lnet/klnds/o2iblnd/o2iblnd.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
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  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/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 lnd_t the_o2iblnd = {
  44         .lnd_type       = O2IBLND,
  45         .lnd_startup    = kiblnd_startup,
  46         .lnd_shutdown   = kiblnd_shutdown,
  47         .lnd_ctl        = kiblnd_ctl,
  48         .lnd_query      = kiblnd_query,
  49         .lnd_send       = kiblnd_send,
  50         .lnd_recv       = kiblnd_recv,
  51 };
  52
  53 kib_data_t              kiblnd_data;
  54
  55 __u32
  56 kiblnd_cksum (void *ptr, int nob)
  57 {
  58         char  *c  = ptr;
  59         __u32  sum = 0;
  60
  61         while (nob-- > 0)
  62                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  63
  64         /* ensure I don't return 0 (== no checksum) */
  65         return (sum == 0) ? 1 : sum;
  66 }
  67
  68 static char *
  69 kiblnd_msgtype2str(int type)
  70 {
  71         switch (type) {
  72         case IBLND_MSG_CONNREQ:
  73                 return "CONNREQ";
  74
  75         case IBLND_MSG_CONNACK:
  76                 return "CONNACK";
  77
  78         case IBLND_MSG_NOOP:
  79                 return "NOOP";
  80
  81         case IBLND_MSG_IMMEDIATE:
  82                 return "IMMEDIATE";
  83
  84         case IBLND_MSG_PUT_REQ:
  85                 return "PUT_REQ";
  86
  87         case IBLND_MSG_PUT_NAK:
  88                 return "PUT_NAK";
  89
  90         case IBLND_MSG_PUT_ACK:
  91                 return "PUT_ACK";
  92
  93         case IBLND_MSG_PUT_DONE:
  94                 return "PUT_DONE";
  95
  96         case IBLND_MSG_GET_REQ:
  97                 return "GET_REQ";
  98
  99         case IBLND_MSG_GET_DONE:
 100                 return "GET_DONE";
 101
 102         default:
 103                 return "???";
 104         }
 105 }
 106
 107 static int
 108 kiblnd_msgtype2size(int type)
 109 {
 110         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 111
 112         switch (type) {
 113         case IBLND_MSG_CONNREQ:
 114         case IBLND_MSG_CONNACK:
 115                 return hdr_size + sizeof(kib_connparams_t);
 116
 117         case IBLND_MSG_NOOP:
 118                 return hdr_size;
 119
 120         case IBLND_MSG_IMMEDIATE:
 121                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 122
 123         case IBLND_MSG_PUT_REQ:
 124                 return hdr_size + sizeof(kib_putreq_msg_t);
 125
 126         case IBLND_MSG_PUT_ACK:
 127                 return hdr_size + sizeof(kib_putack_msg_t);
 128
 129         case IBLND_MSG_GET_REQ:
 130                 return hdr_size + sizeof(kib_get_msg_t);
 131
 132         case IBLND_MSG_PUT_NAK:
 133         case IBLND_MSG_PUT_DONE:
 134         case IBLND_MSG_GET_DONE:
 135                 return hdr_size + sizeof(kib_completion_msg_t);
 136         default:
 137                 return -1;
 138         }
 139 }
 140
 141 static int
 142 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 143 {
 144         kib_rdma_desc_t   *rd;
 145         int                nob;
 146         int                n;
 147         int                i;
 148
 149         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 150                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 151
 152         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 153                               &msg->ibm_u.get.ibgm_rd :
 154                               &msg->ibm_u.putack.ibpam_rd;
 155
 156         if (flip) {
 157                 __swab32s(&rd->rd_key);
 158                 __swab32s(&rd->rd_nfrags);
 159         }
 160
 161         n = rd->rd_nfrags;
 162
 163         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 164                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 165                        n, IBLND_MAX_RDMA_FRAGS);
 166                 return 1;
 167         }
 168
 169         nob = offsetof (kib_msg_t, ibm_u) +
 170               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 171
 172         if (msg->ibm_nob < nob) {
 173                 CERROR("Short %s: %d(%d)\n",
 174                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 175                 return 1;
 176         }
 177
 178         if (!flip)
 179                 return 0;
 180
 181         for (i = 0; i < n; i++) {
 182                 __swab32s(&rd->rd_frags[i].rf_nob);
 183                 __swab64s(&rd->rd_frags[i].rf_addr);
 184         }
 185
 186         return 0;
 187 }
 188
 189 void
 190 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 191                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 192 {
 193         kib_net_t *net = ni->ni_data;
 194
 195         /* CAVEAT EMPTOR! all message fields not set here should have been
 196          * initialised previously. */
 197         msg->ibm_magic    = IBLND_MSG_MAGIC;
 198         msg->ibm_version  = version;
 199         /*   ibm_type */
 200         msg->ibm_credits  = credits;
 201         /*   ibm_nob */
 202         msg->ibm_cksum    = 0;
 203         msg->ibm_srcnid   = ni->ni_nid;
 204         msg->ibm_srcstamp = net->ibn_incarnation;
 205         msg->ibm_dstnid   = dstnid;
 206         msg->ibm_dststamp = dststamp;
 207
 208         if (*kiblnd_tunables.kib_cksum) {
 209                 /* NB ibm_cksum zero while computing cksum */
 210                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 211         }
 212 }
 213
 214 int
 215 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 216 {
 217         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 218         __u32     msg_cksum;
 219         __u16     version;
 220         int       msg_nob;
 221         int       flip;
 222
 223         /* 6 bytes are enough to have received magic + version */
 224         if (nob < 6) {
 225                 CERROR("Short message: %d\n", nob);
 226                 return -EPROTO;
 227         }
 228
 229         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 230                 flip = 0;
 231         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 232                 flip = 1;
 233         } else {
 234                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 235                 return -EPROTO;
 236         }
 237
 238         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 239         if (version != IBLND_MSG_VERSION &&
 240             version != IBLND_MSG_VERSION_1) {
 241                 CERROR("Bad version: %x\n", version);
 242                 return -EPROTO;
 243         }
 244
 245         if (nob < hdr_size) {
 246                 CERROR("Short message: %d\n", nob);
 247                 return -EPROTO;
 248         }
 249
 250         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 251         if (msg_nob > nob) {
 252                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 253                 return -EPROTO;
 254         }
 255
 256         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 257          * gets flipped */
 258         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 259         msg->ibm_cksum = 0;
 260         if (msg_cksum != 0 &&
 261             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 262                 CERROR("Bad checksum\n");
 263                 return -EPROTO;
 264         }
 265
 266         msg->ibm_cksum = msg_cksum;
 267
 268         if (flip) {
 269                 /* leave magic unflipped as a clue to peer endianness */
 270                 msg->ibm_version = version;
 271                 CLASSERT (sizeof(msg->ibm_type) == 1);
 272                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 273                 msg->ibm_nob     = msg_nob;
 274                 __swab64s(&msg->ibm_srcnid);
 275                 __swab64s(&msg->ibm_srcstamp);
 276                 __swab64s(&msg->ibm_dstnid);
 277                 __swab64s(&msg->ibm_dststamp);
 278         }
 279
 280         if (msg->ibm_srcnid == LNET_NID_ANY) {
 281                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 282                 return -EPROTO;
 283         }
 284
 285         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 286                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 287                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 288                 return -EPROTO;
 289         }
 290
 291         switch (msg->ibm_type) {
 292         default:
 293                 CERROR("Unknown message type %x\n", msg->ibm_type);
 294                 return -EPROTO;
 295
 296         case IBLND_MSG_NOOP:
 297         case IBLND_MSG_IMMEDIATE:
 298         case IBLND_MSG_PUT_REQ:
 299                 break;
 300
 301         case IBLND_MSG_PUT_ACK:
 302         case IBLND_MSG_GET_REQ:
 303                 if (kiblnd_unpack_rd(msg, flip))
 304                         return -EPROTO;
 305                 break;
 306
 307         case IBLND_MSG_PUT_NAK:
 308         case IBLND_MSG_PUT_DONE:
 309         case IBLND_MSG_GET_DONE:
 310                 if (flip)
 311                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 312                 break;
 313
 314         case IBLND_MSG_CONNREQ:
 315         case IBLND_MSG_CONNACK:
 316                 if (flip) {
 317                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 318                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 319                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 320                 }
 321                 break;
 322         }
 323         return 0;
 324 }
 325
 326 int
 327 kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 328 {
 329         kib_peer_t     *peer;
 330         kib_net_t      *net = ni->ni_data;
 331         unsigned long   flags;
 332
 333         LASSERT (net != NULL);
 334         LASSERT (nid != LNET_NID_ANY);
 335
 336         LIBCFS_ALLOC(peer, sizeof(*peer));
 337         if (peer == NULL) {
 338                 CERROR("Cannot allocate peer\n");
 339                 return -ENOMEM;
 340         }
 341
 342         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 343
 344         peer->ibp_ni = ni;
 345         peer->ibp_nid = nid;
 346         peer->ibp_error = 0;
 347         peer->ibp_last_alive = 0;
 348         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 349
 350         INIT_LIST_HEAD(&peer->ibp_list);       /* not in the peer table yet */
 351         INIT_LIST_HEAD(&peer->ibp_conns);
 352         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 353
 354         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 355
 356         /* always called with a ref on ni, which prevents ni being shutdown */
 357         LASSERT (net->ibn_shutdown == 0);
 358
 359         /* npeers only grows with the global lock held */
 360         atomic_inc(&net->ibn_npeers);
 361
 362         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 363
 364         *peerp = peer;
 365         return 0;
 366 }
 367
 368 void
 369 kiblnd_destroy_peer (kib_peer_t *peer)
 370 {
 371         kib_net_t *net = peer->ibp_ni->ni_data;
 372
 373         LASSERT (net != NULL);
 374         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 375         LASSERT (!kiblnd_peer_active(peer));
 376         LASSERT (peer->ibp_connecting == 0);
 377         LASSERT (peer->ibp_accepting == 0);
 378         LASSERT (list_empty(&peer->ibp_conns));
 379         LASSERT (list_empty(&peer->ibp_tx_queue));
 380
 381         LIBCFS_FREE(peer, sizeof(*peer));
 382
 383         /* NB a peer's connections keep a reference on their peer until
 384          * they are destroyed, so we can be assured that _all_ state to do
