2080         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear, pointers NULL etc */
2081
2082         msg->msg_type = type;
2083         msg->msg_private = private;
2084         msg->msg_receiving = 1;
2085         msg->msg_len = msg->msg_wanted = payload_length;
2086         msg->msg_offset = 0;
2087         msg->msg_hdr = *hdr;
2088
2089         LNET_LOCK();
2090         rc = lnet_nid2peer_locked(&msg->msg_rxpeer, from_nid);
2091         if (rc != 0) {
2092                 LNET_UNLOCK();
2093                 CERROR("%s, src %s: Dropping %s "
2094                        "(error %d looking up sender)\n",
2095                        libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
2096                        lnet_msgtyp2str(type), rc);
2097                 goto free_drop;
2098         }
2099         LNET_UNLOCK();
2100
2101 #ifndef __KERNEL__
2102         LASSERT (for_me);
2103 #else
2104         if (!for_me) {
2105                 msg->msg_target.pid = le32_to_cpu(hdr->dest_pid);
2106                 msg->msg_target.nid = dest_nid;
2107                 msg->msg_routing = 1;
2108                 msg->msg_offset = 0;
2109
2110                 LNET_LOCK();
2111                 if (msg->msg_rxpeer->lp_rtrcredits <= 0 ||
2112                     lnet_msg2bufpool(msg)->rbp_credits <= 0) {
2113                         rc = lnet_eager_recv_locked(msg);
2114                         if (rc != 0) {
2115                                 LNET_UNLOCK();
2116                                 goto free_drop;
2117                         }
2118                 }
2119
2120                 lnet_commit_routedmsg(msg);
2121                 rc = lnet_post_routed_recv_locked(msg, 0);
2122                 LNET_UNLOCK();
2123
2124                 if (rc == 0)
2125                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
2126                                      0, payload_length, payload_length);
2127                 return 0;
2128         }
2129 #endif
2130         /* convert common msg->hdr fields to host byteorder */
2131         msg->msg_hdr.type = type;
2132         msg->msg_hdr.src_nid = src_nid;
2133         msg->msg_hdr.src_pid = le32_to_cpu(msg->msg_hdr.src_pid);
2134         msg->msg_hdr.dest_nid = dest_nid;
2135         msg->msg_hdr.dest_pid = le32_to_cpu(msg->msg_hdr.dest_pid);
2136         msg->msg_hdr.payload_length = payload_length;
2137
2138         msg->msg_ev.sender = from_nid;
2139
2140         switch (type) {
2141         case LNET_MSG_ACK:
2142                 rc = lnet_parse_ack(ni, msg);
2143                 break;
2144         case LNET_MSG_PUT:
2145                 rc = lnet_parse_put(ni, msg);
2146                 break;
2147         case LNET_MSG_GET:
2148                 rc = lnet_parse_get(ni, msg, rdma_req);
2149                 break;
2150         case LNET_MSG_REPLY:
2151                 rc = lnet_parse_reply(ni, msg);
2152                 break;
2153         default:
2154                 LASSERT(0);
2155                 goto free_drop;  /* prevent an unused label if !kernel */
2156         }
2157
2158         if (rc == 0)
2159                 return 0;
2160
2161         LASSERT (rc == ENOENT);
2162
2163  free_drop:
2164         LASSERT (msg->msg_md == NULL);
2165         LNET_LOCK();
2166         if (msg->msg_rxpeer != NULL) {
2167                 lnet_peer_decref_locked(msg->msg_rxpeer);
2168                 msg->msg_rxpeer = NULL;
2169         }
2170         lnet_msg_free(msg);                     /* expects LNET_LOCK held */
2171         LNET_UNLOCK();
2172
2173  drop:
2174         lnet_drop_message(ni, private, payload_length);
2175         return 0;
2176 }
2177
2178 int
2179 LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
2180         lnet_process_id_t target, unsigned int portal,
2181         __u64 match_bits, unsigned int offset,
2182         __u64 hdr_data)
2183 {
2184         lnet_msg_t       *msg;
2185         lnet_libmd_t     *md;
2186         int               rc;
2187
2188         LASSERT (the_lnet.ln_init);
2189         LASSERT (the_lnet.ln_refcount > 0);
2190
2191         if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2192             fail_peer (target.nid, 1))          /* shall we now? */
2193         {
2194                 CERROR("Dropping PUT to %s: simulated failure\n",
2195                        libcfs_id2str(target));
2196                 return -EIO;
2197         }
2198
2199         msg = lnet_msg_alloc();
2200         if (msg == NULL) {
2201                 CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
2202                        libcfs_id2str(target));
2203                 return -ENOMEM;
2204         }
2205
2206         LNET_LOCK();
2207
2208         md = lnet_handle2md(&mdh);
2209         if (md == NULL || md->md_threshold == 0) {
2210                 lnet_msg_free(msg);
2211                 LNET_UNLOCK();
