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         cfs_atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 349
 350         CFS_INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 351         CFS_INIT_LIST_HEAD(&peer->ibp_conns);
 352         CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue);
 353
 354         cfs_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         cfs_atomic_inc(&net->ibn_npeers);
 361
 362         cfs_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 (cfs_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 (cfs_list_empty(&peer->ibp_conns));
 379         LASSERT (cfs_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         cfs_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         cfs_list_t       *peer_list = kiblnd_nid2peerlist(nid);
 396         cfs_list_t       *tmp;
 397         kib_peer_t       *peer;
 398
 399         cfs_list_for_each (tmp, peer_list) {
 400
 401                 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
 402
 403                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 404                          peer->ibp_accepting > 0 ||
 405                          !cfs_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                        cfs_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 (cfs_list_empty(&peer->ibp_conns));
 423
 424         LASSERT (kiblnd_peer_active(peer));
 425         cfs_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         cfs_list_t            *ptmp;
 436         int                    i;
 437         unsigned long          flags;
 438
 439         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 440
 441         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 442
 443                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 444
 445                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 446                         LASSERT (peer->ibp_connecting > 0 ||
 447                                  peer->ibp_accepting > 0 ||
 448                                  !cfs_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 = cfs_atomic_read(&peer->ibp_refcount);
 458
 459                         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 460                                                    flags);
 461                         return 0;
 462                 }
 463         }
 464
 465         cfs_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         cfs_list_t           *ctmp;
 473         cfs_list_t           *cnxt;
 474         kib_conn_t           *conn;
 475
 476         if (cfs_list_empty(&peer->ibp_conns)) {
 477                 kiblnd_unlink_peer_locked(peer);
 478         } else {
 479                 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 480                         conn = cfs_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         cfs_list_t            *ptmp;
 495         cfs_list_t            *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         cfs_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                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 514                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 515                         LASSERT (peer->ibp_connecting > 0 ||
 516                                  peer->ibp_accepting > 0 ||
 517                                  !cfs_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 (!cfs_list_empty(&peer->ibp_tx_queue)) {
 526                                 LASSERT (cfs_list_empty(&peer->ibp_conns));
 527
 528                                 cfs_list_splice_init(&peer->ibp_tx_queue,
 529                                                      &zombies);
 530                         }
 531
 532                         kiblnd_del_peer_locked(peer);
 533                         rc = 0;         /* matched something */
 534                 }
 535         }
 536
 537         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 538
 539         kiblnd_txlist_done(ni, &zombies, -EIO);
 540
 541         return rc;
 542 }
 543
 544 kib_conn_t *
 545 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 546 {
 547         kib_peer_t            *peer;
 548         cfs_list_t            *ptmp;
 549         kib_conn_t            *conn;
 550         cfs_list_t            *ctmp;
 551         int                    i;
 552         unsigned long          flags;
 553
 554         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 555
 556         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 557                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 558
 559                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 560                         LASSERT (peer->ibp_connecting > 0 ||
 561                                  peer->ibp_accepting > 0 ||
 562                                  !cfs_list_empty(&peer->ibp_conns));
 563
 564                         if (peer->ibp_ni != ni)
 565                                 continue;
 566
 567                         cfs_list_for_each (ctmp, &peer->ibp_conns) {
 568                                 if (index-- > 0)
 569                                         continue;
 570
 571                                 conn = cfs_list_entry(ctmp, kib_conn_t,
 572                                                       ibc_list);
 573                                 kiblnd_conn_addref(conn);
 574                                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 575                                                            flags);
 576                                 return conn;
 577                         }
 578                 }
 579         }
 580
 581         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 582         return NULL;
 583 }
 584
 585 void
 586 kiblnd_debug_rx (kib_rx_t *rx)
 587 {
 588         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 589                rx, rx->rx_status, rx->rx_msg->ibm_type,
 590                rx->rx_msg->ibm_credits);
 591 }
 592
 593 void
 594 kiblnd_debug_tx (kib_tx_t *tx)
 595 {
 596         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 597                "cookie "LPX64" msg %s%s type %x cred %d\n",
 598                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 599                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 600                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 601                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 602                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 603 }
 604
 605 void
 606 kiblnd_debug_conn (kib_conn_t *conn)
 607 {
 608         cfs_list_t           *tmp;
 609         int                   i;
 610
 611         cfs_spin_lock(&conn->ibc_lock);
 612
 613         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 614                cfs_atomic_read(&conn->ibc_refcount), conn,
 615                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 616         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 617                conn->ibc_state, conn->ibc_noops_posted,
 618                conn->ibc_nsends_posted, conn->ibc_credits,
 619                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 620         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 621
 622         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 623         cfs_list_for_each(tmp, &conn->ibc_early_rxs)
 624                 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
 625
 626         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 627         cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 628                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 629
 630         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 631         cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 632                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 633
 634         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 635         cfs_list_for_each(tmp, &conn->ibc_tx_queue)
 636                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 637
 638         CDEBUG(D_CONSOLE, "   active_txs:\n");
 639         cfs_list_for_each(tmp, &conn->ibc_active_txs)
 640                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 641
 642         CDEBUG(D_CONSOLE, "   rxs:\n");
 643         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 644                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 645
 646         cfs_spin_unlock(&conn->ibc_lock);
 647 }
 648
 649 int
 650 kiblnd_translate_mtu(int value)
 651 {
 652         switch (value) {
 653         default:
 654                 return -1;
 655         case 0:
 656                 return 0;
 657         case 256:
 658                 return IB_MTU_256;
 659         case 512:
 660                 return IB_MTU_512;
 661         case 1024:
 662                 return IB_MTU_1024;
 663         case 2048:
 664                 return IB_MTU_2048;
 665         case 4096:
 666                 return IB_MTU_4096;
 667         }
 668 }
 669
 670 static void
 671 kiblnd_setup_mtu(struct rdma_cm_id *cmid)
 672 {
 673         unsigned long flags;
 674         int           mtu;
 675
 676         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 677         if (cmid->route.path_rec == NULL)
 678                 return;
 679
 680         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 681
 682         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 683         LASSERT (mtu >= 0);
 684         if (mtu != 0)
 685                 cmid->route.path_rec->mtu = mtu;
 686
 687         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 688 }
 689
 690 kib_conn_t *
 691 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 692                    int state, int version)
 693 {
 694         /* CAVEAT EMPTOR:
 695          * If the new conn is created successfully it takes over the caller's
 696          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 697          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 698          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 699          * to destroy 'cmid' here since I'm called from the CM which still has
