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, cfs_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 = last_alive;
1099
1100         /* peer is not persistent in hash, trigger peer creation
1101          * and connection establishment with a NULL tx */
1102         if (peer == NULL)
1103                 kiblnd_launch_tx(ni, NULL, nid);
1104         return;
1105 }
1106
1107 void
1108 kiblnd_free_pages (kib_pages_t *p)
1109 {
1110         int         npages = p->ibp_npages;
1111         int         i;
1112
1113         LASSERT (p->ibp_device == NULL);
1114
1115         for (i = 0; i < npages; i++)
1116                 if (p->ibp_pages[i] != NULL)
1117                         __free_page(p->ibp_pages[i]);
1118
1119         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1120 }
1121
1122 int
1123 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1124 {
1125         kib_pages_t   *p;
1126         int            i;
1127
1128         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1129         if (p == NULL) {
1130                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1131                 return -ENOMEM;
1132         }
1133
1134         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1135         p->ibp_npages = npages;
1136
1137         for (i = 0; i < npages; i++) {
1138                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1139                 if (p->ibp_pages[i] == NULL) {
1140                         CERROR("Can't allocate page %d of %d\n", i, npages);
1141                         kiblnd_free_pages(p);
1142                         return -ENOMEM;
1143                 }
1144         }
1145
1146         *pp = p;
1147         return 0;
1148 }
1149
1150 void
1151 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1152 {
1153         kib_rx_t *rx;
1154         int       i;
1155
1156         LASSERT (conn->ibc_rxs != NULL);
1157         LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1158
1159         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1160                 rx = &conn->ibc_rxs[i];
1161
1162                 LASSERT (rx->rx_nob >= 0); /* not posted */
1163
1164                 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1165                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1166                                                           rx->rx_msgaddr),
1167                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1168         }
1169
1170         conn->ibc_rx_pages->ibp_device = NULL;
1171
1172         kiblnd_free_pages(conn->ibc_rx_pages);
1173
1174         conn->ibc_rx_pages = NULL;
1175 }
1176
1177 void
1178 kiblnd_map_rx_descs(kib_conn_t *conn)
1179 {
1180         kib_rx_t       *rx;
1181         struct page    *pg;
1182         int             pg_off;
1183         int             ipg;
1184         int             i;
1185
1186         for (pg_off = ipg = i = 0;
1187              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1188                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1189                 rx = &conn->ibc_rxs[i];
1190
1191                 rx->rx_conn = conn;
1192                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1193
1194                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1195                                                        rx->rx_msg, IBLND_MSG_SIZE,
1196                                                        DMA_FROM_DEVICE);
1197                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1198                                                    rx->rx_msgaddr));
1199                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1200
1201                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1202                        i, rx->rx_msg, rx->rx_msgaddr,
1203                        lnet_page2phys(pg) + pg_off);
1204
1205                 pg_off += IBLND_MSG_SIZE;
1206                 LASSERT (pg_off <= PAGE_SIZE);
1207
1208                 if (pg_off == PAGE_SIZE) {
1209                         pg_off = 0;
1210                         ipg++;
1211                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1212                 }
1213         }
1214
1215         conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1216 }
1217
1218 static void
1219 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1220 {
1221         kib_tx_t       *tx;
1222         int             i;
1223
1224         LASSERT (tpo->tpo_pool.po_allocated == 0);
1225         LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1226
1227         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1228                 tx = &tpo->tpo_tx_descs[i];
1229                 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1230                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1231                                                           tx->tx_msgaddr),
1232                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1233         }
1234         tpo->tpo_tx_pages->ibp_device = NULL;
1235 }
1236
1237 static void
1238 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1239 {
1240         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1241         kib_pool_t     *pool  = &tpo->tpo_pool;
1242         kib_net_t      *net   = pool->po_owner->ps_net;
1243         struct page    *page;
1244         kib_tx_t       *tx;
1245         int             page_offset;
1246         int             ipage;
1247         int             i;
1248
1249         LASSERT (net != NULL);
1250         LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1251                  net->ibn_dev->ibd_cmid->device != NULL);
1252
1253         /* pre-mapped messages are not bigger than 1 page */
1254         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1255
1256         /* No fancy arithmetic when we do the buffer calculations */
1257         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1258
1259         txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1260
1261         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1262                 page = txpgs->ibp_pages[ipage];
1263                 tx = &tpo->tpo_tx_descs[i];
1264
1265                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1266                                            page_offset);
1267
1268                 tx->tx_msgaddr = kiblnd_dma_map_single(
1269                         txpgs->ibp_device, tx->tx_msg,
1270                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1271                 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1272                                                    tx->tx_msgaddr));
1273                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1274
1275                 list_add(&tx->tx_list, &pool->po_free_list);
1276
1277                 page_offset += IBLND_MSG_SIZE;