 385          * with this peer has been cleaned up when its refcount drops to
 386          * zero. */
 387         atomic_dec(&net->ibn_npeers);
 388 }
 389
 390 kib_peer_t *
 391 kiblnd_find_peer_locked (lnet_nid_t nid)
 392 {
 393         /* the caller is responsible for accounting the additional reference
 394          * that this creates */
 395         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 396         struct list_head *tmp;
 397         kib_peer_t       *peer;
 398
 399         list_for_each (tmp, peer_list) {
 400
 401                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 402
 403                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 404                          peer->ibp_accepting > 0 ||
 405                          !list_empty(&peer->ibp_conns));  /* active conn */
 406
 407                 if (peer->ibp_nid != nid)
 408                         continue;
 409
 410                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 411                        peer, libcfs_nid2str(nid),
 412                        atomic_read(&peer->ibp_refcount),
 413                        peer->ibp_version);
 414                 return peer;
 415         }
 416         return NULL;
 417 }
 418
 419 void
 420 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 421 {
 422         LASSERT (list_empty(&peer->ibp_conns));
 423
 424         LASSERT (kiblnd_peer_active(peer));
 425         list_del_init(&peer->ibp_list);
 426         /* lose peerlist's ref */
 427         kiblnd_peer_decref(peer);
 428 }
 429
 430 int
 431 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 432                       lnet_nid_t *nidp, int *count)
 433 {
 434         kib_peer_t        *peer;
 435         struct list_head  *ptmp;
 436         int                i;
 437         unsigned long      flags;
 438
 439         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 440
 441         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 442
 443                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 444
 445                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 446                         LASSERT (peer->ibp_connecting > 0 ||
 447                                  peer->ibp_accepting > 0 ||
 448                                  !list_empty(&peer->ibp_conns));
 449
 450                         if (peer->ibp_ni != ni)
 451                                 continue;
 452
 453                         if (index-- > 0)
 454                                 continue;
 455
 456                         *nidp = peer->ibp_nid;
 457                         *count = atomic_read(&peer->ibp_refcount);
 458
 459                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 460                                                flags);
 461                         return 0;
 462                 }
 463         }
 464
 465         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 466         return -ENOENT;
 467 }
 468
 469 void
 470 kiblnd_del_peer_locked (kib_peer_t *peer)
 471 {
 472         struct list_head *ctmp;
 473         struct list_head *cnxt;
 474         kib_conn_t       *conn;
 475
 476         if (list_empty(&peer->ibp_conns)) {
 477                 kiblnd_unlink_peer_locked(peer);
 478         } else {
 479                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 480                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 481
 482                         kiblnd_close_conn_locked(conn, 0);
 483                 }
 484                 /* NB closing peer's last conn unlinked it. */
 485         }
 486         /* NB peer now unlinked; might even be freed if the peer table had the
 487          * last ref on it. */
 488 }
 489
 490 int
 491 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 492 {
 493         CFS_LIST_HEAD     (zombies);
 494         struct list_head  *ptmp;
 495         struct list_head  *pnxt;
 496         kib_peer_t        *peer;
 497         int                lo;
 498         int                hi;
 499         int                i;
 500         unsigned long      flags;
 501         int                rc = -ENOENT;
 502
 503         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 504
 505         if (nid != LNET_NID_ANY) {
 506                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 507         } else {
 508                 lo = 0;
 509                 hi = kiblnd_data.kib_peer_hash_size - 1;
 510         }
 511
 512         for (i = lo; i <= hi; i++) {
 513                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 514                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 515                         LASSERT (peer->ibp_connecting > 0 ||
 516                                  peer->ibp_accepting > 0 ||
 517                                  !list_empty(&peer->ibp_conns));
 518
 519                         if (peer->ibp_ni != ni)
 520                                 continue;
 521
 522                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 523                                 continue;
 524
 525                         if (!list_empty(&peer->ibp_tx_queue)) {
 526                                 LASSERT (list_empty(&peer->ibp_conns));
 527
 528                                 list_splice_init(&peer->ibp_tx_queue, &zombies);
 529                         }
 530
 531                         kiblnd_del_peer_locked(peer);
 532                         rc = 0;         /* matched something */
 533                 }
 534         }
 535
 536         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 537
 538         kiblnd_txlist_done(ni, &zombies, -EIO);
 539
 540         return rc;
 541 }
 542
 543 kib_conn_t *
 544 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 545 {
 546         kib_peer_t        *peer;
 547         struct list_head  *ptmp;
 548         kib_conn_t        *conn;
 549         struct list_head  *ctmp;
 550         int                i;
 551         unsigned long      flags;
 552
 553         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 554
 555         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 556                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 557
 558                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 559                         LASSERT (peer->ibp_connecting > 0 ||
 560                                  peer->ibp_accepting > 0 ||
 561                                  !list_empty(&peer->ibp_conns));
 562
 563                         if (peer->ibp_ni != ni)
 564                                 continue;
 565
 566                         list_for_each (ctmp, &peer->ibp_conns) {
 567                                 if (index-- > 0)
 568                                         continue;
 569
 570                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 571                                 kiblnd_conn_addref(conn);
 572                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 573                                                        flags);
 574                                 return conn;
 575                         }
 576                 }
 577         }
 578
 579         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 580         return NULL;
 581 }
 582
 583 void
 584 kiblnd_debug_rx (kib_rx_t *rx)
 585 {
 586         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 587                rx, rx->rx_status, rx->rx_msg->ibm_type,
 588                rx->rx_msg->ibm_credits);
 589 }
 590
 591 void
 592 kiblnd_debug_tx (kib_tx_t *tx)
 593 {
 594         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 595                "cookie "LPX64" msg %s%s type %x cred %d\n",
 596                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 597                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 598                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 599                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 600                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 601 }
 602
 603 void
 604 kiblnd_debug_conn (kib_conn_t *conn)
 605 {
 606         struct list_head *tmp;
 607         int               i;
 608
 609         spin_lock(&conn->ibc_lock);
 610
 611         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 612                atomic_read(&conn->ibc_refcount), conn,
 613                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 614         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 615                conn->ibc_state, conn->ibc_noops_posted,
 616                conn->ibc_nsends_posted, conn->ibc_credits,
 617                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 618         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 619
 620         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 621         list_for_each(tmp, &conn->ibc_early_rxs)
 622                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 623
 624         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 625         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 629         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 630                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 631
 632         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 633         list_for_each(tmp, &conn->ibc_tx_queue)
 634                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 635
 636         CDEBUG(D_CONSOLE, "   active_txs:\n");
 637         list_for_each(tmp, &conn->ibc_active_txs)
 638                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 639
 640         CDEBUG(D_CONSOLE, "   rxs:\n");
 641         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 642                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 643
 644         spin_unlock(&conn->ibc_lock);
 645 }
 646
 647 int
 648 kiblnd_translate_mtu(int value)
 649 {
 650         switch (value) {
 651         default:
 652                 return -1;
 653         case 0:
 654                 return 0;
 655         case 256:
 656                 return IB_MTU_256;
 657         case 512:
 658                 return IB_MTU_512;
 659         case 1024:
 660                 return IB_MTU_1024;
 661         case 2048:
 662                 return IB_MTU_2048;
 663         case 4096:
 664                 return IB_MTU_4096;
 665         }
 666 }
 667
 668 static void
 669 kiblnd_setup_mtu(struct rdma_cm_id *cmid)
 670 {
 671         unsigned long flags;
 672         int           mtu;
 673
 674         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 675         if (cmid->route.path_rec == NULL)
 676                 return;
 677
 678         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 679
 680         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 681         LASSERT (mtu >= 0);
 682         if (mtu != 0)
 683                 cmid->route.path_rec->mtu = mtu;
 684
 685         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 686 }
 687
 688 kib_conn_t *
 689 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 690                    int state, int version)
 691 {
 692         /* CAVEAT EMPTOR:
 693          * If the new conn is created successfully it takes over the caller's
 694          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 695          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 696          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 697          * to destroy 'cmid' here since I'm called from the CM which still has
 698          * its ref on 'cmid'). */
 699         kib_net_t              *net = peer->ibp_ni->ni_data;