 700          * its ref on 'cmid'). */
 701         kib_net_t              *net = peer->ibp_ni->ni_data;
 702         struct ib_qp_init_attr *init_qp_attr;
 703         kib_conn_t             *conn;
 704         struct ib_cq           *cq;
 705         unsigned long           flags;
 706         int                     rc;
 707         int                     i;
 708
 709         LASSERT (net != NULL);
 710         LASSERT (!cfs_in_interrupt());
 711
 712         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 713         if (init_qp_attr == NULL) {
 714                 CERROR("Can't allocate qp_attr for %s\n",
 715                        libcfs_nid2str(peer->ibp_nid));
 716                 goto failed_0;
 717         }
 718
 719         LIBCFS_ALLOC(conn, sizeof(*conn));
 720         if (conn == NULL) {
 721                 CERROR("Can't allocate connection for %s\n",
 722                        libcfs_nid2str(peer->ibp_nid));
 723                 goto failed_1;
 724         }
 725
 726         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 727
 728         conn->ibc_state = IBLND_CONN_INIT;
 729         conn->ibc_version = version;
 730         conn->ibc_peer = peer;                  /* I take the caller's ref */
 731         cmid->context = conn;                   /* for future CM callbacks */
 732         conn->ibc_cmid = cmid;
 733
 734         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 735         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 736         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 737         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 738         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 739         cfs_spin_lock_init(&conn->ibc_lock);
 740
 741         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 742         if (conn->ibc_connvars == NULL) {
 743                 CERROR("Can't allocate in-progress connection state\n");
 744                 goto failed_2;
 745         }
 746         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 747
 748         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 749         if (conn->ibc_rxs == NULL) {
 750                 CERROR("Cannot allocate RX buffers\n");
 751                 goto failed_2;
 752         }
 753         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 754
 755         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
 756                                 IBLND_RX_MSG_PAGES(version));
 757         if (rc != 0)
 758                 goto failed_2;
 759
 760         kiblnd_map_rx_descs(conn);
 761
 762 #ifdef HAVE_OFED_IB_COMP_VECTOR
 763         cq = ib_create_cq(cmid->device,
 764                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 765                           IBLND_CQ_ENTRIES(version), 0);
 766 #else
 767         cq = ib_create_cq(cmid->device,
 768                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 769                           IBLND_CQ_ENTRIES(version));
 770 #endif
 771         if (IS_ERR(cq)) {
 772                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 773                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 774                 goto failed_2;
 775         }
 776
 777         conn->ibc_cq = cq;
 778
 779         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 780         if (rc != 0) {
 781                 CERROR("Can't request completion notificiation: %d\n", rc);
 782                 goto failed_2;
 783         }
 784
 785         kiblnd_setup_mtu(cmid);
 786
 787         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 788         init_qp_attr->event_handler = kiblnd_qp_event;
 789         init_qp_attr->qp_context = conn;
 790         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 791         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 792         init_qp_attr->cap.max_send_sge = 1;
 793         init_qp_attr->cap.max_recv_sge = 1;
 794         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 795         init_qp_attr->qp_type = IB_QPT_RC;
 796         init_qp_attr->send_cq = cq;
 797         init_qp_attr->recv_cq = cq;
 798
 799         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 800         if (rc != 0) {
 801                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 802                        rc, init_qp_attr->cap.max_send_wr,
 803                        init_qp_attr->cap.max_recv_wr);
 804                 goto failed_2;
 805         }
 806
 807         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 808
 809         /* 1 ref for caller and each rxmsg */
 810         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 811         conn->ibc_nrx = IBLND_RX_MSGS(version);
 812
 813         /* post receives */
 814         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 815                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 816                                     IBLND_POSTRX_NO_CREDIT);
 817                 if (rc != 0) {
 818                         CERROR("Can't post rxmsg: %d\n", rc);
 819
 820                         /* Make posted receives complete */
 821                         kiblnd_abort_receives(conn);
 822
 823                         /* correct # of posted buffers
 824                          * NB locking needed now I'm racing with completion */
 825                         cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
 826                                               flags);
 827                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 828                         cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 829                                                    flags);
 830
 831                         /* cmid will be destroyed by CM(ofed) after cm_callback
 832                          * returned, so we can't refer it anymore
 833                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 834                         rdma_destroy_qp(conn->ibc_cmid);
 835                         conn->ibc_cmid = NULL;
 836
 837                         /* Drop my own and unused rxbuffer refcounts */
 838                         while (i++ <= IBLND_RX_MSGS(version))
 839                                 kiblnd_conn_decref(conn);
 840
 841                         return NULL;
 842                 }
 843         }
 844
 845         /* Init successful! */
 846         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 847                  state == IBLND_CONN_PASSIVE_WAIT);
 848         conn->ibc_state = state;
 849
 850         /* 1 more conn */
 851         cfs_atomic_inc(&net->ibn_nconns);
 852         return conn;
 853
 854  failed_2:
 855         kiblnd_destroy_conn(conn);
 856  failed_1:
 857         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 858  failed_0:
 859         return NULL;
 860 }
 861
 862 void
 863 kiblnd_destroy_conn (kib_conn_t *conn)
 864 {
 865         struct rdma_cm_id *cmid = conn->ibc_cmid;
 866         kib_peer_t        *peer = conn->ibc_peer;
 867         int                rc;
 868
 869         LASSERT (!cfs_in_interrupt());
 870         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 871         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 872         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 873         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 874         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 875         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 876         LASSERT (conn->ibc_noops_posted == 0);
 877         LASSERT (conn->ibc_nsends_posted == 0);
 878
 879         switch (conn->ibc_state) {
 880         default:
 881                 /* conn must be completely disengaged from the network */
 882                 LBUG();
 883
 884         case IBLND_CONN_DISCONNECTED:
 885                 /* connvars should have been freed already */
 886                 LASSERT (conn->ibc_connvars == NULL);
 887                 break;
 888
 889         case IBLND_CONN_INIT:
 890                 break;
 891         }
 892
 893         /* conn->ibc_cmid might be destroyed by CM already */
 894         if (cmid != NULL && cmid->qp != NULL)
 895                 rdma_destroy_qp(cmid);
 896
 897         if (conn->ibc_cq != NULL) {
 898                 rc = ib_destroy_cq(conn->ibc_cq);
 899                 if (rc != 0)
 900                         CWARN("Error destroying CQ: %d\n", rc);
 901         }
 902
 903         if (conn->ibc_rx_pages != NULL)
 904                 kiblnd_unmap_rx_descs(conn);
 905
 906         if (conn->ibc_rxs != NULL) {
 907                 LIBCFS_FREE(conn->ibc_rxs,
 908                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 909         }
 910
 911         if (conn->ibc_connvars != NULL)
 912                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 913
 914         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 915         if (conn->ibc_state != IBLND_CONN_INIT) {
 916                 kib_net_t *net = peer->ibp_ni->ni_data;
 917
 918                 kiblnd_peer_decref(peer);
 919                 rdma_destroy_id(cmid);
 920                 cfs_atomic_dec(&net->ibn_nconns);
 921         }
 922
 923         LIBCFS_FREE(conn, sizeof(*conn));
 924 }
 925
 926 int
 927 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 928 {
 929         kib_conn_t             *conn;
 930         cfs_list_t             *ctmp;
 931         cfs_list_t             *cnxt;
 932         int                     count = 0;
 933
 934         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 935                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 936
 937                 CDEBUG(D_NET, "Closing conn -> %s, "
 938                               "version: %x, reason: %d\n",
 939                        libcfs_nid2str(peer->ibp_nid),
 940                        conn->ibc_version, why);
 941
 942                 kiblnd_close_conn_locked(conn, why);
 943                 count++;
 944         }
 945
 946         return count;
 947 }
 948
 949 int