2212
2213                 CERROR("Dropping PUT to %s: MD invalid\n",
2214                        libcfs_id2str(target));
2215                 return -ENOENT;
2216         }
2217
2218         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
2219
2220         lnet_commit_md(md, msg);
2221
2222         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
2223
2224         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
2225         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
2226         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
2227         msg->msg_hdr.msg.put.hdr_data = hdr_data;
2228
2229         /* NB handles only looked up by creator (no flips) */
2230         if (ack == LNET_ACK_REQ) {
2231                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
2232                         the_lnet.ln_interface_cookie;
2233                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
2234                         md->md_lh.lh_cookie;
2235         } else {
2236                 msg->msg_hdr.msg.put.ack_wmd = LNET_WIRE_HANDLE_NONE;
2237         }
2238
2239         msg->msg_ev.type = LNET_EVENT_SEND;
2240         msg->msg_ev.initiator.nid = LNET_NID_ANY;
2241         msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2242         msg->msg_ev.target = target;
2243         msg->msg_ev.sender = LNET_NID_ANY;
2244         msg->msg_ev.pt_index = portal;
2245         msg->msg_ev.match_bits = match_bits;
2246         msg->msg_ev.rlength = md->md_length;
2247         msg->msg_ev.mlength = md->md_length;
2248         msg->msg_ev.offset = offset;
2249         msg->msg_ev.hdr_data = hdr_data;
2250
2251         lnet_md_deconstruct(md, &msg->msg_ev.md);
2252         lnet_md2handle(&msg->msg_ev.md_handle, md);
2253
2254         the_lnet.ln_counters.send_count++;
2255         the_lnet.ln_counters.send_length += md->md_length;
2256
2257         LNET_UNLOCK();
2258
2259         rc = lnet_send(self, msg);
2260         if (rc != 0) {
2261                 CERROR("Error sending PUT to %s: %d\n",
2262                        libcfs_id2str(target), rc);
2263                 lnet_finalize (NULL, msg, rc);
2264         }
2265
2266         /* completion will be signalled by an event */
2267         return 0;
2268 }
2269
2270 lnet_msg_t *
2271 lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
2272 {
2273         /* The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
2274          * returns a msg for the LND to pass to lnet_finalize() when the sink
2275          * data has been received.
2276          *
2277          * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
2278          * lnet_finalize() is called on it, so the LND must call this first */
2279
2280         lnet_msg_t        *msg = lnet_msg_alloc();
2281         lnet_libmd_t      *getmd = getmsg->msg_md;
2282         lnet_process_id_t  peer_id = getmsg->msg_target;
2283
2284         LASSERT (!getmsg->msg_target_is_router);
2285         LASSERT (!getmsg->msg_routing);
2286
2287         LNET_LOCK();
2288
2289         LASSERT (getmd->md_refcount > 0);
2290
2291         if (msg == NULL) {
2292                 CERROR ("%s: Dropping REPLY from %s: can't allocate msg\n",
2293                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id));
2294                 goto drop;
2295         }
2296
2297         if (getmd->md_threshold == 0) {
2298                 CERROR ("%s: Dropping REPLY from %s for inactive MD %p\n",
2299                         libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id),
2300                         getmd);
2301                 goto drop_msg;
2302         }
2303
2304         LASSERT (getmd->md_offset == 0);
2305
2306         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
2307                libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
2308
2309         lnet_commit_md (getmd, msg);
2310
2311         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
2312
2313         msg->msg_ev.type = LNET_EVENT_REPLY;
2314         msg->msg_ev.initiator = peer_id;
2315         msg->msg_ev.sender = peer_id.nid;  /* optimized GETs can't be routed */
2316         msg->msg_ev.rlength = msg->msg_ev.mlength = getmd->md_length;
2317         msg->msg_ev.offset = 0;
2318
2319         lnet_md_deconstruct(getmd, &msg->msg_ev.md);
2320         lnet_md2handle(&msg->msg_ev.md_handle, getmd);
2321
2322         the_lnet.ln_counters.recv_count++;
2323         the_lnet.ln_counters.recv_length += getmd->md_length;
2324
2325         LNET_UNLOCK();
2326
2327         return msg;
2328
2329  drop_msg:
2330         lnet_msg_free(msg);
2331  drop:
2332         the_lnet.ln_counters.drop_count++;
2333         the_lnet.ln_counters.drop_length += getmd->md_length;
2334
2335         LNET_UNLOCK ();
2336
2337         return NULL;
2338 }
2339
2340 void