 700         struct ib_qp_init_attr *init_qp_attr;
 701         kib_conn_t             *conn;
 702         struct ib_cq           *cq;
 703         unsigned long           flags;
 704         int                     rc;
 705         int                     i;
 706
 707         LASSERT (net != NULL);
 708         LASSERT (!in_interrupt());
 709
 710         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 711         if (init_qp_attr == NULL) {
 712                 CERROR("Can't allocate qp_attr for %s\n",
 713                        libcfs_nid2str(peer->ibp_nid));
 714                 goto failed_0;
 715         }
 716
 717         LIBCFS_ALLOC(conn, sizeof(*conn));
 718         if (conn == NULL) {
 719                 CERROR("Can't allocate connection for %s\n",
 720                        libcfs_nid2str(peer->ibp_nid));
 721                 goto failed_1;
 722         }
 723
 724         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 725
 726         conn->ibc_state = IBLND_CONN_INIT;
 727         conn->ibc_version = version;
 728         conn->ibc_peer = peer;                  /* I take the caller's ref */
 729         cmid->context = conn;                   /* for future CM callbacks */
 730         conn->ibc_cmid = cmid;
 731
 732         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 733         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 734         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 735         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 736         INIT_LIST_HEAD(&conn->ibc_active_txs);
 737         spin_lock_init(&conn->ibc_lock);
 738
 739         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 740         if (conn->ibc_connvars == NULL) {
 741                 CERROR("Can't allocate in-progress connection state\n");
 742                 goto failed_2;
 743         }
 744         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 745
 746         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 747         if (conn->ibc_rxs == NULL) {
 748                 CERROR("Cannot allocate RX buffers\n");
 749                 goto failed_2;
 750         }
 751         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 752
 753         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
 754                                 IBLND_RX_MSG_PAGES(version));
 755         if (rc != 0)
 756                 goto failed_2;
 757
 758         kiblnd_map_rx_descs(conn);
 759
 760 #ifdef HAVE_OFED_IB_COMP_VECTOR
 761         cq = ib_create_cq(cmid->device,
 762                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 763                           IBLND_CQ_ENTRIES(version), 0);
 764 #else
 765         cq = ib_create_cq(cmid->device,
 766                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 767                           IBLND_CQ_ENTRIES(version));
 768 #endif
 769         if (IS_ERR(cq)) {
 770                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 771                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 772                 goto failed_2;
 773         }
 774
 775         conn->ibc_cq = cq;
 776
 777         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 778         if (rc != 0) {
 779                 CERROR("Can't request completion notificiation: %d\n", rc);
 780                 goto failed_2;
 781         }
 782
 783         kiblnd_setup_mtu(cmid);
 784
 785         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 786         init_qp_attr->event_handler = kiblnd_qp_event;
 787         init_qp_attr->qp_context = conn;
 788         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 789         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 790         init_qp_attr->cap.max_send_sge = 1;
 791         init_qp_attr->cap.max_recv_sge = 1;
 792         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 793         init_qp_attr->qp_type = IB_QPT_RC;
 794         init_qp_attr->send_cq = cq;
 795         init_qp_attr->recv_cq = cq;
 796
 797         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 798         if (rc != 0) {
 799                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 800                        rc, init_qp_attr->cap.max_send_wr,
 801                        init_qp_attr->cap.max_recv_wr);
 802                 goto failed_2;
 803         }
 804
 805         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 806
 807         /* 1 ref for caller and each rxmsg */
 808         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 809         conn->ibc_nrx = IBLND_RX_MSGS(version);
 810
 811         /* post receives */
 812         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 813                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 814                                     IBLND_POSTRX_NO_CREDIT);
 815                 if (rc != 0) {
 816                         CERROR("Can't post rxmsg: %d\n", rc);
 817
 818                         /* Make posted receives complete */
 819                         kiblnd_abort_receives(conn);
 820
 821                         /* correct # of posted buffers
 822                          * NB locking needed now I'm racing with completion */
 823                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
 824                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 825                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 826                                                flags);
 827
 828                         /* cmid will be destroyed by CM(ofed) after cm_callback
 829                          * returned, so we can't refer it anymore
 830                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 831                         rdma_destroy_qp(conn->ibc_cmid);
 832                         conn->ibc_cmid = NULL;
 833
 834                         /* Drop my own and unused rxbuffer refcounts */
 835                         while (i++ <= IBLND_RX_MSGS(version))
 836                                 kiblnd_conn_decref(conn);
 837
 838                         return NULL;
 839                 }
 840         }
 841
 842         /* Init successful! */
 843         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 844                  state == IBLND_CONN_PASSIVE_WAIT);
 845         conn->ibc_state = state;
 846
 847         /* 1 more conn */
 848         atomic_inc(&net->ibn_nconns);
 849         return conn;
 850
 851  failed_2:
 852         kiblnd_destroy_conn(conn);
 853  failed_1:
 854         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 855  failed_0:
 856         return NULL;
 857 }
 858
 859 void
 860 kiblnd_destroy_conn (kib_conn_t *conn)
 861 {
 862         struct rdma_cm_id *cmid = conn->ibc_cmid;
 863         kib_peer_t        *peer = conn->ibc_peer;
 864         int                rc;
 865
 866         LASSERT (!in_interrupt());
 867         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 868         LASSERT (list_empty(&conn->ibc_early_rxs));
 869         LASSERT (list_empty(&conn->ibc_tx_queue));
 870         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 871         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 872         LASSERT (list_empty(&conn->ibc_active_txs));
 873         LASSERT (conn->ibc_noops_posted == 0);
 874         LASSERT (conn->ibc_nsends_posted == 0);
 875
 876         switch (conn->ibc_state) {
 877         default:
 878                 /* conn must be completely disengaged from the network */
 879                 LBUG();
 880
 881         case IBLND_CONN_DISCONNECTED:
 882                 /* connvars should have been freed already */
 883                 LASSERT (conn->ibc_connvars == NULL);
 884                 break;
 885
 886         case IBLND_CONN_INIT:
 887                 break;
 888         }
 889
 890         /* conn->ibc_cmid might be destroyed by CM already */
 891         if (cmid != NULL && cmid->qp != NULL)
 892                 rdma_destroy_qp(cmid);
 893
 894         if (conn->ibc_cq != NULL) {
 895                 rc = ib_destroy_cq(conn->ibc_cq);
 896                 if (rc != 0)
 897                         CWARN("Error destroying CQ: %d\n", rc);
 898         }
 899
 900         if (conn->ibc_rx_pages != NULL)
 901                 kiblnd_unmap_rx_descs(conn);
 902
 903         if (conn->ibc_rxs != NULL) {
 904                 LIBCFS_FREE(conn->ibc_rxs,
 905                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 906         }
 907
 908         if (conn->ibc_connvars != NULL)
 909                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 910
 911         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 912         if (conn->ibc_state != IBLND_CONN_INIT) {
 913                 kib_net_t *net = peer->ibp_ni->ni_data;
 914
 915                 kiblnd_peer_decref(peer);
 916                 rdma_destroy_id(cmid);
 917                 atomic_dec(&net->ibn_nconns);
 918         }
 919
 920         LIBCFS_FREE(conn, sizeof(*conn));
 921 }
 922
 923 int
 924 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 925 {
 926         kib_conn_t         *conn;
 927         struct list_head   *ctmp;
 928         struct list_head   *cnxt;
 929         int                 count = 0;
 930
 931         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 932                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 933
 934                 CDEBUG(D_NET, "Closing conn -> %s, "
 935                               "version: %x, reason: %d\n",
 936                        libcfs_nid2str(peer->ibp_nid),
 937                        conn->ibc_version, why);
 938
 939                 kiblnd_close_conn_locked(conn, why);
 940                 count++;
 941         }
 942
 943         return count;
 944 }
 945
 946 int
 947 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 948                                  int version, __u64 incarnation)
 949 {
 950         kib_conn_t         *conn;
 951         struct list_head   *ctmp;
 952         struct list_head   *cnxt;
 953         int                 count = 0;
 954
 955         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 956                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 957
 958                 if (conn->ibc_version     == version &&
 959                     conn->ibc_incarnation == incarnation)
 960                         continue;
 961
 962                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
 963                               "incarnation:"LPX64"(%x, "LPX64")\n",
 964                        libcfs_nid2str(peer->ibp_nid),
 965                        conn->ibc_version, conn->ibc_incarnation,
 966                        version, incarnation);
 967
 968                 kiblnd_close_conn_locked(conn, -ESTALE);
 969                 count++;
 970         }
 971
 972         return count;
 973 }
 974
 975 int
 976 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 977 {
 978         kib_peer_t         *peer;
 979         struct list_head   *ptmp;
 980         struct list_head   *pnxt;
 981         int                 lo;
 982         int                 hi;
 983         int                 i;
 984         unsigned long       flags;
 985         int                 count = 0;
 986
 987         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 988
 989         if (nid != LNET_NID_ANY)
 990                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 991         else {
 992                 lo = 0;
 993                 hi = kiblnd_data.kib_peer_hash_size - 1;
 994         }
 995
 996         for (i = lo; i <= hi; i++) {
 997                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 998