 950 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 951                                  int version, __u64 incarnation)
 952 {
 953         kib_conn_t             *conn;
 954         cfs_list_t             *ctmp;
 955         cfs_list_t             *cnxt;
 956         int                     count = 0;
 957
 958         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 959                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 960
 961                 if (conn->ibc_version     == version &&
 962                     conn->ibc_incarnation == incarnation)
 963                         continue;
 964
 965                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
 966                               "incarnation:"LPX64"(%x, "LPX64")\n",
 967                        libcfs_nid2str(peer->ibp_nid),
 968                        conn->ibc_version, conn->ibc_incarnation,
 969                        version, incarnation);
 970
 971                 kiblnd_close_conn_locked(conn, -ESTALE);
 972                 count++;
 973         }
 974
 975         return count;
 976 }
 977
 978 int
 979 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 980 {
 981         kib_peer_t             *peer;
 982         cfs_list_t             *ptmp;
 983         cfs_list_t             *pnxt;
 984         int                     lo;
 985         int                     hi;
 986         int                     i;
 987         unsigned long           flags;
 988         int                     count = 0;
 989
 990         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 991
 992         if (nid != LNET_NID_ANY)
 993                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 994         else {
 995                 lo = 0;
 996                 hi = kiblnd_data.kib_peer_hash_size - 1;
 997         }
 998
 999         for (i = lo; i <= hi; i++) {
1000                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1001
1002                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1003                         LASSERT (peer->ibp_connecting > 0 ||
1004                                  peer->ibp_accepting > 0 ||
1005                                  !cfs_list_empty(&peer->ibp_conns));
1006
1007                         if (peer->ibp_ni != ni)
1008                                 continue;
1009
1010                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1011                                 continue;
1012
1013                         count += kiblnd_close_peer_conns_locked(peer, 0);
1014                 }
1015         }
1016
1017         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1018
1019         /* wildcards always succeed */
1020         if (nid == LNET_NID_ANY)
1021                 return 0;
1022
1023         return (count == 0) ? -ENOENT : 0;
1024 }
1025
1026 int
1027 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1028 {
1029         struct libcfs_ioctl_data *data = arg;
1030         int                       rc = -EINVAL;
1031
1032         switch(cmd) {
1033         case IOC_LIBCFS_GET_PEER: {
1034                 lnet_nid_t   nid = 0;
1035                 int          count = 0;
1036
1037                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1038                                           &nid, &count);
1039                 data->ioc_nid    = nid;
1040                 data->ioc_count  = count;
1041                 break;
1042         }
1043
1044         case IOC_LIBCFS_DEL_PEER: {
1045                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1046                 break;
1047         }
1048         case IOC_LIBCFS_GET_CONN: {
1049                 kib_conn_t *conn;
1050
1051                 rc = 0;
1052                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1053                 if (conn == NULL) {
1054                         rc = -ENOENT;
1055                         break;
1056                 }
1057
1058                 LASSERT (conn->ibc_cmid != NULL);
1059                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1060                 if (conn->ibc_cmid->route.path_rec == NULL)
1061                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1062                 else
1063                         data->ioc_u32[0] =
1064                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1065                 kiblnd_conn_decref(conn);
1066                 break;
1067         }
1068         case IOC_LIBCFS_CLOSE_CONNECTION: {
1069                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1070                 break;
1071         }
1072
1073         default:
1074                 break;
1075         }
1076
1077         return rc;
1078 }
1079
1080 void
1081 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1082 {
1083         cfs_time_t         last_alive = 0;
1084         cfs_rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1085         kib_peer_t        *peer;
1086         unsigned long      flags;
1087
1088         cfs_read_lock_irqsave(glock, flags);
1089
1090         peer = kiblnd_find_peer_locked(nid);
1091         if (peer != NULL) {
1092                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1093                          peer->ibp_accepting > 0 ||
1094                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1095                 last_alive = peer->ibp_last_alive;
1096         }
1097
1098         cfs_read_unlock_irqrestore(glock, flags);
1099
1100         if (last_alive != 0)
1101                 *when = last_alive;
1102
1103         /* peer is not persistent in hash, trigger peer creation
1104          * and connection establishment with a NULL tx */
1105         if (peer == NULL)
1106                 kiblnd_launch_tx(ni, NULL, nid);
1107         return;
1108 }
1109
1110 void
1111 kiblnd_free_pages (kib_pages_t *p)
1112 {
1113         int         npages = p->ibp_npages;
1114         int         i;
1115
1116         LASSERT (p->ibp_device == NULL);
1117
1118         for (i = 0; i < npages; i++)
1119                 if (p->ibp_pages[i] != NULL)
1120                         __free_page(p->ibp_pages[i]);
1121
1122         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1123 }
1124
1125 int
1126 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1127 {
1128         kib_pages_t   *p;
1129         int            i;
1130
1131         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1132         if (p == NULL) {
1133                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1134                 return -ENOMEM;
1135         }
1136
1137         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1138         p->ibp_npages = npages;
1139
1140         for (i = 0; i < npages; i++) {
1141                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1142                 if (p->ibp_pages[i] == NULL) {
1143                         CERROR("Can't allocate page %d of %d\n", i, npages);
1144                         kiblnd_free_pages(p);
1145                         return -ENOMEM;
1146                 }
1147         }
1148
1149         *pp = p;
1150         return 0;
1151 }
1152
1153 void
1154 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1155 {
1156         kib_rx_t *rx;
1157         int       i;
1158
1159         LASSERT (conn->ibc_rxs != NULL);
1160         LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1161
1162         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1163                 rx = &conn->ibc_rxs[i];
1164
1165                 LASSERT (rx->rx_nob >= 0); /* not posted */
1166
1167                 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1168                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1169                                                           rx->rx_msgaddr),
1170                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1171         }
1172
1173         conn->ibc_rx_pages->ibp_device = NULL;
1174
1175         kiblnd_free_pages(conn->ibc_rx_pages);
1176
1177         conn->ibc_rx_pages = NULL;
1178 }
1179
1180 void
1181 kiblnd_map_rx_descs(kib_conn_t *conn)
1182 {
1183         kib_rx_t       *rx;
1184         struct page    *pg;
1185         int             pg_off;
1186         int             ipg;
1187         int             i;
1188
1189         for (pg_off = ipg = i = 0;
1190              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1191                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1192                 rx = &conn->ibc_rxs[i];
1193
1194                 rx->rx_conn = conn;
1195                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1196
1197                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1198                                                        rx->rx_msg, IBLND_MSG_SIZE,
1199                                                        DMA_FROM_DEVICE);
1200                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1201                                                    rx->rx_msgaddr));
1202                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1203
1204                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1205                        i, rx->rx_msg, rx->rx_msgaddr,
1206                        lnet_page2phys(pg) + pg_off);
1207
1208                 pg_off += IBLND_MSG_SIZE;
1209                 LASSERT (pg_off <= PAGE_SIZE);
1210
1211                 if (pg_off == PAGE_SIZE) {
1212                         pg_off = 0;
1213                         ipg++;
1214                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1215                 }
1216         }
1217
1218         conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1219 }
1220
1221 static void
1222 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1223 {
1224         kib_tx_t       *tx;
1225         int             i;
1226
1227         LASSERT (tpo->tpo_pool.po_allocated == 0);
1228         LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1229
1230         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1231                 tx = &tpo->tpo_tx_descs[i];