1278                 LASSERT (page_offset <= PAGE_SIZE);
1279
1280                 if (page_offset == PAGE_SIZE) {
1281                         page_offset = 0;
1282                         ipage++;
1283                         LASSERT (ipage <= txpgs->ibp_npages);
1284                 }
1285         }
1286 }
1287
1288 struct ib_mr *
1289 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1290 {
1291         __u64   index;
1292
1293         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1294
1295         if (net->ibn_dev->ibd_nmrs == 1)
1296                 return net->ibn_dev->ibd_mrs[0];
1297
1298         index = addr >> net->ibn_dev->ibd_mr_shift;
1299
1300         if (index <  net->ibn_dev->ibd_nmrs &&
1301             index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1302                 return net->ibn_dev->ibd_mrs[index];
1303
1304         return NULL;
1305 }
1306
1307 struct ib_mr *
1308 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1309 {
1310         struct ib_mr *prev_mr;
1311         struct ib_mr *mr;
1312         int           i;
1313
1314         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1315
1316         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1317             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1318                 return NULL;
1319
1320         if (net->ibn_dev->ibd_nmrs == 1)
1321                 return net->ibn_dev->ibd_mrs[0];
1322
1323         for (i = 0, mr = prev_mr = NULL;
1324              i < rd->rd_nfrags; i++) {
1325                 mr = kiblnd_find_dma_mr(net,
1326                                         rd->rd_frags[i].rf_addr,
1327                                         rd->rd_frags[i].rf_nob);
1328                 if (prev_mr == NULL)
1329                         prev_mr = mr;
1330
1331                 if (mr == NULL || prev_mr != mr) {
1332                         /* Can't covered by one single MR */
1333                         mr = NULL;
1334                         break;
1335                 }
1336         }
1337
1338         return mr;
1339 }
1340
1341 void
1342 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1343 {
1344         LASSERT (pool->fpo_map_count == 0);
1345
1346         if (pool->fpo_fmr_pool != NULL)
1347                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1348
1349         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1350 }
1351
1352 void
1353 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1354 {
1355         kib_fmr_pool_t *pool;
1356
1357         while (!list_empty(head)) {
1358                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1359                 list_del(&pool->fpo_list);
1360                 kiblnd_destroy_fmr_pool(pool);
1361         }
1362 }
1363
1364 int
1365 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1366 {
1367         /* FMR pool for RDMA */
1368         kib_fmr_pool_t          *fpo;
1369         struct ib_fmr_pool_param param = {
1370                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1371                 .page_shift        = PAGE_SHIFT,
1372                 .access            = (IB_ACCESS_LOCAL_WRITE |
1373                                       IB_ACCESS_REMOTE_WRITE),
1374                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1375                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1376                 .flush_function    = NULL,
1377                 .flush_arg         = NULL,
1378                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1379         int rc;
1380
1381         LASSERT (fps->fps_net->ibn_dev != NULL &&
1382                  fps->fps_net->ibn_dev->ibd_pd != NULL);
1383
1384         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1385         if (fpo == NULL)
1386                 return -ENOMEM;
1387
1388         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1389         fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, &param);
1390         if (IS_ERR(fpo->fpo_fmr_pool)) {
1391                 CERROR("Failed to create FMR pool: %ld\n",
1392                        PTR_ERR(fpo->fpo_fmr_pool));
1393                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1394                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1395                 return rc;
1396         }
1397
1398         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1399         fpo->fpo_owner    = fps;
1400         *pp_fpo = fpo;
1401
1402         return 0;
1403 }
1404
1405 static void
1406 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1407 {
1408         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1409 }
1410
1411 static int
1412 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1413 {
1414         kib_fmr_pool_t *fpo;
1415         int             rc;
1416
1417         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1418
1419         fps->fps_net = net;
1420         spin_lock_init(&fps->fps_lock);
1421         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1422         rc = kiblnd_create_fmr_pool(fps, &fpo);
1423         if (rc == 0)
1424                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1425
1426         return rc;
1427 }
1428
1429 void
1430 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1431 {
1432         CFS_LIST_HEAD     (zombies);
1433         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1434         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1435         kib_fmr_pool_t    *tmp;
1436         int                rc;
1437
1438         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1439         LASSERT (rc == 0);
1440
1441         if (status != 0) {
1442                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1443                 LASSERT (rc == 0);
1444         }
1445
1446         fmr->fmr_pool = NULL;
1447         fmr->fmr_pfmr = NULL;
1448
1449         spin_lock(&fps->fps_lock);
1450         fpo->fpo_map_count --;  /* decref the pool */
1451
1452         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1453                 /* the first pool is persistent */
1454                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1455                         continue;
1456
1457                 if (fpo->fpo_map_count == 0 &&  /* no more reference */
1458                     cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1459                         list_move(&fpo->fpo_list, &zombies);
1460                         fps->fps_version ++;
1461                 }
1462         }
1463         spin_unlock(&fps->fps_lock);
1464
1465         if (!list_empty(&zombies))
1466                 kiblnd_destroy_fmr_pool_list(&zombies);
1467 }
1468
1469 int
1470 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1471                     __u64 iov, kib_fmr_t *fmr)
1472 {
1473         struct ib_pool_fmr *pfmr;
1474         kib_fmr_pool_t     *fpo;
1475         __u64               version;