 999                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1000                         LASSERT (peer->ibp_connecting > 0 ||
1001                                  peer->ibp_accepting > 0 ||
1002                                  !list_empty(&peer->ibp_conns));
1003
1004                         if (peer->ibp_ni != ni)
1005                                 continue;
1006
1007                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1008                                 continue;
1009
1010                         count += kiblnd_close_peer_conns_locked(peer, 0);
1011                 }
1012         }
1013
1014         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1015
1016         /* wildcards always succeed */
1017         if (nid == LNET_NID_ANY)
1018                 return 0;
1019
1020         return (count == 0) ? -ENOENT : 0;
1021 }
1022
1023 int
1024 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1025 {
1026         struct libcfs_ioctl_data *data = arg;
1027         int                       rc = -EINVAL;
1028
1029         switch(cmd) {
1030         case IOC_LIBCFS_GET_PEER: {
1031                 lnet_nid_t   nid = 0;
1032                 int          count = 0;
1033
1034                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1035                                           &nid, &count);
1036                 data->ioc_nid    = nid;
1037                 data->ioc_count  = count;
1038                 break;
1039         }
1040
1041         case IOC_LIBCFS_DEL_PEER: {
1042                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1043                 break;
1044         }
1045         case IOC_LIBCFS_GET_CONN: {
1046                 kib_conn_t *conn;
1047
1048                 rc = 0;
1049                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1050                 if (conn == NULL) {
1051                         rc = -ENOENT;
1052                         break;
1053                 }
1054
1055                 LASSERT (conn->ibc_cmid != NULL);
1056                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1057                 if (conn->ibc_cmid->route.path_rec == NULL)
1058                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1059                 else
1060                         data->ioc_u32[0] =
1061                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1062                 kiblnd_conn_decref(conn);
1063                 break;
1064         }
1065         case IOC_LIBCFS_CLOSE_CONNECTION: {
1066                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1067                 break;
1068         }
1069
1070         default:
1071                 break;
1072         }
1073
1074         return rc;
1075 }
1076
1077 void
1078 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, time_t *when)
1079 {
1080         cfs_time_t     last_alive = 0;
1081         rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1082         kib_peer_t    *peer;
1083         unsigned long  flags;
1084
1085         read_lock_irqsave(glock, flags);
1086
1087         peer = kiblnd_find_peer_locked(nid);
1088         if (peer != NULL) {
1089                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1090                          peer->ibp_accepting > 0 ||
1091                          !list_empty(&peer->ibp_conns));  /* active conn */
1092                 last_alive = peer->ibp_last_alive;
1093         }
1094
1095         read_unlock_irqrestore(glock, flags);
1096
1097         if (last_alive != 0)
1098                 *when = cfs_time_current_sec() -
1099                         cfs_duration_sec(cfs_time_current() - last_alive);
1100
1101         /* peer is not persistent in hash, trigger peer creation
1102          * and connection establishment with a NULL tx */
1103         if (peer == NULL)
1104                 kiblnd_launch_tx(ni, NULL, nid);
1105         return;
1106 }
1107
1108 void
1109 kiblnd_free_pages (kib_pages_t *p)
1110 {
1111         int         npages = p->ibp_npages;
1112         int         i;
1113
1114         LASSERT (p->ibp_device == NULL);
1115
1116         for (i = 0; i < npages; i++)
1117                 if (p->ibp_pages[i] != NULL)
1118                         __free_page(p->ibp_pages[i]);
1119
1120         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1121 }
1122
1123 int
1124 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1125 {
1126         kib_pages_t   *p;
1127         int            i;
1128
1129         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1130         if (p == NULL) {
1131                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1132                 return -ENOMEM;
1133         }
1134
1135         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1136         p->ibp_npages = npages;
1137
1138         for (i = 0; i < npages; i++) {
1139                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1140                 if (p->ibp_pages[i] == NULL) {
1141                         CERROR("Can't allocate page %d of %d\n", i, npages);
1142                         kiblnd_free_pages(p);
1143                         return -ENOMEM;
1144                 }
1145         }
1146
1147         *pp = p;
1148         return 0;
1149 }
1150
1151 void
1152 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1153 {
1154         kib_rx_t *rx;
1155         int       i;
1156
1157         LASSERT (conn->ibc_rxs != NULL);
1158         LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1159
1160         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1161                 rx = &conn->ibc_rxs[i];
1162
1163                 LASSERT (rx->rx_nob >= 0); /* not posted */
1164
1165                 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1166                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1167                                                           rx->rx_msgaddr),
1168                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1169         }
1170
1171         conn->ibc_rx_pages->ibp_device = NULL;
1172
1173         kiblnd_free_pages(conn->ibc_rx_pages);
1174
1175         conn->ibc_rx_pages = NULL;
1176 }
1177
1178 void
1179 kiblnd_map_rx_descs(kib_conn_t *conn)
1180 {
1181         kib_rx_t       *rx;
1182         struct page    *pg;
1183         int             pg_off;
1184         int             ipg;
1185         int             i;
1186
1187         for (pg_off = ipg = i = 0;
1188              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1189                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1190                 rx = &conn->ibc_rxs[i];
1191
1192                 rx->rx_conn = conn;
1193                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1194
1195                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1196                                                        rx->rx_msg, IBLND_MSG_SIZE,
1197                                                        DMA_FROM_DEVICE);
1198                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1199                                                    rx->rx_msgaddr));
1200                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1201
1202                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1203                        i, rx->rx_msg, rx->rx_msgaddr,
1204                        lnet_page2phys(pg) + pg_off);
1205
1206                 pg_off += IBLND_MSG_SIZE;
1207                 LASSERT (pg_off <= PAGE_SIZE);
1208
1209                 if (pg_off == PAGE_SIZE) {
1210                         pg_off = 0;
1211                         ipg++;
1212                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1213                 }
1214         }
1215
1216         conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1217 }
1218
1219 static void
1220 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1221 {
1222         kib_tx_t       *tx;
1223         int             i;
1224
1225         LASSERT (tpo->tpo_pool.po_allocated == 0);
1226         LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1227
1228         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1229                 tx = &tpo->tpo_tx_descs[i];
1230                 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1231                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1232                                                           tx->tx_msgaddr),
1233                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1234         }
1235         tpo->tpo_tx_pages->ibp_device = NULL;
1236 }
1237
1238 static void
1239 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1240 {
1241         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1242         kib_pool_t     *pool  = &tpo->tpo_pool;
1243         kib_net_t      *net   = pool->po_owner->ps_net;
1244         struct page    *page;
1245         kib_tx_t       *tx;
1246         int             page_offset;
1247         int             ipage;
1248         int             i;
1249
1250         LASSERT (net != NULL);
1251         LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1252                  net->ibn_dev->ibd_cmid->device != NULL);
1253
1254         /* pre-mapped messages are not bigger than 1 page */
1255         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1256
1257         /* No fancy arithmetic when we do the buffer calculations */
1258         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1259
1260         txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1261
1262         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1263                 page = txpgs->ibp_pages[ipage];
1264                 tx = &tpo->tpo_tx_descs[i];
1265
1266                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1267                                            page_offset);
1268
1269                 tx->tx_msgaddr = kiblnd_dma_map_single(
1270                         txpgs->ibp_device, tx->tx_msg,
1271                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1272                 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1273                                                    tx->tx_msgaddr));
1274                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1275
1276                 list_add(&tx->tx_list, &pool->po_free_list);
1277
1278                 page_offset += IBLND_MSG_SIZE;
1279                 LASSERT (page_offset <= PAGE_SIZE);
1280
1281                 if (page_offset == PAGE_SIZE) {
1282                         page_offset = 0;
1283                         ipage++;
1284                         LASSERT (ipage <= txpgs->ibp_npages);
1285                 }
1286         }
1287 }
1288
1289 struct ib_mr *
1290 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1291 {
1292         __u64   index;
1293
1294         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1295
1296         if (net->ibn_dev->ibd_nmrs == 1)
1297                 return net->ibn_dev->ibd_mrs[0];
1298
1299         index = addr >> net->ibn_dev->ibd_mr_shift;
1300
1301         if (index <  net->ibn_dev->ibd_nmrs &&
1302             index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1303                 return net->ibn_dev->ibd_mrs[index];
1304
1305         return NULL;
1306 }
1307
1308 struct ib_mr *
1309 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1310 {
1311         struct ib_mr *prev_mr;
1312         struct ib_mr *mr;
1313         int           i;
1314
1315         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1316
1317         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1318             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1319                 return NULL;
1320
1321         if (net->ibn_dev->ibd_nmrs == 1)