1232                 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1233                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1234                                                           tx->tx_msgaddr),
1235                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1236         }
1237         tpo->tpo_tx_pages->ibp_device = NULL;
1238 }
1239
1240 static void
1241 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1242 {
1243         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1244         kib_pool_t     *pool  = &tpo->tpo_pool;
1245         kib_net_t      *net   = pool->po_owner->ps_net;
1246         struct page    *page;
1247         kib_tx_t       *tx;
1248         int             page_offset;
1249         int             ipage;
1250         int             i;
1251
1252         LASSERT (net != NULL);
1253         LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1254                  net->ibn_dev->ibd_cmid->device != NULL);
1255
1256         /* pre-mapped messages are not bigger than 1 page */
1257         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1258
1259         /* No fancy arithmetic when we do the buffer calculations */
1260         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1261
1262         txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1263
1264         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1265                 page = txpgs->ibp_pages[ipage];
1266                 tx = &tpo->tpo_tx_descs[i];
1267
1268                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1269                                            page_offset);
1270
1271                 tx->tx_msgaddr = kiblnd_dma_map_single(
1272                         txpgs->ibp_device, tx->tx_msg,
1273                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1274                 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1275                                                    tx->tx_msgaddr));
1276                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1277
1278                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1279
1280                 page_offset += IBLND_MSG_SIZE;
1281                 LASSERT (page_offset <= PAGE_SIZE);
1282
1283                 if (page_offset == PAGE_SIZE) {
1284                         page_offset = 0;
1285                         ipage++;
1286                         LASSERT (ipage <= txpgs->ibp_npages);
1287                 }
1288         }
1289 }
1290
1291 struct ib_mr *
1292 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1293 {
1294         __u64   index;
1295
1296         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1297
1298         if (net->ibn_dev->ibd_nmrs == 1)
1299                 return net->ibn_dev->ibd_mrs[0];
1300
1301         index = addr >> net->ibn_dev->ibd_mr_shift;
1302
1303         if (index <  net->ibn_dev->ibd_nmrs &&
1304             index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1305                 return net->ibn_dev->ibd_mrs[index];
1306
1307         return NULL;
1308 }
1309
1310 struct ib_mr *
1311 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1312 {
1313         struct ib_mr *prev_mr;
1314         struct ib_mr *mr;
1315         int           i;
1316
1317         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1318
1319         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1320             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1321                 return NULL;
1322
1323         if (net->ibn_dev->ibd_nmrs == 1)
1324                 return net->ibn_dev->ibd_mrs[0];
1325
1326         for (i = 0, mr = prev_mr = NULL;
1327              i < rd->rd_nfrags; i++) {
1328                 mr = kiblnd_find_dma_mr(net,
1329                                         rd->rd_frags[i].rf_addr,
1330                                         rd->rd_frags[i].rf_nob);
1331                 if (prev_mr == NULL)
1332                         prev_mr = mr;
1333
1334                 if (mr == NULL || prev_mr != mr) {
1335                         /* Can't covered by one single MR */
1336                         mr = NULL;
1337                         break;
1338                 }
1339         }
1340
1341         return mr;
1342 }
1343
1344 void
1345 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1346 {
1347         LASSERT (pool->fpo_map_count == 0);
1348
1349         if (pool->fpo_fmr_pool != NULL)
1350                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1351
1352         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1353 }
1354
1355 void
1356 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1357 {
1358         kib_fmr_pool_t *pool;
1359
1360         while (!cfs_list_empty(head)) {
1361                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1362                 cfs_list_del(&pool->fpo_list);
1363                 kiblnd_destroy_fmr_pool(pool);
1364         }
1365 }
1366
1367 int
1368 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1369 {
1370         /* FMR pool for RDMA */
1371         kib_fmr_pool_t          *fpo;
1372         struct ib_fmr_pool_param param = {
1373                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1374                 .page_shift        = PAGE_SHIFT,
1375                 .access            = (IB_ACCESS_LOCAL_WRITE |
1376                                       IB_ACCESS_REMOTE_WRITE),
1377                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1378                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1379                 .flush_function    = NULL,
1380                 .flush_arg         = NULL,
1381                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1382         int rc;
1383
1384         LASSERT (fps->fps_net->ibn_dev != NULL &&
1385                  fps->fps_net->ibn_dev->ibd_pd != NULL);
1386
1387         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1388         if (fpo == NULL)
1389                 return -ENOMEM;
1390
1391         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1392         fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, &param);
1393         if (IS_ERR(fpo->fpo_fmr_pool)) {
1394                 CERROR("Failed to create FMR pool: %ld\n",
1395                        PTR_ERR(fpo->fpo_fmr_pool));
1396                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1397                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1398                 return rc;
1399         }
1400
1401         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1402         fpo->fpo_owner    = fps;
1403         *pp_fpo = fpo;
1404
1405         return 0;
1406 }
1407
1408 static void
1409 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1410 {
1411         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1412 }
1413
1414 static int
1415 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1416 {
1417         kib_fmr_pool_t *fpo;
1418         int             rc;
1419
1420         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1421
1422         fps->fps_net = net;
1423         cfs_spin_lock_init(&fps->fps_lock);
1424         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1425         rc = kiblnd_create_fmr_pool(fps, &fpo);
1426         if (rc == 0)
1427                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1428
1429         return rc;
1430 }
1431
1432 void
1433 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1434 {
1435         CFS_LIST_HEAD     (zombies);
1436         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1437         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1438         kib_fmr_pool_t    *tmp;
1439         int                rc;
1440
1441         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1442         LASSERT (rc == 0);
1443
1444         if (status != 0) {
1445                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1446                 LASSERT (rc == 0);
1447         }
1448
1449         fmr->fmr_pool = NULL;
1450         fmr->fmr_pfmr = NULL;
1451
1452         cfs_spin_lock(&fps->fps_lock);
1453         fpo->fpo_map_count --;  /* decref the pool */
1454
1455         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1456                 /* the first pool is persistent */
1457                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1458                         continue;
1459
1460                 if (fpo->fpo_map_count == 0 &&  /* no more reference */
1461                     cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1462                         cfs_list_move(&fpo->fpo_list, &zombies);
1463                         fps->fps_version ++;
1464                 }
1465         }
1466         cfs_spin_unlock(&fps->fps_lock);
1467
1468         if (!cfs_list_empty(&zombies))
1469                 kiblnd_destroy_fmr_pool_list(&zombies);
1470 }
1471
1472 int
1473 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1474                     __u64 iov, kib_fmr_t *fmr)
1475 {
1476         struct ib_pool_fmr *pfmr;
1477         kib_fmr_pool_t     *fpo;
1478         __u64               version;
1479         int                 rc;
1480
1481         LASSERT (fps->fps_net->ibn_with_fmr);
1482  again:
1483         cfs_spin_lock(&fps->fps_lock);
1484         version = fps->fps_version;
1485         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1486                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1487                 fpo->fpo_map_count ++;
1488                 cfs_spin_unlock(&fps->fps_lock);
1489
1490                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1491                                             pages, npages, iov);
1492                 if (likely(!IS_ERR(pfmr))) {
1493                         fmr->fmr_pool = fpo;
1494                         fmr->fmr_pfmr = pfmr;
1495                         return 0;
1496                 }
1497
1498                 cfs_spin_lock(&fps->fps_lock);
1499                 fpo->fpo_map_count --;
1500                 if (PTR_ERR(pfmr) != -EAGAIN) {
1501                         cfs_spin_unlock(&fps->fps_lock);
1502                         return PTR_ERR(pfmr);
1503                 }