1476         int                 rc;
1477
1478         LASSERT (fps->fps_net->ibn_with_fmr);
1479  again:
1480         spin_lock(&fps->fps_lock);
1481         version = fps->fps_version;
1482         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1483                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1484                 fpo->fpo_map_count ++;
1485                 spin_unlock(&fps->fps_lock);
1486
1487                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1488                                             pages, npages, iov);
1489                 if (likely(!IS_ERR(pfmr))) {
1490                         fmr->fmr_pool = fpo;
1491                         fmr->fmr_pfmr = pfmr;
1492                         return 0;
1493                 }
1494
1495                 spin_lock(&fps->fps_lock);
1496                 fpo->fpo_map_count --;
1497                 if (PTR_ERR(pfmr) != -EAGAIN) {
1498                         spin_unlock(&fps->fps_lock);
1499                         return PTR_ERR(pfmr);
1500                 }
1501
1502                 /* EAGAIN and ... */
1503                 if (version != fps->fps_version) {
1504                         spin_unlock(&fps->fps_lock);
1505                         goto again;
1506                 }
1507         }
1508
1509         if (fps->fps_increasing) {
1510                 spin_unlock(&fps->fps_lock);
1511                 CDEBUG(D_NET, "Another thread is allocating new "
1512                               "FMR pool, waiting for her to complete\n");
1513                 schedule();
1514                 goto again;
1515
1516         }
1517
1518         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1519                 /* someone failed recently */
1520                 spin_unlock(&fps->fps_lock);
1521                 return -EAGAIN;
1522         }
1523
1524         fps->fps_increasing = 1;
1525         spin_unlock(&fps->fps_lock);
1526
1527         CDEBUG(D_NET, "Allocate new FMR pool\n");
1528         rc = kiblnd_create_fmr_pool(fps, &fpo);
1529         spin_lock(&fps->fps_lock);
1530         fps->fps_increasing = 0;
1531         if (rc == 0) {
1532                 fps->fps_version ++;
1533                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1534         } else {
1535                 fps->fps_next_retry = cfs_time_shift(10);
1536         }
1537         spin_unlock(&fps->fps_lock);
1538
1539         goto again;
1540 }
1541
1542 static void
1543 kiblnd_fini_pool(kib_pool_t *pool)
1544 {
1545         LASSERT (list_empty(&pool->po_free_list));
1546         LASSERT (pool->po_allocated == 0);
1547
1548         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1549 }
1550
1551 static void
1552 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1553 {
1554         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1555
1556         memset(pool, 0, sizeof(kib_pool_t));
1557         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1558         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1559         pool->po_owner    = ps;
1560         pool->po_size     = size;
1561 }
1562
1563 void
1564 kiblnd_destroy_pool_list(kib_poolset_t *ps, struct list_head *head)
1565 {
1566         kib_pool_t *pool;
1567
1568         while (!list_empty(head)) {
1569                 pool = list_entry(head->next, kib_pool_t, po_list);
1570                 list_del(&pool->po_list);
1571                 ps->ps_pool_destroy(pool);
1572         }
1573 }
1574
1575 static void
1576 kiblnd_fini_pool_set(kib_poolset_t *ps)
1577 {
1578         kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1579 }
1580
1581 static int
1582 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1583                      char *name, int size,
1584                      kib_ps_pool_create_t po_create,
1585                      kib_ps_pool_destroy_t po_destroy,
1586                      kib_ps_node_init_t nd_init,
1587                      kib_ps_node_fini_t nd_fini)
1588 {
1589         kib_pool_t    *pool;
1590         int            rc;
1591
1592         memset(ps, 0, sizeof(kib_poolset_t));
1593
1594         ps->ps_net          = net;
1595         ps->ps_pool_create  = po_create;
1596         ps->ps_pool_destroy = po_destroy;
1597         ps->ps_node_init    = nd_init;
1598         ps->ps_node_fini    = nd_fini;
1599         ps->ps_pool_size    = size;
1600         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1601         spin_lock_init(&ps->ps_lock);
1602         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1603
1604         rc = ps->ps_pool_create(ps, size, &pool);
1605         if (rc == 0)
1606                 list_add(&pool->po_list, &ps->ps_pool_list);
1607         else
1608                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1609
1610         return rc;
1611 }
1612
1613 void
1614 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1615 {
1616         CFS_LIST_HEAD  (zombies);
1617         kib_poolset_t  *ps = pool->po_owner;
1618         kib_pool_t     *tmp;
1619         cfs_time_t      now = cfs_time_current();
1620
1621         spin_lock(&ps->ps_lock);
1622
1623         if (ps->ps_node_fini != NULL)
1624                 ps->ps_node_fini(pool, node);
1625
1626         LASSERT (pool->po_allocated > 0);
1627         list_add(node, &pool->po_free_list);
1628         pool->po_allocated --;
1629
1630         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1631                 /* the first pool is persistent */
1632                 if (ps->ps_pool_list.next == &pool->po_list)
1633                         continue;
1634
1635                 if (pool->po_allocated == 0 &&
1636                     cfs_time_aftereq(now, pool->po_deadline))
1637                         list_move(&pool->po_list, &zombies);
1638         }
1639         spin_unlock(&ps->ps_lock);
1640
1641         if (!list_empty(&zombies))
1642                 kiblnd_destroy_pool_list(ps, &zombies);
1643 }
1644
1645 struct list_head *
1646 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1647 {
1648         struct list_head  *node;
1649         kib_pool_t        *pool;
1650         int                rc;
1651
1652  again:
1653         spin_lock(&ps->ps_lock);
1654         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1655                 if (list_empty(&pool->po_free_list))
1656                         continue;
1657
1658                 pool->po_allocated ++;
1659                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1660                 node = pool->po_free_list.next;
1661                 list_del(node);
1662
1663                 if (ps->ps_node_init != NULL) {
1664                         /* still hold the lock */
1665                         ps->ps_node_init(pool, node);
1666                 }
1667                 spin_unlock(&ps->ps_lock);
1668                 return node;
1669         }
1670
1671         /* no available tx pool and ... */