1322                 return net->ibn_dev->ibd_mrs[0];
1323
1324         for (i = 0, mr = prev_mr = NULL;
1325              i < rd->rd_nfrags; i++) {
1326                 mr = kiblnd_find_dma_mr(net,
1327                                         rd->rd_frags[i].rf_addr,
1328                                         rd->rd_frags[i].rf_nob);
1329                 if (prev_mr == NULL)
1330                         prev_mr = mr;
1331
1332                 if (mr == NULL || prev_mr != mr) {
1333                         /* Can't covered by one single MR */
1334                         mr = NULL;
1335                         break;
1336                 }
1337         }
1338
1339         return mr;
1340 }
1341
1342 void
1343 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1344 {
1345         LASSERT (pool->fpo_map_count == 0);
1346
1347         if (pool->fpo_fmr_pool != NULL)
1348                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1349
1350         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1351 }
1352
1353 void
1354 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1355 {
1356         kib_fmr_pool_t *pool;
1357
1358         while (!list_empty(head)) {
1359                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1360                 list_del(&pool->fpo_list);
1361                 kiblnd_destroy_fmr_pool(pool);
1362         }
1363 }
1364
1365 int
1366 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1367 {
1368         /* FMR pool for RDMA */
1369         kib_fmr_pool_t          *fpo;
1370         struct ib_fmr_pool_param param = {
1371                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1372                 .page_shift        = PAGE_SHIFT,
1373                 .access            = (IB_ACCESS_LOCAL_WRITE |
1374                                       IB_ACCESS_REMOTE_WRITE),
1375                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1376                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1377                 .flush_function    = NULL,
1378                 .flush_arg         = NULL,
1379                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1380         int rc;
1381
1382         LASSERT (fps->fps_net->ibn_dev != NULL &&
1383                  fps->fps_net->ibn_dev->ibd_pd != NULL);
1384
1385         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1386         if (fpo == NULL)
1387                 return -ENOMEM;
1388
1389         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1390         fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, &param);
1391         if (IS_ERR(fpo->fpo_fmr_pool)) {
1392                 CERROR("Failed to create FMR pool: %ld\n",
1393                        PTR_ERR(fpo->fpo_fmr_pool));
1394                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1395                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1396                 return rc;
1397         }
1398
1399         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1400         fpo->fpo_owner    = fps;
1401         *pp_fpo = fpo;
1402
1403         return 0;
1404 }
1405
1406 static void
1407 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1408 {
1409         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1410 }
1411
1412 static int
1413 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1414 {
1415         kib_fmr_pool_t *fpo;
1416         int             rc;
1417
1418         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1419
1420         fps->fps_net = net;
1421         spin_lock_init(&fps->fps_lock);
1422         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1423         rc = kiblnd_create_fmr_pool(fps, &fpo);
1424         if (rc == 0)
1425                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1426
1427         return rc;
1428 }
1429
1430 void
1431 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1432 {
1433         CFS_LIST_HEAD     (zombies);
1434         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1435         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1436         kib_fmr_pool_t    *tmp;
1437         int                rc;
1438
1439         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1440         LASSERT (rc == 0);
1441
1442         if (status != 0) {
1443                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1444                 LASSERT (rc == 0);
1445         }
1446
1447         fmr->fmr_pool = NULL;
1448         fmr->fmr_pfmr = NULL;
1449
1450         spin_lock(&fps->fps_lock);
1451         fpo->fpo_map_count --;  /* decref the pool */
1452
1453         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1454                 /* the first pool is persistent */
1455                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1456                         continue;
1457
1458                 if (fpo->fpo_map_count == 0 &&  /* no more reference */
1459                     cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1460                         list_move(&fpo->fpo_list, &zombies);
1461                         fps->fps_version ++;
1462                 }
1463         }
1464         spin_unlock(&fps->fps_lock);
1465
1466         if (!list_empty(&zombies))
1467                 kiblnd_destroy_fmr_pool_list(&zombies);
1468 }
1469
1470 int
1471 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1472                     __u64 iov, kib_fmr_t *fmr)
1473 {
1474         struct ib_pool_fmr *pfmr;
1475         kib_fmr_pool_t     *fpo;
1476         __u64               version;
1477         int                 rc;
1478
1479         LASSERT (fps->fps_net->ibn_with_fmr);
1480  again:
1481         spin_lock(&fps->fps_lock);
1482         version = fps->fps_version;
1483         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1484                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1485                 fpo->fpo_map_count ++;
1486                 spin_unlock(&fps->fps_lock);
1487
1488                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1489                                             pages, npages, iov);
1490                 if (likely(!IS_ERR(pfmr))) {
1491                         fmr->fmr_pool = fpo;
1492                         fmr->fmr_pfmr = pfmr;
1493                         return 0;
1494                 }
1495
1496                 spin_lock(&fps->fps_lock);
1497                 fpo->fpo_map_count --;
1498                 if (PTR_ERR(pfmr) != -EAGAIN) {
1499                         spin_unlock(&fps->fps_lock);
1500                         return PTR_ERR(pfmr);
1501                 }
1502
1503                 /* EAGAIN and ... */
1504                 if (version != fps->fps_version) {
1505                         spin_unlock(&fps->fps_lock);
1506                         goto again;
1507                 }
1508         }
1509
1510         if (fps->fps_increasing) {
1511                 spin_unlock(&fps->fps_lock);
1512                 CDEBUG(D_NET, "Another thread is allocating new "
1513                               "FMR pool, waiting for her to complete\n");
1514                 schedule();
1515                 goto again;
1516
1517         }
1518
1519         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1520                 /* someone failed recently */
1521                 spin_unlock(&fps->fps_lock);
1522                 return -EAGAIN;
1523         }
1524
1525         fps->fps_increasing = 1;
1526         spin_unlock(&fps->fps_lock);
1527
1528         CDEBUG(D_NET, "Allocate new FMR pool\n");
1529         rc = kiblnd_create_fmr_pool(fps, &fpo);
1530         spin_lock(&fps->fps_lock);
1531         fps->fps_increasing = 0;
1532         if (rc == 0) {
1533                 fps->fps_version ++;
1534                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1535         } else {
1536                 fps->fps_next_retry = cfs_time_shift(10);
1537         }
1538         spin_unlock(&fps->fps_lock);
1539
1540         goto again;
1541 }
1542
1543 static void
1544 kiblnd_fini_pool(kib_pool_t *pool)
1545 {
1546         LASSERT (list_empty(&pool->po_free_list));
1547         LASSERT (pool->po_allocated == 0);
1548
1549         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1550 }
1551
1552 static void
1553 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1554 {
1555         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1556
1557         memset(pool, 0, sizeof(kib_pool_t));
1558         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1559         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1560         pool->po_owner    = ps;
1561         pool->po_size     = size;
1562 }
1563
1564 void
1565 kiblnd_destroy_pool_list(kib_poolset_t *ps, struct list_head *head)
1566 {
1567         kib_pool_t *pool;
1568
1569         while (!list_empty(head)) {
1570                 pool = list_entry(head->next, kib_pool_t, po_list);
1571                 list_del(&pool->po_list);
1572                 ps->ps_pool_destroy(pool);
1573         }
1574 }
1575
1576 static void
1577 kiblnd_fini_pool_set(kib_poolset_t *ps)
1578 {
1579         kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1580 }
1581
1582 static int
1583 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1584                      char *name, int size,
1585                      kib_ps_pool_create_t po_create,
1586                      kib_ps_pool_destroy_t po_destroy,
1587                      kib_ps_node_init_t nd_init,
1588                      kib_ps_node_fini_t nd_fini)
1589 {
1590         kib_pool_t    *pool;
1591         int            rc;
1592
1593         memset(ps, 0, sizeof(kib_poolset_t));
1594
1595         ps->ps_net          = net;
1596         ps->ps_pool_create  = po_create;
1597         ps->ps_pool_destroy = po_destroy;
1598         ps->ps_node_init    = nd_init;
1599         ps->ps_node_fini    = nd_fini;
1600         ps->ps_pool_size    = size;
1601         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1602         spin_lock_init(&ps->ps_lock);
1603         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1604
1605         rc = ps->ps_pool_create(ps, size, &pool);
1606         if (rc == 0)
1607                 list_add(&pool->po_list, &ps->ps_pool_list);
1608         else
1609                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1610
1611         return rc;
1612 }
1613
1614 void
1615 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1616 {
1617         CFS_LIST_HEAD  (zombies);
1618         kib_poolset_t  *ps = pool->po_owner;
1619         kib_pool_t     *tmp;
1620         cfs_time_t      now = cfs_time_current();
1621
1622         spin_lock(&ps->ps_lock);
1623
1624         if (ps->ps_node_fini != NULL)
1625                 ps->ps_node_fini(pool, node);
1626
1627         LASSERT (pool->po_allocated > 0);
1628         list_add(node, &pool->po_free_list);
1629         pool->po_allocated --;
1630
1631         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1632                 /* the first pool is persistent */
1633                 if (ps->ps_pool_list.next == &pool->po_list)
1634                         continue;
1635
1636                 if (pool->po_allocated == 0 &&
1637                     cfs_time_aftereq(now, pool->po_deadline))
1638                         list_move(&pool->po_list, &zombies);
1639         }
1640         spin_unlock(&ps->ps_lock);
1641
1642         if (!list_empty(&zombies))
1643                 kiblnd_destroy_pool_list(ps, &zombies);
1644 }
1645
1646 struct list_head *
1647 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1648 {
1649         struct list_head  *node;
1650         kib_pool_t        *pool;
1651         int                rc;
1652
1653  again:
1654         spin_lock(&ps->ps_lock);