2341 lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
2342 {
2343         /* Set the REPLY length, now the RDMA that elides the REPLY message has
2344          * completed and I know it. */
2345         LASSERT (reply != NULL);
2346         LASSERT (reply->msg_type == LNET_MSG_GET);
2347         LASSERT (reply->msg_ev.type == LNET_EVENT_REPLY);
2348
2349         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
2350          * the end of my buffer, I might as well be dead. */
2351         LASSERT (len <= reply->msg_ev.mlength);
2352
2353         reply->msg_ev.mlength = len;
2354 }
2355
2356 int
2357 LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
2358         lnet_process_id_t target, unsigned int portal,
2359         __u64 match_bits, unsigned int offset)
2360 {
2361         lnet_msg_t       *msg;
2362         lnet_libmd_t     *md;
2363         int               rc;
2364
2365         LASSERT (the_lnet.ln_init);
2366         LASSERT (the_lnet.ln_refcount > 0);
2367
2368         if (!list_empty (&the_lnet.ln_test_peers) && /* normally we don't */
2369             fail_peer (target.nid, 1))          /* shall we now? */
2370         {
2371                 CERROR("Dropping GET to %s: simulated failure\n",
2372                        libcfs_id2str(target));
2373                 return -EIO;
2374         }
2375
2376         msg = lnet_msg_alloc();
2377         if (msg == NULL) {
2378                 CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
2379                        libcfs_id2str(target));
2380                 return -ENOMEM;
2381         }
2382
2383         LNET_LOCK();
2384
2385         md = lnet_handle2md(&mdh);
2386         if (md == NULL || md->md_threshold == 0) {
2387                 lnet_msg_free(msg);
2388                 LNET_UNLOCK();
2389
2390                 CERROR("Dropping GET to %s: MD invalid\n",
2391                        libcfs_id2str(target));
2392                 return -ENOENT;
2393         }
2394
2395         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
2396
2397         lnet_commit_md(md, msg);
2398
2399         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
2400
2401         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
2402         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
2403         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
2404         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
2405
2406         /* NB handles only looked up by creator (no flips) */
2407         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
2408                 the_lnet.ln_interface_cookie;
2409         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
2410                 md->md_lh.lh_cookie;
2411
2412         msg->msg_ev.type = LNET_EVENT_SEND;
2413         msg->msg_ev.initiator.nid = LNET_NID_ANY;
2414         msg->msg_ev.initiator.pid = the_lnet.ln_pid;
2415         msg->msg_ev.target = target;
2416         msg->msg_ev.sender = LNET_NID_ANY;
2417         msg->msg_ev.pt_index = portal;
2418         msg->msg_ev.match_bits = match_bits;
2419         msg->msg_ev.rlength = md->md_length;
2420         msg->msg_ev.mlength = md->md_length;
2421         msg->msg_ev.offset = offset;
2422         msg->msg_ev.hdr_data = 0;
2423
2424         lnet_md_deconstruct(md, &msg->msg_ev.md);
2425         lnet_md2handle(&msg->msg_ev.md_handle, md);
2426
2427         the_lnet.ln_counters.send_count++;
2428
2429         LNET_UNLOCK();
2430
2431         rc = lnet_send(self, msg);
2432         if (rc < 0) {
2433                 CERROR("error sending GET to %s: %d\n",
2434                        libcfs_id2str(target), rc);
2435                 lnet_finalize (NULL, msg, rc);
2436         }
2437
2438         /* completion will be signalled by an event */
2439         return 0;
2440 }
2441
2442 int
2443 LNetDist (lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
2444 {
2445         struct list_head *e;
2446         lnet_ni_t        *ni;
2447         lnet_route_t     *route;
2448         lnet_remotenet_t *rnet;
2449         __u32             dstnet = LNET_NIDNET(dstnid);
2450         int               hops;
2451         __u32             order = 2;
2452
2453         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
2454          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
2455          * keep order 0 free for 0@lo and order 1 free for a local NID
2456          * match */
2457
2458         LASSERT (the_lnet.ln_init);
2459         LASSERT (the_lnet.ln_refcount > 0);
2460
2461         LNET_LOCK();
2462
2463         list_for_each (e, &the_lnet.ln_nis) {
2464                 ni = list_entry(e, lnet_ni_t, ni_list);
2465
2466                 if (ni->ni_nid == dstnid ||
2467                     (the_lnet.ln_ptlcompat > 0 &&
2468                      LNET_NIDNET(dstnid) == 0 &&
2469                      LNET_NIDADDR(dstnid) == LNET_NIDADDR(ni->ni_nid) &&