1672         if (ps->ps_increasing) {
1673                 /* another thread is allocating a new pool */
1674                 spin_unlock(&ps->ps_lock);
1675                 CDEBUG(D_NET, "Another thread is allocating new "
1676                        "%s pool, waiting for her to complete\n",
1677                        ps->ps_name);
1678                 schedule();
1679                 goto again;
1680         }
1681
1682         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1683                 /* someone failed recently */
1684                 spin_unlock(&ps->ps_lock);
1685                 return NULL;
1686         }
1687
1688         ps->ps_increasing = 1;
1689         spin_unlock(&ps->ps_lock);
1690
1691         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1692
1693         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1694
1695         spin_lock(&ps->ps_lock);
1696         ps->ps_increasing = 0;
1697         if (rc == 0) {
1698                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1699         } else {
1700                 /* retry 10 seconds later */
1701                 ps->ps_next_retry = cfs_time_shift(10);
1702                 CERROR("Can't allocate new %s pool because out of memory\n",
1703                        ps->ps_name);
1704         }
1705         spin_unlock(&ps->ps_lock);
1706
1707         goto again;
1708 }
1709
1710 void
1711 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1712 {
1713         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1714         struct ib_mr        *mr  = pmr->pmr_mr;
1715
1716         pmr->pmr_mr = NULL;
1717         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1718         if (mr != NULL)
1719                 ib_dereg_mr(mr);
1720 }
1721
1722 int
1723 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1724                     __u64 *iova, kib_phys_mr_t **pp_pmr)
1725 {
1726         kib_phys_mr_t       *pmr;
1727         struct list_head    *node;
1728         int                  rc;
1729         int                  i;
1730
1731         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1732         if (node == NULL) {
1733                 CERROR("Failed to allocate PMR descriptor\n");
1734                 return -ENOMEM;
1735         }
1736
1737         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1738         for (i = 0; i < rd->rd_nfrags; i ++) {
1739                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1740                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1741         }
1742
1743         pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1744                                      pmr->pmr_ipb, rd->rd_nfrags,
1745                                      IB_ACCESS_LOCAL_WRITE |
1746                                      IB_ACCESS_REMOTE_WRITE,
1747                                      iova);
1748         if (!IS_ERR(pmr->pmr_mr)) {
1749                 pmr->pmr_iova = *iova;
1750                 *pp_pmr = pmr;
1751                 return 0;
1752         }
1753
1754         rc = PTR_ERR(pmr->pmr_mr);
1755         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1756
1757         pmr->pmr_mr = NULL;
1758         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1759
1760         return rc;
1761 }
1762
1763 static void
1764 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1765 {
1766         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1767         kib_phys_mr_t  *pmr;
1768
1769         LASSERT (pool->po_allocated == 0);
1770
1771         while (!list_empty(&pool->po_free_list)) {
1772                 pmr = list_entry(pool->po_free_list.next,
1773                                  kib_phys_mr_t, pmr_list);
1774
1775                 LASSERT (pmr->pmr_mr == NULL);
1776                 list_del(&pmr->pmr_list);
1777
1778                 if (pmr->pmr_ipb != NULL) {
1779                         LIBCFS_FREE(pmr->pmr_ipb,
1780                                     IBLND_MAX_RDMA_FRAGS *
1781                                     sizeof(struct ib_phys_buf));
1782                 }
1783
1784                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1785         }
1786
1787         kiblnd_fini_pool(pool);
1788         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1789 }
1790
1791 static int
1792 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1793 {
1794         kib_pmr_pool_t      *ppo;
1795         kib_pool_t          *pool;
1796         kib_phys_mr_t       *pmr;
1797         int                  i;
1798
1799         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1800         if (ppo == NULL) {
1801                 CERROR("Failed to allocate PMR pool\n");
1802                 return -ENOMEM;
1803         }
1804
1805         pool = &ppo->ppo_pool;
1806         kiblnd_init_pool(ps, pool, size);
1807
1808         for (i = 0; i < size; i++) {
1809                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1810                 if (pmr == NULL)
1811                         break;
1812
1813                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1814                 pmr->pmr_pool = ppo;
1815                 LIBCFS_ALLOC(pmr->pmr_ipb,
1816                              IBLND_MAX_RDMA_FRAGS *
1817                              sizeof(struct ib_phys_buf));
1818                 if (pmr->pmr_ipb == NULL)
1819                         break;
1820
1821                 list_add(&pmr->pmr_list, &pool->po_free_list);
1822         }
1823
1824         if (i < size) {
1825                 ps->ps_pool_destroy(pool);
1826                 return -ENOMEM;
1827         }
1828
1829         *pp_po = pool;
1830         return 0;
1831 }
1832
1833 static void
1834 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1835 {
1836         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1837         int             i;
1838
1839         LASSERT (pool->po_allocated == 0);
1840
1841         if (tpo->tpo_tx_pages != NULL) {
1842                 if (tpo->tpo_tx_pages->ibp_device != NULL)
1843                         kiblnd_unmap_tx_pool(tpo);
1844                 kiblnd_free_pages(tpo->tpo_tx_pages);
1845         }
1846
1847         if (tpo->tpo_tx_descs == NULL)
1848                 goto out;
1849
1850         for (i = 0; i < pool->po_size; i++) {
1851                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1852
1853                 list_del(&tx->tx_list);
1854                 if (tx->tx_pages != NULL)
1855                         LIBCFS_FREE(tx->tx_pages,
1856                                     LNET_MAX_IOV *
1857                                     sizeof(*tx->tx_pages));
1858                 if (tx->tx_frags != NULL)
1859                         LIBCFS_FREE(tx->tx_frags,
1860                                     IBLND_MAX_RDMA_FRAGS *
1861                                             sizeof(*tx->tx_frags));