1655         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1656                 if (list_empty(&pool->po_free_list))
1657                         continue;
1658
1659                 pool->po_allocated ++;
1660                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1661                 node = pool->po_free_list.next;
1662                 list_del(node);
1663
1664                 if (ps->ps_node_init != NULL) {
1665                         /* still hold the lock */
1666                         ps->ps_node_init(pool, node);
1667                 }
1668                 spin_unlock(&ps->ps_lock);
1669                 return node;
1670         }
1671
1672         /* no available tx pool and ... */
1673         if (ps->ps_increasing) {
1674                 /* another thread is allocating a new pool */
1675                 spin_unlock(&ps->ps_lock);
1676                 CDEBUG(D_NET, "Another thread is allocating new "
1677                        "%s pool, waiting for her to complete\n",
1678                        ps->ps_name);
1679                 schedule();
1680                 goto again;
1681         }
1682
1683         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1684                 /* someone failed recently */
1685                 spin_unlock(&ps->ps_lock);
1686                 return NULL;
1687         }
1688
1689         ps->ps_increasing = 1;
1690         spin_unlock(&ps->ps_lock);
1691
1692         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1693
1694         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1695
1696         spin_lock(&ps->ps_lock);
1697         ps->ps_increasing = 0;
1698         if (rc == 0) {
1699                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1700         } else {
1701                 /* retry 10 seconds later */
1702                 ps->ps_next_retry = cfs_time_shift(10);
1703                 CERROR("Can't allocate new %s pool because out of memory\n",
1704                        ps->ps_name);
1705         }
1706         spin_unlock(&ps->ps_lock);
1707
1708         goto again;
1709 }
1710
1711 void
1712 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1713 {
1714         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1715         struct ib_mr        *mr  = pmr->pmr_mr;
1716
1717         pmr->pmr_mr = NULL;
1718         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1719         if (mr != NULL)
1720                 ib_dereg_mr(mr);
1721 }
1722
1723 int
1724 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1725                     __u64 *iova, kib_phys_mr_t **pp_pmr)
1726 {
1727         kib_phys_mr_t       *pmr;
1728         struct list_head    *node;
1729         int                  rc;
1730         int                  i;
1731
1732         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1733         if (node == NULL) {
1734                 CERROR("Failed to allocate PMR descriptor\n");
1735                 return -ENOMEM;
1736         }
1737
1738         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1739         for (i = 0; i < rd->rd_nfrags; i ++) {
1740                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1741                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1742         }
1743
1744         pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1745                                      pmr->pmr_ipb, rd->rd_nfrags,
1746                                      IB_ACCESS_LOCAL_WRITE |
1747                                      IB_ACCESS_REMOTE_WRITE,
1748                                      iova);
1749         if (!IS_ERR(pmr->pmr_mr)) {
1750                 pmr->pmr_iova = *iova;
1751                 *pp_pmr = pmr;
1752                 return 0;
1753         }
1754
1755         rc = PTR_ERR(pmr->pmr_mr);
1756         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1757
1758         pmr->pmr_mr = NULL;
1759         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1760
1761         return rc;
1762 }
1763
1764 static void
1765 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1766 {
1767         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1768         kib_phys_mr_t  *pmr;
1769
1770         LASSERT (pool->po_allocated == 0);
1771
1772         while (!list_empty(&pool->po_free_list)) {
1773                 pmr = list_entry(pool->po_free_list.next,
1774                                  kib_phys_mr_t, pmr_list);
1775
1776                 LASSERT (pmr->pmr_mr == NULL);
1777                 list_del(&pmr->pmr_list);
1778
1779                 if (pmr->pmr_ipb != NULL) {
1780                         LIBCFS_FREE(pmr->pmr_ipb,
1781                                     IBLND_MAX_RDMA_FRAGS *
1782                                     sizeof(struct ib_phys_buf));
1783                 }
1784
1785                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1786         }
1787
1788         kiblnd_fini_pool(pool);
1789         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1790 }
1791
1792 static int
1793 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1794 {
1795         kib_pmr_pool_t      *ppo;
1796         kib_pool_t          *pool;
1797         kib_phys_mr_t       *pmr;
1798         int                  i;
1799
1800         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1801         if (ppo == NULL) {
1802                 CERROR("Failed to allocate PMR pool\n");
1803                 return -ENOMEM;
1804         }
1805
1806         pool = &ppo->ppo_pool;
1807         kiblnd_init_pool(ps, pool, size);
1808
1809         for (i = 0; i < size; i++) {
1810                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1811                 if (pmr == NULL)
1812                         break;
1813
1814                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1815                 pmr->pmr_pool = ppo;
1816                 LIBCFS_ALLOC(pmr->pmr_ipb,
1817                              IBLND_MAX_RDMA_FRAGS *
1818                              sizeof(struct ib_phys_buf));
1819                 if (pmr->pmr_ipb == NULL)
1820                         break;
1821
1822                 list_add(&pmr->pmr_list, &pool->po_free_list);
1823         }
1824
1825         if (i < size) {
1826                 ps->ps_pool_destroy(pool);
1827                 return -ENOMEM;
1828         }
1829
1830         *pp_po = pool;
1831         return 0;
1832 }
1833
1834 static void
1835 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1836 {
1837         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1838         int             i;
1839
1840         LASSERT (pool->po_allocated == 0);
1841
1842         if (tpo->tpo_tx_pages != NULL) {
1843                 if (tpo->tpo_tx_pages->ibp_device != NULL)
1844                         kiblnd_unmap_tx_pool(tpo);
1845                 kiblnd_free_pages(tpo->tpo_tx_pages);
1846         }
1847
1848         if (tpo->tpo_tx_descs == NULL)
1849                 goto out;
1850
1851         for (i = 0; i < pool->po_size; i++) {
1852                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1853
1854                 list_del(&tx->tx_list);
1855                 if (tx->tx_pages != NULL)
1856                         LIBCFS_FREE(tx->tx_pages,
1857                                     LNET_MAX_IOV *
1858                                     sizeof(*tx->tx_pages));
1859                 if (tx->tx_frags != NULL)
1860                         LIBCFS_FREE(tx->tx_frags,
1861                                     IBLND_MAX_RDMA_FRAGS *
1862                                             sizeof(*tx->tx_frags));
1863                 if (tx->tx_wrq != NULL)
1864                         LIBCFS_FREE(tx->tx_wrq,
1865                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1866                                     sizeof(*tx->tx_wrq));
1867                 if (tx->tx_sge != NULL)
1868                         LIBCFS_FREE(tx->tx_sge,
1869                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1870                                     sizeof(*tx->tx_sge));
1871                 if (tx->tx_rd != NULL)
1872                         LIBCFS_FREE(tx->tx_rd,
1873                                     offsetof(kib_rdma_desc_t,
1874                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1875         }
1876
1877         LIBCFS_FREE(tpo->tpo_tx_descs,
1878                     pool->po_size * sizeof(kib_tx_t));
1879 out:
1880         kiblnd_fini_pool(pool);
1881         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1882 }
1883
1884 static int
1885 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1886 {
1887         int            i;
1888         int            npg;
1889         kib_pool_t    *pool;
1890         kib_tx_pool_t *tpo;
1891
1892         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1893         if (tpo == NULL) {
1894                 CERROR("Failed to allocate TX pool\n");
1895                 return -ENOMEM;
1896         }
1897
1898         pool = &tpo->tpo_pool;
1899         kiblnd_init_pool(ps, pool, size);
1900         tpo->tpo_tx_descs = NULL;
1901         tpo->tpo_tx_pages = NULL;
1902
1903         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1904         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1905                 CERROR("Can't allocate tx pages: %d\n", npg);
1906                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1907                 return -ENOMEM;
1908         }
1909
1910         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1911         if (tpo->tpo_tx_descs == NULL) {
1912                 CERROR("Can't allocate %d tx descriptors\n", size);
1913                 ps->ps_pool_destroy(pool);
1914                 return -ENOMEM;
1915         }
1916
1917         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1918
1919         for (i = 0; i < size; i++) {
1920                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1921
1922                 tx->tx_pool = tpo;
1923                 if (ps->ps_net->ibn_with_fmr){
1924                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1925                                      sizeof(*tx->tx_pages));
1926                         if (tx->tx_pages == NULL)
1927                                 break;
1928                 }
1929
1930                 LIBCFS_ALLOC(tx->tx_frags,
1931                              IBLND_MAX_RDMA_FRAGS *
1932                              sizeof(*tx->tx_frags));
1933                 if (tx->tx_frags == NULL)
1934                         break;
1935
1936                 LIBCFS_ALLOC(tx->tx_wrq,
1937                              (1 + IBLND_MAX_RDMA_FRAGS) *
1938                              sizeof(*tx->tx_wrq));
1939                 if (tx->tx_wrq == NULL)
1940                         break;
1941
1942                 LIBCFS_ALLOC(tx->tx_sge,
1943                              (1 + IBLND_MAX_RDMA_FRAGS) *
1944                              sizeof(*tx->tx_sge));
1945                 if (tx->tx_sge == NULL)
1946                         break;
1947
1948                 LIBCFS_ALLOC(tx->tx_rd,
1949                              offsetof(kib_rdma_desc_t,
1950                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1951                 if (tx->tx_rd == NULL)
1952                         break;
1953         }
1954
1955         if (i == size) {
1956                 kiblnd_map_tx_pool(tpo);
1957                 *pp_po = pool;
1958                 return 0;
1959         }
1960
1961         ps->ps_pool_destroy(pool);
1962         return -ENOMEM;
1963 }
1964
1965 static void
1966 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1967 {
1968         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, tps_poolset);