1504
1505                 /* EAGAIN and ... */
1506                 if (version != fps->fps_version) {
1507                         cfs_spin_unlock(&fps->fps_lock);
1508                         goto again;
1509                 }
1510         }
1511
1512         if (fps->fps_increasing) {
1513                 cfs_spin_unlock(&fps->fps_lock);
1514                 CDEBUG(D_NET, "Another thread is allocating new "
1515                               "FMR pool, waiting for her to complete\n");
1516                 cfs_schedule();
1517                 goto again;
1518
1519         }
1520
1521         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1522                 /* someone failed recently */
1523                 cfs_spin_unlock(&fps->fps_lock);
1524                 return -EAGAIN;
1525         }
1526
1527         fps->fps_increasing = 1;
1528         cfs_spin_unlock(&fps->fps_lock);
1529
1530         CDEBUG(D_NET, "Allocate new FMR pool\n");
1531         rc = kiblnd_create_fmr_pool(fps, &fpo);
1532         cfs_spin_lock(&fps->fps_lock);
1533         fps->fps_increasing = 0;
1534         if (rc == 0) {
1535                 fps->fps_version ++;
1536                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1537         } else {
1538                 fps->fps_next_retry = cfs_time_shift(10);
1539         }
1540         cfs_spin_unlock(&fps->fps_lock);
1541
1542         goto again;
1543 }
1544
1545 static void
1546 kiblnd_fini_pool(kib_pool_t *pool)
1547 {
1548         LASSERT (cfs_list_empty(&pool->po_free_list));
1549         LASSERT (pool->po_allocated == 0);
1550
1551         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1552 }
1553
1554 static void
1555 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1556 {
1557         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1558
1559         memset(pool, 0, sizeof(kib_pool_t));
1560         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1561         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1562         pool->po_owner    = ps;
1563         pool->po_size     = size;
1564 }
1565
1566 void
1567 kiblnd_destroy_pool_list(kib_poolset_t *ps, cfs_list_t *head)
1568 {
1569         kib_pool_t *pool;
1570
1571         while (!cfs_list_empty(head)) {
1572                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1573                 cfs_list_del(&pool->po_list);
1574                 ps->ps_pool_destroy(pool);
1575         }
1576 }
1577
1578 static void
1579 kiblnd_fini_pool_set(kib_poolset_t *ps)
1580 {
1581         kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1582 }
1583
1584 static int
1585 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1586                      char *name, int size,
1587                      kib_ps_pool_create_t po_create,
1588                      kib_ps_pool_destroy_t po_destroy,
1589                      kib_ps_node_init_t nd_init,
1590                      kib_ps_node_fini_t nd_fini)
1591 {
1592         kib_pool_t    *pool;
1593         int            rc;
1594
1595         memset(ps, 0, sizeof(kib_poolset_t));
1596
1597         ps->ps_net          = net;
1598         ps->ps_pool_create  = po_create;
1599         ps->ps_pool_destroy = po_destroy;
1600         ps->ps_node_init    = nd_init;
1601         ps->ps_node_fini    = nd_fini;
1602         ps->ps_pool_size    = size;
1603         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1604         cfs_spin_lock_init(&ps->ps_lock);
1605         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1606
1607         rc = ps->ps_pool_create(ps, size, &pool);
1608         if (rc == 0)
1609                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1610         else
1611                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1612
1613         return rc;
1614 }
1615
1616 void
1617 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1618 {
1619         CFS_LIST_HEAD  (zombies);
1620         kib_poolset_t  *ps = pool->po_owner;
1621         kib_pool_t     *tmp;
1622         cfs_time_t      now = cfs_time_current();
1623
1624         cfs_spin_lock(&ps->ps_lock);
1625
1626         if (ps->ps_node_fini != NULL)
1627                 ps->ps_node_fini(pool, node);
1628
1629         LASSERT (pool->po_allocated > 0);
1630         cfs_list_add(node, &pool->po_free_list);
1631         pool->po_allocated --;
1632
1633         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1634                 /* the first pool is persistent */
1635                 if (ps->ps_pool_list.next == &pool->po_list)
1636                         continue;
1637
1638                 if (pool->po_allocated == 0 &&
1639                     cfs_time_aftereq(now, pool->po_deadline))
1640                         cfs_list_move(&pool->po_list, &zombies);
1641         }
1642         cfs_spin_unlock(&ps->ps_lock);
1643
1644         if (!cfs_list_empty(&zombies))
1645                 kiblnd_destroy_pool_list(ps, &zombies);
1646 }
1647
1648 cfs_list_t *
1649 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1650 {
1651         cfs_list_t            *node;
1652         kib_pool_t            *pool;
1653         int                    rc;
1654
1655  again:
1656         cfs_spin_lock(&ps->ps_lock);
1657         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1658                 if (cfs_list_empty(&pool->po_free_list))
1659                         continue;
1660
1661                 pool->po_allocated ++;
1662                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1663                 node = pool->po_free_list.next;
1664                 cfs_list_del(node);
1665
1666                 if (ps->ps_node_init != NULL) {
1667                         /* still hold the lock */
1668                         ps->ps_node_init(pool, node);
1669                 }
1670                 cfs_spin_unlock(&ps->ps_lock);
1671                 return node;
1672         }
1673
1674         /* no available tx pool and ... */
1675         if (ps->ps_increasing) {
1676                 /* another thread is allocating a new pool */
1677                 cfs_spin_unlock(&ps->ps_lock);
1678                 CDEBUG(D_NET, "Another thread is allocating new "
1679                        "%s pool, waiting for her to complete\n",
1680                        ps->ps_name);
1681                 cfs_schedule();
1682                 goto again;
1683         }
1684
1685         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1686                 /* someone failed recently */
1687                 cfs_spin_unlock(&ps->ps_lock);
1688                 return NULL;
1689         }
1690
1691         ps->ps_increasing = 1;
1692         cfs_spin_unlock(&ps->ps_lock);
1693
1694         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1695
1696         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1697
1698         cfs_spin_lock(&ps->ps_lock);
1699         ps->ps_increasing = 0;
1700         if (rc == 0) {
1701                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1702         } else {
1703                 /* retry 10 seconds later */
1704                 ps->ps_next_retry = cfs_time_shift(10);
1705                 CERROR("Can't allocate new %s pool because out of memory\n",
1706                        ps->ps_name);
1707         }
1708         cfs_spin_unlock(&ps->ps_lock);
1709
1710         goto again;
1711 }
1712
1713 void
1714 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1715 {
1716         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1717         struct ib_mr        *mr  = pmr->pmr_mr;
1718
1719         pmr->pmr_mr = NULL;
1720         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1721         if (mr != NULL)
1722                 ib_dereg_mr(mr);
1723 }
1724
1725 int
1726 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1727                     __u64 *iova, kib_phys_mr_t **pp_pmr)
1728 {
1729         kib_phys_mr_t *pmr;
1730         cfs_list_t    *node;
1731         int            rc;
1732         int            i;
1733
1734         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1735         if (node == NULL) {
1736                 CERROR("Failed to allocate PMR descriptor\n");
1737                 return -ENOMEM;
1738         }
1739
1740         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1741         for (i = 0; i < rd->rd_nfrags; i ++) {
1742                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1743                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1744         }
1745
1746         pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1747                                      pmr->pmr_ipb, rd->rd_nfrags,
1748                                      IB_ACCESS_LOCAL_WRITE |
1749                                      IB_ACCESS_REMOTE_WRITE,
1750                                      iova);
1751         if (!IS_ERR(pmr->pmr_mr)) {
1752                 pmr->pmr_iova = *iova;
1753                 *pp_pmr = pmr;
1754                 return 0;
1755         }
1756
1757         rc = PTR_ERR(pmr->pmr_mr);
1758         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1759
1760         pmr->pmr_mr = NULL;
1761         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1762
1763         return rc;
1764 }
1765
1766 static void
1767 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1768 {
1769         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1770         kib_phys_mr_t  *pmr;
1771
1772         LASSERT (pool->po_allocated == 0);
1773
1774         while (!cfs_list_empty(&pool->po_free_list)) {
1775                 pmr = cfs_list_entry(pool->po_free_list.next,
1776                                      kib_phys_mr_t, pmr_list);
1777
1778                 LASSERT (pmr->pmr_mr == NULL);
1779                 cfs_list_del(&pmr->pmr_list);
1780
1781                 if (pmr->pmr_ipb != NULL) {
1782                         LIBCFS_FREE(pmr->pmr_ipb,
1783                                     IBLND_MAX_RDMA_FRAGS *
1784                                     sizeof(struct ib_phys_buf));
1785                 }
1786
1787                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1788         }
1789