1862                 if (tx->tx_wrq != NULL)
1863                         LIBCFS_FREE(tx->tx_wrq,
1864                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1865                                     sizeof(*tx->tx_wrq));
1866                 if (tx->tx_sge != NULL)
1867                         LIBCFS_FREE(tx->tx_sge,
1868                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1869                                     sizeof(*tx->tx_sge));
1870                 if (tx->tx_rd != NULL)
1871                         LIBCFS_FREE(tx->tx_rd,
1872                                     offsetof(kib_rdma_desc_t,
1873                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1874         }
1875
1876         LIBCFS_FREE(tpo->tpo_tx_descs,
1877                     pool->po_size * sizeof(kib_tx_t));
1878 out:
1879         kiblnd_fini_pool(pool);
1880         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1881 }
1882
1883 static int
1884 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1885 {
1886         int            i;
1887         int            npg;
1888         kib_pool_t    *pool;
1889         kib_tx_pool_t *tpo;
1890
1891         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1892         if (tpo == NULL) {
1893                 CERROR("Failed to allocate TX pool\n");
1894                 return -ENOMEM;
1895         }
1896
1897         pool = &tpo->tpo_pool;
1898         kiblnd_init_pool(ps, pool, size);
1899         tpo->tpo_tx_descs = NULL;
1900         tpo->tpo_tx_pages = NULL;
1901
1902         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1903         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1904                 CERROR("Can't allocate tx pages: %d\n", npg);
1905                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1906                 return -ENOMEM;
1907         }
1908
1909         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1910         if (tpo->tpo_tx_descs == NULL) {
1911                 CERROR("Can't allocate %d tx descriptors\n", size);
1912                 ps->ps_pool_destroy(pool);
1913                 return -ENOMEM;
1914         }
1915
1916         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1917
1918         for (i = 0; i < size; i++) {
1919                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1920
1921                 tx->tx_pool = tpo;
1922                 if (ps->ps_net->ibn_with_fmr){
1923                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1924                                      sizeof(*tx->tx_pages));
1925                         if (tx->tx_pages == NULL)
1926                                 break;
1927                 }
1928
1929                 LIBCFS_ALLOC(tx->tx_frags,
1930                              IBLND_MAX_RDMA_FRAGS *
1931                              sizeof(*tx->tx_frags));
1932                 if (tx->tx_frags == NULL)
1933                         break;
1934
1935                 LIBCFS_ALLOC(tx->tx_wrq,
1936                              (1 + IBLND_MAX_RDMA_FRAGS) *
1937                              sizeof(*tx->tx_wrq));
1938                 if (tx->tx_wrq == NULL)
1939                         break;
1940
1941                 LIBCFS_ALLOC(tx->tx_sge,
1942                              (1 + IBLND_MAX_RDMA_FRAGS) *
1943                              sizeof(*tx->tx_sge));
1944                 if (tx->tx_sge == NULL)
1945                         break;
1946
1947                 LIBCFS_ALLOC(tx->tx_rd,
1948                              offsetof(kib_rdma_desc_t,
1949                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1950                 if (tx->tx_rd == NULL)
1951                         break;
1952         }
1953
1954         if (i == size) {
1955                 kiblnd_map_tx_pool(tpo);
1956                 *pp_po = pool;
1957                 return 0;
1958         }
1959
1960         ps->ps_pool_destroy(pool);
1961         return -ENOMEM;
1962 }
1963
1964 static void
1965 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1966 {
1967         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t, tps_poolset);
1968         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
1969
1970         tx->tx_cookie = tps->tps_next_tx_cookie ++;
1971 }
1972
1973 void
1974 kiblnd_ni_fini_pools(kib_net_t *net)
1975 {
1976         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1977         if (net->ibn_with_fmr)
1978                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1979         else if (net->ibn_with_pmr)
1980                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1981 }
1982
1983 int
1984 kiblnd_net_init_pools(kib_net_t *net)
1985 {
1986         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1987         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1988         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
1989         int                rc;
1990
1991         if (*kiblnd_tunables.kib_fmr_pool_size <
1992             *kiblnd_tunables.kib_ntx / 4) {
1993                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1994                        *kiblnd_tunables.kib_fmr_pool_size,
1995                        *kiblnd_tunables.kib_ntx / 4);
1996                 return -EINVAL;
1997         }
1998
1999         if (*kiblnd_tunables.kib_pmr_pool_size <
2000             *kiblnd_tunables.kib_ntx / 4) {
2001                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2002                        *kiblnd_tunables.kib_pmr_pool_size,
2003                        *kiblnd_tunables.kib_ntx / 4);
2004                 return -EINVAL;
2005         }
2006
2007         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2008             net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2009                                            * so we always create FMR/PMR pool and
2010                                            * map-on-demand if premapping failed */
2011                 rc = kiblnd_init_fmr_pool_set(fps, net);
2012                 if (rc == 0) {
2013                         net->ibn_with_fmr = 1;
2014                 } else if (rc == -ENOSYS) {
2015                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2016                                                   *kiblnd_tunables.kib_pmr_pool_size,
2017                                                   kiblnd_create_pmr_pool,
2018                                                   kiblnd_destroy_pmr_pool,
2019                                                   NULL, NULL);
2020                         if (rc == 0)
2021                                 net->ibn_with_pmr = 1;
2022                 }
2023                 if (rc != 0)
2024                         return rc;
2025         }
2026
2027         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2028                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2029                                   kiblnd_tx_init, NULL);