1969         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
1970
1971         tx->tx_cookie = tps->tps_next_tx_cookie ++;
1972 }
1973
1974 void
1975 kiblnd_ni_fini_pools(kib_net_t *net)
1976 {
1977         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1978         if (net->ibn_with_fmr)
1979                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1980         else if (net->ibn_with_pmr)
1981                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1982 }
1983
1984 int
1985 kiblnd_net_init_pools(kib_net_t *net)
1986 {
1987         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1988         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1989         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
1990         int                rc;
1991
1992         if (*kiblnd_tunables.kib_fmr_pool_size <
1993             *kiblnd_tunables.kib_ntx / 4) {
1994                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1995                        *kiblnd_tunables.kib_fmr_pool_size,
1996                        *kiblnd_tunables.kib_ntx / 4);
1997                 return -EINVAL;
1998         }
1999
2000         if (*kiblnd_tunables.kib_pmr_pool_size <
2001             *kiblnd_tunables.kib_ntx / 4) {
2002                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2003                        *kiblnd_tunables.kib_pmr_pool_size,
2004                        *kiblnd_tunables.kib_ntx / 4);
2005                 return -EINVAL;
2006         }
2007
2008         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2009             net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2010                                            * so we always create FMR/PMR pool and
2011                                            * map-on-demand if premapping failed */
2012                 rc = kiblnd_init_fmr_pool_set(fps, net);
2013                 if (rc == 0) {
2014                         net->ibn_with_fmr = 1;
2015                 } else if (rc == -ENOSYS) {
2016                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2017                                                   *kiblnd_tunables.kib_pmr_pool_size,
2018                                                   kiblnd_create_pmr_pool,
2019                                                   kiblnd_destroy_pmr_pool,
2020                                                   NULL, NULL);
2021                         if (rc == 0)
2022                                 net->ibn_with_pmr = 1;
2023                 }
2024                 if (rc != 0)
2025                         return rc;
2026         }
2027
2028         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2029                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2030                                   kiblnd_tx_init, NULL);
2031         if (rc == 0)
2032                 return 0;
2033
2034         if (net->ibn_with_fmr)
2035                 kiblnd_fini_fmr_pool_set(fps);
2036         else if (net->ibn_with_pmr)
2037                 kiblnd_fini_pool_set(&pps->pps_poolset);
2038
2039         return rc;
2040 }
2041
2042 void
2043 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2044 {
2045         int     i;
2046
2047         if (ibdev->ibd_mrs == NULL)
2048                 return;
2049
2050         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2051                 if (ibdev->ibd_mrs[i] == NULL)
2052                         break;
2053
2054                 ib_dereg_mr(ibdev->ibd_mrs[i]);
2055         }
2056
2057         LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2058         ibdev->ibd_mrs = NULL;
2059 }
2060
2061 static int
2062 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2063 {
2064         struct ib_device_attr *attr;
2065         int                    rc;
2066
2067         /* It's safe to assume a HCA can handle a page size
2068          * matching that of the native system */
2069         ibdev->ibd_page_shift = PAGE_SHIFT;
2070         ibdev->ibd_page_size  = 1 << PAGE_SHIFT;
2071         ibdev->ibd_page_mask  = ~((__u64)ibdev->ibd_page_size - 1);
2072
2073         LIBCFS_ALLOC(attr, sizeof(*attr));
2074         if (attr == NULL) {
2075                 CERROR("Out of memory\n");
2076                 return -ENOMEM;
2077         }
2078
2079         rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2080         if (rc == 0)
2081                 ibdev->ibd_mr_size = attr->max_mr_size;
2082
2083         LIBCFS_FREE(attr, sizeof(*attr));
2084
2085         if (rc != 0) {
2086                 CERROR("Failed to query IB device: %d\n", rc);
2087                 return rc;
2088         }
2089
2090 #ifdef HAVE_OFED_TRANSPORT_IWARP
2091         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2092          * and we have reported the bug, remove these in the future when Chelsio
2093          * bug got fixed. */
2094         if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2095             RDMA_TRANSPORT_IWARP)
2096                 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2097 #endif
2098
2099         if (ibdev->ibd_mr_size == ~0ULL) {
2100                 ibdev->ibd_mr_shift = 64;
2101                 return 0;
2102         }
2103
2104         for (ibdev->ibd_mr_shift = 0;
2105              ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2106                 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2107                     ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2108                         return 0;
2109         }
2110
2111         CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2112         return -EINVAL;
2113 }
2114
2115 int
2116 kiblnd_dev_setup(kib_dev_t *ibdev)
2117 {
2118         struct ib_mr *mr;
2119         int           i;
2120         int           rc;
2121         __u64         mm_size;
2122         __u64         mr_size;
2123         int           acflags = IB_ACCESS_LOCAL_WRITE |
2124                                 IB_ACCESS_REMOTE_WRITE;
2125
2126         rc = kiblnd_dev_get_attr(ibdev);
2127         if (rc != 0)
2128                 return rc;
2129
2130         if (ibdev->ibd_mr_shift == 64) {
2131                 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2132                 if (ibdev->ibd_mrs == NULL) {
2133                         CERROR("Failed to allocate MRs table\n");
2134                         return -ENOMEM;
2135                 }
2136
2137                 ibdev->ibd_mrs[0] = NULL;
2138                 ibdev->ibd_nmrs   = 1;
2139
2140                 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2141                 if (IS_ERR(mr)) {
2142                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2143                         kiblnd_dev_cleanup(ibdev);
2144                         return PTR_ERR(mr);
2145                 }
2146
2147                 ibdev->ibd_mrs[0] = mr;
2148
2149                 goto out;
2150         }
2151
2152         mr_size = (1ULL << ibdev->ibd_mr_shift);
2153         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2154
2155         ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2156
2157         if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2158                 /* it's 4T..., assume we will re-code at that time */
2159                 CERROR("Can't support memory size: x"LPX64
2160                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2161                 return -EINVAL;
2162         }
2163
2164         /* create an array of MRs to cover all memory */
2165         LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2166         if (ibdev->ibd_mrs == NULL) {
2167                 CERROR("Failed to allocate MRs' table\n");
2168                 return -ENOMEM;
2169         }
2170
2171         memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2172
2173         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2174                 struct ib_phys_buf ipb;
2175                 __u64              iova;
2176
2177                 ipb.size = ibdev->ibd_mr_size;
2178                 ipb.addr = i * mr_size;
2179                 iova     = ipb.addr;
2180
2181                 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2182                 if (IS_ERR(mr)) {
2183                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2184                                " size "LPX64" : %ld\n",
2185                                ipb.addr, ipb.size, PTR_ERR(mr));
2186                         kiblnd_dev_cleanup(ibdev);
2187                         return PTR_ERR(mr);
2188                 }
2189
2190                 LASSERT (iova == ipb.addr);
2191
2192                 ibdev->ibd_mrs[i] = mr;
2193         }
2194
2195 out:
2196         CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2197                           LPX64", array size: %d\n",
2198                           ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2199
2200         list_add_tail(&ibdev->ibd_list,
2201                       &kiblnd_data.kib_devs);
2202         return 0;
2203 }
2204
2205 void
2206 kiblnd_destroy_dev (kib_dev_t *dev)
2207 {
2208         LASSERT (dev->ibd_nnets == 0);
2209
2210         if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
2211                 list_del_init(&dev->ibd_list);
2212
2213         kiblnd_dev_cleanup(dev);
2214
2215         if (dev->ibd_pd != NULL)
2216                 ib_dealloc_pd(dev->ibd_pd);
2217
2218         if (dev->ibd_cmid != NULL)
2219                 rdma_destroy_id(dev->ibd_cmid);
2220
2221         LIBCFS_FREE(dev, sizeof(*dev));
2222 }
2223
2224 void
2225 kiblnd_base_shutdown (void)
2226 {
2227         int i;
2228
2229         LASSERT (list_empty(&kiblnd_data.kib_devs));
2230
2231         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2232                atomic_read(&libcfs_kmemory));
2233
2234         switch (kiblnd_data.kib_init) {
2235         default:
2236                 LBUG();
2237
2238         case IBLND_INIT_ALL:
2239         case IBLND_INIT_DATA:
2240                 LASSERT (kiblnd_data.kib_peers != NULL);
2241                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2242                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2243                 }
2244                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2245                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2246
2247                 /* flag threads to terminate; wake and wait for them to die */
2248                 kiblnd_data.kib_shutdown = 1;
2249                 wake_up_all(&kiblnd_data.kib_sched_waitq);
2250                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2251
2252                 i = 2;
2253                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2254                         i++;
2255                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2256                                "Waiting for %d threads to terminate\n",
2257                                atomic_read(&kiblnd_data.kib_nthreads));
2258                         cfs_pause(cfs_time_seconds(1));
2259                 }
2260
2261                 /* fall through */
2262
2263         case IBLND_INIT_NOTHING:
2264                 break;
2265         }
2266
2267         if (kiblnd_data.kib_peers != NULL)
2268                 LIBCFS_FREE(kiblnd_data.kib_peers,
2269                             sizeof(struct list_head) *
2270                             kiblnd_data.kib_peer_hash_size);
2271
2272         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2273                atomic_read(&libcfs_kmemory));
2274
2275         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2276         PORTAL_MODULE_UNUSE;
2277 }
2278
2279 void
2280 kiblnd_shutdown (lnet_ni_t *ni)
2281 {