1790         kiblnd_fini_pool(pool);
1791         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1792 }
1793
1794 static int
1795 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1796 {
1797         kib_pmr_pool_t      *ppo;
1798         kib_pool_t          *pool;
1799         kib_phys_mr_t       *pmr;
1800         int                  i;
1801
1802         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1803         if (ppo == NULL) {
1804                 CERROR("Failed to allocate PMR pool\n");
1805                 return -ENOMEM;
1806         }
1807
1808         pool = &ppo->ppo_pool;
1809         kiblnd_init_pool(ps, pool, size);
1810
1811         for (i = 0; i < size; i++) {
1812                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1813                 if (pmr == NULL)
1814                         break;
1815
1816                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1817                 pmr->pmr_pool = ppo;
1818                 LIBCFS_ALLOC(pmr->pmr_ipb,
1819                              IBLND_MAX_RDMA_FRAGS *
1820                              sizeof(struct ib_phys_buf));
1821                 if (pmr->pmr_ipb == NULL)
1822                         break;
1823
1824                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1825         }
1826
1827         if (i < size) {
1828                 ps->ps_pool_destroy(pool);
1829                 return -ENOMEM;
1830         }
1831
1832         *pp_po = pool;
1833         return 0;
1834 }
1835
1836 static void
1837 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1838 {
1839         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1840         int             i;
1841
1842         LASSERT (pool->po_allocated == 0);
1843
1844         if (tpo->tpo_tx_pages != NULL) {
1845                 if (tpo->tpo_tx_pages->ibp_device != NULL)
1846                         kiblnd_unmap_tx_pool(tpo);
1847                 kiblnd_free_pages(tpo->tpo_tx_pages);
1848         }
1849
1850         if (tpo->tpo_tx_descs == NULL)
1851                 goto out;
1852
1853         for (i = 0; i < pool->po_size; i++) {
1854                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1855
1856                 cfs_list_del(&tx->tx_list);
1857                 if (tx->tx_pages != NULL)
1858                         LIBCFS_FREE(tx->tx_pages,
1859                                     LNET_MAX_IOV *
1860                                     sizeof(*tx->tx_pages));
1861                 if (tx->tx_frags != NULL)
1862                         LIBCFS_FREE(tx->tx_frags,
1863                                     IBLND_MAX_RDMA_FRAGS *
1864                                             sizeof(*tx->tx_frags));
1865                 if (tx->tx_wrq != NULL)
1866                         LIBCFS_FREE(tx->tx_wrq,
1867                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1868                                     sizeof(*tx->tx_wrq));
1869                 if (tx->tx_sge != NULL)
1870                         LIBCFS_FREE(tx->tx_sge,
1871                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1872                                     sizeof(*tx->tx_sge));
1873                 if (tx->tx_rd != NULL)
1874                         LIBCFS_FREE(tx->tx_rd,
1875                                     offsetof(kib_rdma_desc_t,
1876                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1877         }
1878
1879         LIBCFS_FREE(tpo->tpo_tx_descs,
1880                     pool->po_size * sizeof(kib_tx_t));
1881 out:
1882         kiblnd_fini_pool(pool);
1883         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1884 }
1885
1886 static int
1887 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1888 {
1889         int            i;
1890         int            npg;
1891         kib_pool_t    *pool;
1892         kib_tx_pool_t *tpo;
1893
1894         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1895         if (tpo == NULL) {
1896                 CERROR("Failed to allocate TX pool\n");
1897                 return -ENOMEM;
1898         }
1899
1900         pool = &tpo->tpo_pool;
1901         kiblnd_init_pool(ps, pool, size);
1902         tpo->tpo_tx_descs = NULL;
1903         tpo->tpo_tx_pages = NULL;
1904
1905         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1906         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1907                 CERROR("Can't allocate tx pages: %d\n", npg);
1908                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1909                 return -ENOMEM;
1910         }
1911
1912         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1913         if (tpo->tpo_tx_descs == NULL) {
1914                 CERROR("Can't allocate %d tx descriptors\n", size);
1915                 ps->ps_pool_destroy(pool);
1916                 return -ENOMEM;
1917         }
1918
1919         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1920
1921         for (i = 0; i < size; i++) {
1922                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1923
1924                 tx->tx_pool = tpo;
1925                 if (ps->ps_net->ibn_with_fmr){
1926                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1927                                      sizeof(*tx->tx_pages));
1928                         if (tx->tx_pages == NULL)
1929                                 break;
1930                 }
1931
1932                 LIBCFS_ALLOC(tx->tx_frags,
1933                              IBLND_MAX_RDMA_FRAGS *
1934                              sizeof(*tx->tx_frags));
1935                 if (tx->tx_frags == NULL)
1936                         break;
1937
1938                 LIBCFS_ALLOC(tx->tx_wrq,
1939                              (1 + IBLND_MAX_RDMA_FRAGS) *
1940                              sizeof(*tx->tx_wrq));
1941                 if (tx->tx_wrq == NULL)
1942                         break;
1943
1944                 LIBCFS_ALLOC(tx->tx_sge,
1945                              (1 + IBLND_MAX_RDMA_FRAGS) *
1946                              sizeof(*tx->tx_sge));
1947                 if (tx->tx_sge == NULL)
1948                         break;
1949
1950                 LIBCFS_ALLOC(tx->tx_rd,
1951                              offsetof(kib_rdma_desc_t,
1952                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1953                 if (tx->tx_rd == NULL)
1954                         break;
1955         }
1956
1957         if (i == size) {
1958                 kiblnd_map_tx_pool(tpo);
1959                 *pp_po = pool;
1960                 return 0;
1961         }
1962
1963         ps->ps_pool_destroy(pool);
1964         return -ENOMEM;
1965 }
1966
1967 static void
1968 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
1969 {
1970         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1971                                              tps_poolset);
1972         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
1973
1974         tx->tx_cookie = tps->tps_next_tx_cookie ++;
1975 }
1976
1977 void
1978 kiblnd_ni_fini_pools(kib_net_t *net)
1979 {
1980         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1981         if (net->ibn_with_fmr)
1982                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1983         else if (net->ibn_with_pmr)
1984                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1985 }
1986
1987 int
1988 kiblnd_net_init_pools(kib_net_t *net)
1989 {
1990         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1991         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1992         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
1993         int                rc;
1994
1995         if (*kiblnd_tunables.kib_fmr_pool_size <
1996             *kiblnd_tunables.kib_ntx / 4) {
1997                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1998                        *kiblnd_tunables.kib_fmr_pool_size,
1999                        *kiblnd_tunables.kib_ntx / 4);
2000                 return -EINVAL;
2001         }
2002
2003         if (*kiblnd_tunables.kib_pmr_pool_size <
2004             *kiblnd_tunables.kib_ntx / 4) {
2005                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2006                        *kiblnd_tunables.kib_pmr_pool_size,
2007                        *kiblnd_tunables.kib_ntx / 4);
2008                 return -EINVAL;
2009         }
2010
2011         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2012             net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2013                                            * so we always create FMR/PMR pool and
2014                                            * map-on-demand if premapping failed */
2015                 rc = kiblnd_init_fmr_pool_set(fps, net);
2016                 if (rc == 0) {
2017                         net->ibn_with_fmr = 1;
2018                 } else if (rc == -ENOSYS) {
2019                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2020                                                   *kiblnd_tunables.kib_pmr_pool_size,
2021                                                   kiblnd_create_pmr_pool,
2022                                                   kiblnd_destroy_pmr_pool,
2023                                                   NULL, NULL);
2024                         if (rc == 0)
2025                                 net->ibn_with_pmr = 1;
2026                 }
2027                 if (rc != 0)
2028                         return rc;
2029         }
2030
2031         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2032                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2033                                   kiblnd_tx_init, NULL);
2034         if (rc == 0)
2035                 return 0;
2036
2037         if (net->ibn_with_fmr)
2038                 kiblnd_fini_fmr_pool_set(fps);
2039         else if (net->ibn_with_pmr)
2040                 kiblnd_fini_pool_set(&pps->pps_poolset);
2041
2042         return rc;
2043 }
2044
2045 void
2046 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2047 {
2048         int     i;
2049
2050         if (ibdev->ibd_mrs == NULL)
2051                 return;
2052