2030         if (rc == 0)
2031                 return 0;
2032
2033         if (net->ibn_with_fmr)
2034                 kiblnd_fini_fmr_pool_set(fps);
2035         else if (net->ibn_with_pmr)
2036                 kiblnd_fini_pool_set(&pps->pps_poolset);
2037
2038         return rc;
2039 }
2040
2041 void
2042 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2043 {
2044         int     i;
2045
2046         if (ibdev->ibd_mrs == NULL)
2047                 return;
2048
2049         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2050                 if (ibdev->ibd_mrs[i] == NULL)
2051                         break;
2052
2053                 ib_dereg_mr(ibdev->ibd_mrs[i]);
2054         }
2055
2056         LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2057         ibdev->ibd_mrs = NULL;
2058 }
2059
2060 static int
2061 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2062 {
2063         struct ib_device_attr *attr;
2064         int                    rc;
2065
2066         /* It's safe to assume a HCA can handle a page size
2067          * matching that of the native system */
2068         ibdev->ibd_page_shift = PAGE_SHIFT;
2069         ibdev->ibd_page_size  = 1 << PAGE_SHIFT;
2070         ibdev->ibd_page_mask  = ~((__u64)ibdev->ibd_page_size - 1);
2071
2072         LIBCFS_ALLOC(attr, sizeof(*attr));
2073         if (attr == NULL) {
2074                 CERROR("Out of memory\n");
2075                 return -ENOMEM;
2076         }
2077
2078         rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2079         if (rc == 0)
2080                 ibdev->ibd_mr_size = attr->max_mr_size;
2081
2082         LIBCFS_FREE(attr, sizeof(*attr));
2083
2084         if (rc != 0) {
2085                 CERROR("Failed to query IB device: %d\n", rc);
2086                 return rc;
2087         }
2088
2089 #ifdef HAVE_OFED_TRANSPORT_IWARP
2090         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2091          * and we have reported the bug, remove these in the future when Chelsio
2092          * bug got fixed. */
2093         if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2094             RDMA_TRANSPORT_IWARP)
2095                 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2096 #endif
2097
2098         if (ibdev->ibd_mr_size == ~0ULL) {
2099                 ibdev->ibd_mr_shift = 64;
2100                 return 0;
2101         }
2102
2103         for (ibdev->ibd_mr_shift = 0;
2104              ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2105                 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2106                     ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2107                         return 0;
2108         }
2109
2110         CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2111         return -EINVAL;
2112 }
2113
2114 int
2115 kiblnd_dev_setup(kib_dev_t *ibdev)
2116 {
2117         struct ib_mr *mr;
2118         int           i;
2119         int           rc;
2120         __u64         mm_size;
2121         __u64         mr_size;
2122         int           acflags = IB_ACCESS_LOCAL_WRITE |
2123                                 IB_ACCESS_REMOTE_WRITE;
2124
2125         rc = kiblnd_dev_get_attr(ibdev);
2126         if (rc != 0)
2127                 return rc;
2128
2129         if (ibdev->ibd_mr_shift == 64) {
2130                 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2131                 if (ibdev->ibd_mrs == NULL) {
2132                         CERROR("Failed to allocate MRs table\n");
2133                         return -ENOMEM;
2134                 }
2135
2136                 ibdev->ibd_mrs[0] = NULL;
2137                 ibdev->ibd_nmrs   = 1;
2138
2139                 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2140                 if (IS_ERR(mr)) {
2141                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2142                         kiblnd_dev_cleanup(ibdev);
2143                         return PTR_ERR(mr);
2144                 }
2145
2146                 ibdev->ibd_mrs[0] = mr;
2147
2148                 goto out;
2149         }
2150
2151         mr_size = (1ULL << ibdev->ibd_mr_shift);
2152         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2153
2154         ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2155
2156         if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2157                 /* it's 4T..., assume we will re-code at that time */
2158                 CERROR("Can't support memory size: x"LPX64
2159                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2160                 return -EINVAL;
2161         }
2162
2163         /* create an array of MRs to cover all memory */
2164         LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2165         if (ibdev->ibd_mrs == NULL) {
2166                 CERROR("Failed to allocate MRs' table\n");
2167                 return -ENOMEM;
2168         }
2169
2170         memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2171
2172         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2173                 struct ib_phys_buf ipb;
2174                 __u64              iova;
2175
2176                 ipb.size = ibdev->ibd_mr_size;
2177                 ipb.addr = i * mr_size;
2178                 iova     = ipb.addr;
2179
2180                 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2181                 if (IS_ERR(mr)) {
2182                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2183                                " size "LPX64" : %ld\n",
2184                                ipb.addr, ipb.size, PTR_ERR(mr));
2185                         kiblnd_dev_cleanup(ibdev);
2186                         return PTR_ERR(mr);
2187                 }
2188
2189                 LASSERT (iova == ipb.addr);
2190
2191                 ibdev->ibd_mrs[i] = mr;
2192         }
2193
2194 out:
2195         CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2196                           LPX64", array size: %d\n",
2197                           ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2198
2199         list_add_tail(&ibdev->ibd_list,
2200                       &kiblnd_data.kib_devs);
2201         return 0;
2202 }
2203
2204 void
2205 kiblnd_destroy_dev (kib_dev_t *dev)
2206 {
2207         LASSERT (dev->ibd_nnets == 0);
2208
2209         if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
2210                 list_del_init(&dev->ibd_list);
2211
2212         kiblnd_dev_cleanup(dev);
2213
2214         if (dev->ibd_pd != NULL)
2215                 ib_dealloc_pd(dev->ibd_pd);
2216
2217         if (dev->ibd_cmid != NULL)
2218                 rdma_destroy_id(dev->ibd_cmid);
2219
2220         LIBCFS_FREE(dev, sizeof(*dev));
2221 }
2222
2223 void
2224 kiblnd_base_shutdown (void)
2225 {
2226         int i;
2227
2228         LASSERT (list_empty(&kiblnd_data.kib_devs));