2282         kib_net_t        *net = ni->ni_data;
2283         rwlock_t         *g_lock = &kiblnd_data.kib_global_lock;
2284         int               i;
2285         unsigned long     flags;
2286
2287         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2288
2289         if (net == NULL)
2290                 goto out;
2291
2292         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2293                atomic_read(&libcfs_kmemory));
2294
2295         write_lock_irqsave(g_lock, flags);
2296         net->ibn_shutdown = 1;
2297         write_unlock_irqrestore(g_lock, flags);
2298
2299         switch (net->ibn_init) {
2300         default:
2301                 LBUG();
2302
2303         case IBLND_INIT_ALL:
2304                 /* nuke all existing peers within this net */
2305                 kiblnd_del_peer(ni, LNET_NID_ANY);
2306
2307                 /* Wait for all peer state to clean up */
2308                 i = 2;
2309                 while (atomic_read(&net->ibn_npeers) != 0) {
2310                         i++;
2311                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2312                                "%s: waiting for %d peers to disconnect\n",
2313                                libcfs_nid2str(ni->ni_nid),
2314                                atomic_read(&net->ibn_npeers));
2315                         cfs_pause(cfs_time_seconds(1));
2316                 }
2317
2318                 kiblnd_ni_fini_pools(net);
2319
2320                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2321                 net->ibn_dev->ibd_nnets--;
2322
2323                 /* fall through */
2324
2325         case IBLND_INIT_NOTHING:
2326                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2327
2328                 if (net->ibn_dev != NULL &&
2329                     net->ibn_dev->ibd_nnets == 0)
2330                         kiblnd_destroy_dev(net->ibn_dev);
2331
2332                 break;
2333         }
2334
2335         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2336                atomic_read(&libcfs_kmemory));
2337
2338         net->ibn_init = IBLND_INIT_NOTHING;
2339         ni->ni_data = NULL;
2340
2341         LIBCFS_FREE(net, sizeof(*net));
2342
2343 out:
2344         if (list_empty(&kiblnd_data.kib_devs))
2345                 kiblnd_base_shutdown();
2346         return;
2347 }
2348
2349 int
2350 kiblnd_base_startup (void)
2351 {
2352         int i;
2353         int rc;
2354
2355         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2356
2357         PORTAL_MODULE_USE;
2358         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2359
2360         rwlock_init(&kiblnd_data.kib_global_lock);
2361
2362         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2363
2364         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2365         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2366                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2367         if (kiblnd_data.kib_peers == NULL) {
2368                 goto failed;
2369         }
2370         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2371                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2372
2373         spin_lock_init(&kiblnd_data.kib_connd_lock);
2374         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2375         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2376         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2377
2378         spin_lock_init(&kiblnd_data.kib_sched_lock);
2379         INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2380         init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
2381
2382         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2383
2384         /* lists/ptrs/locks initialised */
2385         kiblnd_data.kib_init = IBLND_INIT_DATA;
2386         /*****************************************************/
2387
2388         for (i = 0; i < IBLND_N_SCHED; i++) {
2389                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2390                 if (rc != 0) {
2391                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2392                                i, rc);
2393                         goto failed;
2394                 }
2395         }
2396
2397         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2398         if (rc != 0) {
2399                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2400                 goto failed;
2401         }
2402
2403         /* flag everything initialised */
2404         kiblnd_data.kib_init = IBLND_INIT_ALL;
2405         /*****************************************************/
2406
2407         return 0;
2408
2409  failed:
2410         kiblnd_base_shutdown();
2411         return -ENETDOWN;
2412 }
2413
2414 int
2415 kiblnd_startup (lnet_ni_t *ni)
2416 {
2417         char                     *ifname;
2418         kib_dev_t                *ibdev = NULL;
2419         kib_net_t                *net;
2420         struct list_head         *tmp;
2421         struct timeval            tv;
2422         int                       rc;
2423
2424         LASSERT (ni->ni_lnd == &the_o2iblnd);
2425
2426         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2427                 rc = kiblnd_base_startup();
2428                 if (rc != 0)
2429                         return rc;
2430         }
2431
2432         LIBCFS_ALLOC(net, sizeof(*net));
2433         ni->ni_data = net;
2434         if (net == NULL)
2435                 goto failed;
2436
2437         memset(net, 0, sizeof(*net));
2438
2439         do_gettimeofday(&tv);
2440         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2441
2442         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2443         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2444         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2445         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2446
2447         if (ni->ni_interfaces[0] != NULL) {
2448                 /* Use the IPoIB interface specified in 'networks=' */
2449
2450                 CLASSERT (LNET_MAX_INTERFACES > 1);
2451                 if (ni->ni_interfaces[1] != NULL) {
2452                         CERROR("Multiple interfaces not supported\n");
2453                         goto failed;
2454                 }
2455
2456                 ifname = ni->ni_interfaces[0];
2457         } else {
2458                 ifname = *kiblnd_tunables.kib_default_ipif;
2459         }
2460
2461         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2462                 CERROR("IPoIB interface name too long: %s\n", ifname);
2463                 goto failed;
2464         }
2465
2466         list_for_each (tmp, &kiblnd_data.kib_devs) {
2467                 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
2468
2469                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2470                         break;
2471
2472                 ibdev = NULL;
2473         }
2474
2475         if (ibdev == NULL) {
2476                 __u32                     ip;
2477                 __u32                     netmask;
2478                 int                       up;
2479                 struct rdma_cm_id        *id;
2480                 struct ib_pd             *pd;
2481                 struct sockaddr_in        addr;
2482
2483                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2484                 if (rc != 0) {
2485                         CERROR("Can't query IPoIB interface %s: %d\n",
2486                                ifname, rc);
2487                         goto failed;
2488                 }
2489
2490                 if (!up) {
2491                         CERROR("Can't query IPoIB interface %s: it's down\n",
2492                                ifname);
2493                         goto failed;
2494                 }
2495
2496                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2497                 if (ibdev == NULL)
2498                         goto failed;
2499
2500                 memset(ibdev, 0, sizeof(*ibdev));
2501
2502                 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2503                 ibdev->ibd_ifip = ip;
2504                 strcpy(&ibdev->ibd_ifname[0], ifname);
2505
2506                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2507                 if (IS_ERR(id)) {
2508                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2509                         goto failed;
2510                 }
2511
2512                 ibdev->ibd_cmid = id;
2513
2514                 memset(&addr, 0, sizeof(addr));
2515                 addr.sin_family      = AF_INET;
2516                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2517                 addr.sin_addr.s_addr = htonl(ip);
2518
2519                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2520                 if (rc != 0) {
2521                         CERROR("Can't bind to %s: %d\n", ifname, rc);
2522                         goto failed;
2523                 }
2524
2525                 /* Binding should have assigned me an IB device */
2526                 LASSERT (id->device != NULL);
2527                 CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n",
2528                        ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service,
2529                        id->device->name);
2530
2531                 pd = ib_alloc_pd(id->device);
2532                 if (IS_ERR(pd)) {
2533                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2534                         goto failed;
2535                 }
2536
2537                 ibdev->ibd_pd = pd;
2538
2539                 rc = rdma_listen(id, 256);
2540                 if (rc != 0) {
2541                         CERROR("Can't start listener: %d\n", rc);
2542                         goto failed;
2543                 }
2544
2545                 rc = kiblnd_dev_setup(ibdev);
2546                 if (rc != 0) {
2547                         CERROR("Can't setup device: %d\n", rc);
2548                         goto failed;
2549                 }
2550         }
2551
2552         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2553         net->ibn_dev = ibdev;
2554
2555         rc = kiblnd_net_init_pools(net);
2556         if (rc != 0) {
2557                 CERROR("Failed to initialize NI pools: %d\n", rc);
2558                 goto failed;
2559         }
2560         ibdev->ibd_nnets++;
2561         net->ibn_init = IBLND_INIT_ALL;
2562
2563         return 0;
2564
2565 failed:
2566         if (net->ibn_dev == NULL && ibdev != NULL)
2567                 kiblnd_destroy_dev(ibdev);
2568
2569         kiblnd_shutdown(ni);
2570
2571         CDEBUG(D_NET, "kiblnd_startup failed\n");
2572         return -ENETDOWN;
2573 }
2574
2575 void __exit
2576 kiblnd_module_fini (void)
2577 {
2578         lnet_unregister_lnd(&the_o2iblnd);
2579         kiblnd_tunables_fini();
2580 }
2581
2582 int __init
2583 kiblnd_module_init (void)
2584 {
2585         int    rc;
2586
2587         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2588         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2589                   <= IBLND_MSG_SIZE);
2590         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2591                   <= IBLND_MSG_SIZE);
2592
2593         rc = kiblnd_tunables_init();
2594         if (rc != 0)
2595                 return rc;
2596
2597         lnet_register_lnd(&the_o2iblnd);
2598
2599         return 0;
2600 }
2601
2602 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2603 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2604 MODULE_LICENSE("GPL");
2605
2606 module_init(kiblnd_module_init);
2607 module_exit(kiblnd_module_fini);