2053         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2054                 if (ibdev->ibd_mrs[i] == NULL)
2055                         break;
2056
2057                 ib_dereg_mr(ibdev->ibd_mrs[i]);
2058         }
2059
2060         LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2061         ibdev->ibd_mrs = NULL;
2062 }
2063
2064 static int
2065 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2066 {
2067         struct ib_device_attr *attr;
2068         int                    rc;
2069
2070         /* It's safe to assume a HCA can handle a page size
2071          * matching that of the native system */
2072         ibdev->ibd_page_shift = PAGE_SHIFT;
2073         ibdev->ibd_page_size  = 1 << PAGE_SHIFT;
2074         ibdev->ibd_page_mask  = ~((__u64)ibdev->ibd_page_size - 1);
2075
2076         LIBCFS_ALLOC(attr, sizeof(*attr));
2077         if (attr == NULL) {
2078                 CERROR("Out of memory\n");
2079                 return -ENOMEM;
2080         }
2081
2082         rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2083         if (rc == 0)
2084                 ibdev->ibd_mr_size = attr->max_mr_size;
2085
2086         LIBCFS_FREE(attr, sizeof(*attr));
2087
2088         if (rc != 0) {
2089                 CERROR("Failed to query IB device: %d\n", rc);
2090                 return rc;
2091         }
2092
2093 #ifdef HAVE_OFED_TRANSPORT_IWARP
2094         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2095          * and we have reported the bug, remove these in the future when Chelsio
2096          * bug got fixed. */
2097         if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2098             RDMA_TRANSPORT_IWARP)
2099                 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2100 #endif
2101
2102         if (ibdev->ibd_mr_size == ~0ULL) {
2103                 ibdev->ibd_mr_shift = 64;
2104                 return 0;
2105         }
2106
2107         for (ibdev->ibd_mr_shift = 0;
2108              ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2109                 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2110                     ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2111                         return 0;
2112         }
2113
2114         CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2115         return -EINVAL;
2116 }
2117
2118 int
2119 kiblnd_dev_setup(kib_dev_t *ibdev)
2120 {
2121         struct ib_mr *mr;
2122         int           i;
2123         int           rc;
2124         __u64         mm_size;
2125         __u64         mr_size;
2126         int           acflags = IB_ACCESS_LOCAL_WRITE |
2127                                 IB_ACCESS_REMOTE_WRITE;
2128
2129         rc = kiblnd_dev_get_attr(ibdev);
2130         if (rc != 0)
2131                 return rc;
2132
2133         if (ibdev->ibd_mr_shift == 64) {
2134                 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2135                 if (ibdev->ibd_mrs == NULL) {
2136                         CERROR("Failed to allocate MRs table\n");
2137                         return -ENOMEM;
2138                 }
2139
2140                 ibdev->ibd_mrs[0] = NULL;
2141                 ibdev->ibd_nmrs   = 1;
2142
2143                 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2144                 if (IS_ERR(mr)) {
2145                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2146                         kiblnd_dev_cleanup(ibdev);
2147                         return PTR_ERR(mr);
2148                 }
2149
2150                 ibdev->ibd_mrs[0] = mr;
2151
2152                 goto out;
2153         }
2154
2155         mr_size = (1ULL << ibdev->ibd_mr_shift);
2156         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2157
2158         ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2159
2160         if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2161                 /* it's 4T..., assume we will re-code at that time */
2162                 CERROR("Can't support memory size: x"LPX64
2163                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2164                 return -EINVAL;
2165         }
2166
2167         /* create an array of MRs to cover all memory */
2168         LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2169         if (ibdev->ibd_mrs == NULL) {
2170                 CERROR("Failed to allocate MRs' table\n");
2171                 return -ENOMEM;
2172         }
2173
2174         memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2175
2176         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2177                 struct ib_phys_buf ipb;
2178                 __u64              iova;
2179
2180                 ipb.size = ibdev->ibd_mr_size;
2181                 ipb.addr = i * mr_size;
2182                 iova     = ipb.addr;
2183
2184                 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2185                 if (IS_ERR(mr)) {
2186                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2187                                " size "LPX64" : %ld\n",
2188                                ipb.addr, ipb.size, PTR_ERR(mr));
2189                         kiblnd_dev_cleanup(ibdev);
2190                         return PTR_ERR(mr);
2191                 }
2192
2193                 LASSERT (iova == ipb.addr);
2194
2195                 ibdev->ibd_mrs[i] = mr;
2196         }
2197
2198 out:
2199         CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2200                           LPX64", array size: %d\n",
2201                           ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2202
2203         cfs_list_add_tail(&ibdev->ibd_list,
2204                       &kiblnd_data.kib_devs);
2205         return 0;
2206 }
2207
2208 void
2209 kiblnd_destroy_dev (kib_dev_t *dev)
2210 {
2211         LASSERT (dev->ibd_nnets == 0);
2212
2213         if (!cfs_list_empty(&dev->ibd_list)) /* on kib_devs? */
2214                 cfs_list_del_init(&dev->ibd_list);
2215
2216         kiblnd_dev_cleanup(dev);
2217
2218         if (dev->ibd_pd != NULL)
2219                 ib_dealloc_pd(dev->ibd_pd);
2220
2221         if (dev->ibd_cmid != NULL)
2222                 rdma_destroy_id(dev->ibd_cmid);
2223
2224         LIBCFS_FREE(dev, sizeof(*dev));
2225 }
2226
2227 void
2228 kiblnd_base_shutdown (void)
2229 {
2230         int i;
2231
2232         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2233
2234         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2235                cfs_atomic_read(&libcfs_kmemory));
2236
2237         switch (kiblnd_data.kib_init) {
2238         default:
2239                 LBUG();
2240
2241         case IBLND_INIT_ALL:
2242         case IBLND_INIT_DATA:
2243                 LASSERT (kiblnd_data.kib_peers != NULL);
2244                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2245                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2246                 }
2247                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2248                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2249
2250                 /* flag threads to terminate; wake and wait for them to die */
2251                 kiblnd_data.kib_shutdown = 1;
2252                 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2253                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2254
2255                 i = 2;
2256                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2257                         i++;
2258                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2259                                "Waiting for %d threads to terminate\n",
2260                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2261                         cfs_pause(cfs_time_seconds(1));
2262                 }
2263
2264                 /* fall through */
2265
2266         case IBLND_INIT_NOTHING:
2267                 break;
2268         }
2269
2270         if (kiblnd_data.kib_peers != NULL)
2271                 LIBCFS_FREE(kiblnd_data.kib_peers,
2272                             sizeof(cfs_list_t) *
2273                             kiblnd_data.kib_peer_hash_size);
2274
2275         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2276                cfs_atomic_read(&libcfs_kmemory));
2277
2278         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2279         PORTAL_MODULE_UNUSE;
2280 }
2281
2282 void
2283 kiblnd_shutdown (lnet_ni_t *ni)
2284 {
2285         kib_net_t        *net = ni->ni_data;
2286         cfs_rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2287         int               i;
2288         unsigned long     flags;
2289
2290         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2291
2292         if (net == NULL)
2293                 goto out;
2294
2295         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2296                cfs_atomic_read(&libcfs_kmemory));
2297
2298         cfs_write_lock_irqsave(g_lock, flags);
2299         net->ibn_shutdown = 1;
2300         cfs_write_unlock_irqrestore(g_lock, flags);
2301
2302         switch (net->ibn_init) {
2303         default:
2304                 LBUG();
2305
2306         case IBLND_INIT_ALL:
2307                 /* nuke all existing peers within this net */
2308                 kiblnd_del_peer(ni, LNET_NID_ANY);
2309
2310                 /* Wait for all peer state to clean up */
2311                 i = 2;
2312                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2313                         i++;
2314                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2315                                "%s: waiting for %d peers to disconnect\n",
2316                                libcfs_nid2str(ni->ni_nid),
2317                                cfs_atomic_read(&net->ibn_npeers));
2318                         cfs_pause(cfs_time_seconds(1));
2319                 }
2320
2321                 kiblnd_ni_fini_pools(net);
2322
2323                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2324                 net->ibn_dev->ibd_nnets--;
2325
2326                 /* fall through */
2327
2328         case IBLND_INIT_NOTHING:
2329                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2330
2331                 if (net->ibn_dev != NULL &&
2332                     net->ibn_dev->ibd_nnets == 0)
2333                         kiblnd_destroy_dev(net->ibn_dev);
2334
2335                 break;
2336         }
2337
2338         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2339                cfs_atomic_read(&libcfs_kmemory));
2340
2341         net->ibn_init = IBLND_INIT_NOTHING;
2342         ni->ni_data = NULL;
2343
2344         LIBCFS_FREE(net, sizeof(*net));
2345
2346 out:
2347         if (cfs_list_empty(&kiblnd_data.kib_devs))
2348                 kiblnd_base_shutdown();
2349         return;
2350 }
2351
2352 int
2353 kiblnd_base_startup (void)
2354 {
2355         int i;
2356         int rc;
2357
2358         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2359
2360         PORTAL_MODULE_USE;
2361         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2362
2363         cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2364
2365         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2366
2367         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2368         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2369                      sizeof(cfs_list_t) *
2370                             kiblnd_data.kib_peer_hash_size);
2371         if (kiblnd_data.kib_peers == NULL) {
2372                 goto failed;
2373         }
2374         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2375                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2376
2377         cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2378         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2379         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2380         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2381
2382         cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2383         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2384         cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
2385
2386         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2387
2388         /* lists/ptrs/locks initialised */
2389         kiblnd_data.kib_init = IBLND_INIT_DATA;
2390         /*****************************************************/
2391
2392         for (i = 0; i < IBLND_N_SCHED; i++) {
2393                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2394                 if (rc != 0) {
2395                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2396                                i, rc);
2397                         goto failed;
2398                 }
2399         }
2400
2401         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2402         if (rc != 0) {
2403                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2404                 goto failed;
2405         }
2406
2407         /* flag everything initialised */
2408         kiblnd_data.kib_init = IBLND_INIT_ALL;
2409         /*****************************************************/
2410
2411         return 0;
2412
2413  failed:
2414         kiblnd_base_shutdown();
2415         return -ENETDOWN;
2416 }
2417
2418 int
2419 kiblnd_startup (lnet_ni_t *ni)
2420 {
2421         char                     *ifname;
2422         kib_dev_t                *ibdev = NULL;
2423         kib_net_t                *net;
2424         cfs_list_t               *tmp;
2425         struct timeval            tv;
2426         int                       rc;
2427
2428         LASSERT (ni->ni_lnd == &the_o2iblnd);
2429
2430         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2431                 rc = kiblnd_base_startup();
2432                 if (rc != 0)
2433                         return rc;
2434         }
2435
2436         LIBCFS_ALLOC(net, sizeof(*net));
2437         ni->ni_data = net;
2438         if (net == NULL)
2439                 goto failed;
2440
2441         memset(net, 0, sizeof(*net));
2442
2443         cfs_gettimeofday(&tv);
2444         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2445
2446         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2447         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2448         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2449         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2450
2451         if (ni->ni_interfaces[0] != NULL) {
2452                 /* Use the IPoIB interface specified in 'networks=' */
2453
2454                 CLASSERT (LNET_MAX_INTERFACES > 1);
2455                 if (ni->ni_interfaces[1] != NULL) {
2456                         CERROR("Multiple interfaces not supported\n");
2457                         goto failed;
2458                 }
2459
2460                 ifname = ni->ni_interfaces[0];
2461         } else {
2462                 ifname = *kiblnd_tunables.kib_default_ipif;
2463         }
2464
2465         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2466                 CERROR("IPoIB interface name too long: %s\n", ifname);
2467                 goto failed;
2468         }
2469
2470         cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2471                 ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
2472
2473                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2474                         break;
2475
2476                 ibdev = NULL;
2477         }
2478
2479         if (ibdev == NULL) {
2480                 __u32                     ip;
2481                 __u32                     netmask;
2482                 int                       up;
2483                 struct rdma_cm_id        *id;
2484                 struct ib_pd             *pd;
2485                 struct sockaddr_in        addr;
2486
2487                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2488                 if (rc != 0) {
2489                         CERROR("Can't query IPoIB interface %s: %d\n",
2490                                ifname, rc);
2491                         goto failed;
2492                 }
2493
2494                 if (!up) {
2495                         CERROR("Can't query IPoIB interface %s: it's down\n",
2496                                ifname);
2497                         goto failed;
2498                 }
2499
2500                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2501                 if (ibdev == NULL)
2502                         goto failed;
2503
2504                 memset(ibdev, 0, sizeof(*ibdev));
2505
2506                 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2507                 ibdev->ibd_ifip = ip;
2508                 strcpy(&ibdev->ibd_ifname[0], ifname);
2509
2510                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2511                 if (IS_ERR(id)) {
2512                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2513                         goto failed;
2514                 }
2515
2516                 ibdev->ibd_cmid = id;
2517
2518                 memset(&addr, 0, sizeof(addr));
2519                 addr.sin_family      = AF_INET;
2520                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2521                 addr.sin_addr.s_addr = htonl(ip);
2522
2523                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2524                 if (rc != 0) {
2525                         CERROR("Can't bind to %s: %d\n", ifname, rc);
2526                         goto failed;
2527                 }
2528
2529                 /* Binding should have assigned me an IB device */
2530                 LASSERT (id->device != NULL);
2531                 CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n",
2532                        ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service,
2533                        id->device->name);
2534
2535                 pd = ib_alloc_pd(id->device);
2536                 if (IS_ERR(pd)) {
2537                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2538                         goto failed;
2539                 }
2540
2541                 ibdev->ibd_pd = pd;
2542
2543                 rc = rdma_listen(id, 256);
2544                 if (rc != 0) {
2545                         CERROR("Can't start listener: %d\n", rc);
2546                         goto failed;
2547                 }
2548
2549                 rc = kiblnd_dev_setup(ibdev);
2550                 if (rc != 0) {
2551                         CERROR("Can't setup device: %d\n", rc);
2552                         goto failed;
2553                 }
2554         }
2555
2556         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2557         net->ibn_dev = ibdev;
2558
2559         rc = kiblnd_net_init_pools(net);
2560         if (rc != 0) {
2561                 CERROR("Failed to initialize NI pools: %d\n", rc);
2562                 goto failed;
2563         }
2564         ibdev->ibd_nnets++;
2565         net->ibn_init = IBLND_INIT_ALL;
2566
2567         return 0;
2568
2569 failed:
2570         if (net->ibn_dev == NULL && ibdev != NULL)
2571                 kiblnd_destroy_dev(ibdev);
2572
2573         kiblnd_shutdown(ni);
2574
2575         CDEBUG(D_NET, "kiblnd_startup failed\n");
2576         return -ENETDOWN;
2577 }
2578
2579 void __exit
2580 kiblnd_module_fini (void)
2581 {
2582         lnet_unregister_lnd(&the_o2iblnd);
2583         kiblnd_tunables_fini();
2584 }
2585
2586 int __init
2587 kiblnd_module_init (void)
2588 {
2589         int    rc;
2590
2591         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2592         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2593                   <= IBLND_MSG_SIZE);
2594         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2595                   <= IBLND_MSG_SIZE);
2596
2597         rc = kiblnd_tunables_init();
2598         if (rc != 0)
2599                 return rc;
2600
2601         lnet_register_lnd(&the_o2iblnd);
2602
2603         return 0;
2604 }
2605
2606 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2607 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2608 MODULE_LICENSE("GPL");
2609
2610 module_init(kiblnd_module_init);
2611 module_exit(kiblnd_module_fini);