2229
2230         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2231                atomic_read(&libcfs_kmemory));
2232
2233         switch (kiblnd_data.kib_init) {
2234         default:
2235                 LBUG();
2236
2237         case IBLND_INIT_ALL:
2238         case IBLND_INIT_DATA:
2239                 LASSERT (kiblnd_data.kib_peers != NULL);
2240                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2241                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2242                 }
2243                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2244                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2245
2246                 /* flag threads to terminate; wake and wait for them to die */
2247                 kiblnd_data.kib_shutdown = 1;
2248                 wake_up_all(&kiblnd_data.kib_sched_waitq);
2249                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2250
2251                 i = 2;
2252                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2253                         i++;
2254                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2255                                "Waiting for %d threads to terminate\n",
2256                                atomic_read(&kiblnd_data.kib_nthreads));
2257                         cfs_pause(cfs_time_seconds(1));
2258                 }
2259
2260                 /* fall through */
2261
2262         case IBLND_INIT_NOTHING:
2263                 break;
2264         }
2265
2266         if (kiblnd_data.kib_peers != NULL)
2267                 LIBCFS_FREE(kiblnd_data.kib_peers,
2268                             sizeof(struct list_head) *
2269                             kiblnd_data.kib_peer_hash_size);
2270
2271         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2272                atomic_read(&libcfs_kmemory));
2273
2274         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2275         PORTAL_MODULE_UNUSE;
2276 }
2277
2278 void
2279 kiblnd_shutdown (lnet_ni_t *ni)
2280 {
2281         kib_net_t        *net = ni->ni_data;
2282         rwlock_t         *g_lock = &kiblnd_data.kib_global_lock;
2283         int               i;
2284         unsigned long     flags;
2285
2286         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2287
2288         if (net == NULL)
2289                 goto out;
2290
2291         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2292                atomic_read(&libcfs_kmemory));
2293
2294         write_lock_irqsave(g_lock, flags);
2295         net->ibn_shutdown = 1;
2296         write_unlock_irqrestore(g_lock, flags);
2297
2298         switch (net->ibn_init) {
2299         default:
2300                 LBUG();
2301
2302         case IBLND_INIT_ALL:
2303                 /* nuke all existing peers within this net */
2304                 kiblnd_del_peer(ni, LNET_NID_ANY);
2305
2306                 /* Wait for all peer state to clean up */
2307                 i = 2;
2308                 while (atomic_read(&net->ibn_npeers) != 0) {
2309                         i++;
2310                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2311                                "%s: waiting for %d peers to disconnect\n",
2312                                libcfs_nid2str(ni->ni_nid),
2313                                atomic_read(&net->ibn_npeers));
2314                         cfs_pause(cfs_time_seconds(1));
2315                 }
2316
2317                 kiblnd_ni_fini_pools(net);
2318
2319                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2320                 net->ibn_dev->ibd_nnets--;
2321
2322                 /* fall through */
2323
2324         case IBLND_INIT_NOTHING:
2325                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2326
2327                 if (net->ibn_dev != NULL &&
2328                     net->ibn_dev->ibd_nnets == 0)
2329                         kiblnd_destroy_dev(net->ibn_dev);
2330
2331                 break;
2332         }
2333
2334         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2335                atomic_read(&libcfs_kmemory));
2336
2337         net->ibn_init = IBLND_INIT_NOTHING;
2338         ni->ni_data = NULL;
2339
2340         LIBCFS_FREE(net, sizeof(*net));
2341
2342 out:
2343         if (list_empty(&kiblnd_data.kib_devs))
2344                 kiblnd_base_shutdown();
2345         return;
2346 }
2347
2348 int
2349 kiblnd_base_startup (void)
2350 {
2351         int i;
2352         int rc;
2353
2354         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2355
2356         PORTAL_MODULE_USE;
2357         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2358
2359         rwlock_init(&kiblnd_data.kib_global_lock);
2360
2361         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2362
2363         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2364         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2365                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2366         if (kiblnd_data.kib_peers == NULL) {
2367                 goto failed;
2368         }
2369         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2370                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2371
2372         spin_lock_init(&kiblnd_data.kib_connd_lock);
2373         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2374         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2375         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2376
2377         spin_lock_init(&kiblnd_data.kib_sched_lock);
2378         INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2379         init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
2380
2381         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2382
2383         /* lists/ptrs/locks initialised */
2384         kiblnd_data.kib_init = IBLND_INIT_DATA;
2385         /*****************************************************/
2386
2387         for (i = 0; i < IBLND_N_SCHED; i++) {
2388                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2389                 if (rc != 0) {
2390                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2391                                i, rc);
2392                         goto failed;
2393                 }
2394         }
2395
2396         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2397         if (rc != 0) {
2398                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2399                 goto failed;
2400         }
2401
2402         /* flag everything initialised */
2403         kiblnd_data.kib_init = IBLND_INIT_ALL;
2404         /*****************************************************/
2405
2406         return 0;
2407
2408  failed:
2409         kiblnd_base_shutdown();
2410         return -ENETDOWN;
2411 }
2412
2413 int
2414 kiblnd_startup (lnet_ni_t *ni)
2415 {