2470                      LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) != LOLND)) {
2471                         if (srcnidp != NULL)
2472                                 *srcnidp = dstnid;
2473                         if (orderp != NULL) {
2474                                 if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
2475                                         *orderp = 0;
2476                                 else
2477                                         *orderp = 1;
2478                         }
2479                         LNET_UNLOCK();
2480
2481                         return local_nid_dist_zero ? 0 : 1;
2482                 }
2483
2484                 if (LNET_NIDNET(ni->ni_nid) == dstnet ||
2485                     (the_lnet.ln_ptlcompat > 0 &&
2486                      dstnet == 0 &&
2487                      LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) != LOLND)) {
2488                         if (srcnidp != NULL)
2489                                 *srcnidp = ni->ni_nid;
2490                         if (orderp != NULL)
2491                                 *orderp = order;
2492                         LNET_UNLOCK();
2493                         return 1;
2494                 }
2495
2496                 order++;
2497         }
2498
2499         list_for_each (e, &the_lnet.ln_remote_nets) {
2500                 rnet = list_entry(e, lnet_remotenet_t, lrn_list);
2501
2502                 if (rnet->lrn_net == dstnet) {
2503                         LASSERT (!list_empty(&rnet->lrn_routes));
2504                         route = list_entry(rnet->lrn_routes.next,
2505                                            lnet_route_t, lr_list);
2506                         hops = rnet->lrn_hops;
2507                         if (srcnidp != NULL)
2508                                 *srcnidp = route->lr_gateway->lp_ni->ni_nid;
2509                         if (orderp != NULL)
2510                                 *orderp = order;
2511                         LNET_UNLOCK();
2512                         return hops + 1;
2513                 }
2514                 order++;
2515         }
2516
2517         LNET_UNLOCK();
2518         return -EHOSTUNREACH;
2519 }
2520
2521 int
2522 LNetSetAsync(lnet_process_id_t id, int nasync)
2523 {
2524 #ifdef __KERNEL__
2525         return 0;
2526 #else
2527         lnet_ni_t        *ni;
2528         lnet_remotenet_t *rnet;
2529         struct list_head *tmp;
2530         lnet_route_t     *route;
2531         lnet_nid_t       *nids;
2532         int               nnids;
2533         int               maxnids = 256;
2534         int               rc = 0;
2535         int               rc2;
2536
2537         /* Target on a local network? */
2538
2539         ni = lnet_net2ni(LNET_NIDNET(id.nid));
2540         if (ni != NULL) {
2541                 if (ni->ni_lnd->lnd_setasync != NULL)
2542                         rc = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2543                 lnet_ni_decref(ni);
2544                 return rc;
2545         }
2546
2547         /* Target on a remote network: apply to routers */
2548  again:
2549         LIBCFS_ALLOC(nids, maxnids * sizeof(*nids));
2550         if (nids == NULL)
2551                 return -ENOMEM;
2552         nnids = 0;
2553
2554         /* Snapshot all the router NIDs */
2555         LNET_LOCK();
2556         rnet = lnet_find_net_locked(LNET_NIDNET(id.nid));
2557         if (rnet != NULL) {
2558                 list_for_each(tmp, &rnet->lrn_routes) {
2559                         if (nnids == maxnids) {
2560                                 LNET_UNLOCK();
2561                                 LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2562                                 maxnids *= 2;
2563                                 goto again;
2564                         }
2565
2566                         route = list_entry(tmp, lnet_route_t, lr_list);
2567                         nids[nnids++] = route->lr_gateway->lp_nid;
2568                 }
2569         }
2570         LNET_UNLOCK();
2571
2572         /* set async on all the routers */
2573         while (nnids-- > 0) {
2574                 id.pid = LUSTRE_SRV_LNET_PID;
2575                 id.nid = nids[nnids];
2576
2577                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2578                 if (ni == NULL)
2579                         continue;
2580
2581                 if (ni->ni_lnd->lnd_setasync != NULL) {
2582                         rc2 = (ni->ni_lnd->lnd_setasync)(ni, id, nasync);
2583                         if (rc2 != 0)
2584                                 rc = rc2;
2585                 }
2586                 lnet_ni_decref(ni);
2587         }
2588
2589         LIBCFS_FREE(nids, maxnids * sizeof(*nids));
2590         return rc;
2591 #endif
2592 }