2416         char                     *ifname;
2417         kib_dev_t                *ibdev = NULL;
2418         kib_net_t                *net;
2419         struct list_head         *tmp;
2420         struct timeval            tv;
2421         int                       rc;
2422
2423         LASSERT (ni->ni_lnd == &the_o2iblnd);
2424
2425         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2426                 rc = kiblnd_base_startup();
2427                 if (rc != 0)
2428                         return rc;
2429         }
2430
2431         LIBCFS_ALLOC(net, sizeof(*net));
2432         ni->ni_data = net;
2433         if (net == NULL)
2434                 goto failed;
2435
2436         memset(net, 0, sizeof(*net));
2437
2438         do_gettimeofday(&tv);
2439         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2440
2441         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2442         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2443         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2444         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2445
2446         if (ni->ni_interfaces[0] != NULL) {
2447                 /* Use the IPoIB interface specified in 'networks=' */
2448
2449                 CLASSERT (LNET_MAX_INTERFACES > 1);
2450                 if (ni->ni_interfaces[1] != NULL) {
2451                         CERROR("Multiple interfaces not supported\n");
2452                         goto failed;
2453                 }
2454
2455                 ifname = ni->ni_interfaces[0];
2456         } else {
2457                 ifname = *kiblnd_tunables.kib_default_ipif;
2458         }
2459
2460         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2461                 CERROR("IPoIB interface name too long: %s\n", ifname);
2462                 goto failed;
2463         }
2464
2465         list_for_each (tmp, &kiblnd_data.kib_devs) {
2466                 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
2467
2468                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2469                         break;
2470
2471                 ibdev = NULL;
2472         }
2473
2474         if (ibdev == NULL) {
2475                 __u32                     ip;
2476                 __u32                     netmask;
2477                 int                       up;
2478                 struct rdma_cm_id        *id;
2479                 struct ib_pd             *pd;
2480                 struct sockaddr_in        addr;
2481
2482                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2483                 if (rc != 0) {
2484                         CERROR("Can't query IPoIB interface %s: %d\n",
2485                                ifname, rc);
2486                         goto failed;
2487                 }
2488
2489                 if (!up) {
2490                         CERROR("Can't query IPoIB interface %s: it's down\n",
2491                                ifname);
2492                         goto failed;
2493                 }
2494
2495                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2496                 if (ibdev == NULL)
2497                         goto failed;
2498
2499                 memset(ibdev, 0, sizeof(*ibdev));
2500
2501                 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2502                 ibdev->ibd_ifip = ip;
2503                 strcpy(&ibdev->ibd_ifname[0], ifname);
2504
2505                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2506                 if (IS_ERR(id)) {
2507                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2508                         goto failed;
2509                 }
2510
2511                 ibdev->ibd_cmid = id;
2512
2513                 memset(&addr, 0, sizeof(addr));
2514                 addr.sin_family      = AF_INET;
2515                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2516                 addr.sin_addr.s_addr = htonl(ip);
2517
2518                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2519                 if (rc != 0) {
2520                         CERROR("Can't bind to %s: %d\n", ifname, rc);
2521                         goto failed;
2522                 }
2523
2524                 /* Binding should have assigned me an IB device */
2525                 LASSERT (id->device != NULL);
2526                 CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n",
2527                        ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service,
2528                        id->device->name);
2529
2530                 pd = ib_alloc_pd(id->device);
2531                 if (IS_ERR(pd)) {
2532                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2533                         goto failed;
2534                 }
2535
2536                 ibdev->ibd_pd = pd;
2537
2538                 rc = rdma_listen(id, 256);
2539                 if (rc != 0) {
2540                         CERROR("Can't start listener: %d\n", rc);
2541                         goto failed;
2542                 }
2543
2544                 rc = kiblnd_dev_setup(ibdev);
2545                 if (rc != 0) {
2546                         CERROR("Can't setup device: %d\n", rc);
2547                         goto failed;
2548                 }
2549         }
2550
2551         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2552         net->ibn_dev = ibdev;
2553
2554         rc = kiblnd_net_init_pools(net);
2555         if (rc != 0) {
2556                 CERROR("Failed to initialize NI pools: %d\n", rc);
2557                 goto failed;
2558         }
2559         ibdev->ibd_nnets++;
2560         net->ibn_init = IBLND_INIT_ALL;
2561
2562         return 0;
2563
2564 failed:
2565         if (net->ibn_dev == NULL && ibdev != NULL)
2566                 kiblnd_destroy_dev(ibdev);
2567
2568         kiblnd_shutdown(ni);
2569
2570         CDEBUG(D_NET, "kiblnd_startup failed\n");
2571         return -ENETDOWN;
2572 }
2573
2574 void __exit
2575 kiblnd_module_fini (void)
2576 {
2577         lnet_unregister_lnd(&the_o2iblnd);
2578         kiblnd_tunables_fini();
2579 }
2580
2581 int __init
2582 kiblnd_module_init (void)
2583 {
2584         int    rc;
2585
2586         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2587         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2588                   <= IBLND_MSG_SIZE);
2589         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2590                   <= IBLND_MSG_SIZE);
2591
2592         rc = kiblnd_tunables_init();
2593         if (rc != 0)
2594                 return rc;
2595
2596         lnet_register_lnd(&the_o2iblnd);
2597
2598         return 0;
2599 }
2600
2601 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2602 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2603 MODULE_LICENSE("GPL");
2604
2605 module_init(kiblnd_module_init);
2606 module_exit(kiblnd_module_fini);