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 (c) 2007, 2010, Oracle and/or its affiliates. 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_noops:\n");
 627         cfs_list_for_each(tmp, &conn->ibc_tx_noops)
 628                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 629
 630         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 631         cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 632                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 633
 634         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 635         cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 636                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 637
 638         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 639         cfs_list_for_each(tmp, &conn->ibc_tx_queue)
 640                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 641
 642         CDEBUG(D_CONSOLE, "   active_txs:\n");
 643         cfs_list_for_each(tmp, &conn->ibc_active_txs)
 644                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 645
 646         CDEBUG(D_CONSOLE, "   rxs:\n");
 647         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 648                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 649
 650         cfs_spin_unlock(&conn->ibc_lock);
 651 }
 652
 653 int
 654 kiblnd_translate_mtu(int value)
 655 {
 656         switch (value) {
 657         default:
 658                 return -1;
 659         case 0:
 660                 return 0;
 661         case 256:
 662                 return IB_MTU_256;
 663         case 512:
 664                 return IB_MTU_512;
 665         case 1024:
 666                 return IB_MTU_1024;
 667         case 2048:
 668                 return IB_MTU_2048;
 669         case 4096:
 670                 return IB_MTU_4096;
 671         }
 672 }
 673
 674 static void
 675 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 676 {
 677         int           mtu;
 678
 679         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 680         if (cmid->route.path_rec == NULL)
 681                 return;
 682
 683         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 684         LASSERT (mtu >= 0);
 685         if (mtu != 0)
 686                 cmid->route.path_rec->mtu = mtu;
 687 }
 688
 689 kib_conn_t *
 690 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 691                    int state, int version)
 692 {
 693         /* CAVEAT EMPTOR:
 694          * If the new conn is created successfully it takes over the caller's
 695          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 696          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 697          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 698          * to destroy 'cmid' here since I'm called from the CM which still has
 699          * its ref on 'cmid'). */
 700         cfs_rwlock_t           *glock = &kiblnd_data.kib_global_lock;
 701         kib_net_t              *net = peer->ibp_ni->ni_data;
 702         kib_dev_t              *dev = net->ibn_dev;
 703         struct ib_qp_init_attr *init_qp_attr;
 704         kib_conn_t             *conn;
 705         struct ib_cq           *cq;
 706         unsigned long           flags;
 707         int                     rc;
 708         int                     i;
 709
 710         LASSERT (net != NULL);
 711         LASSERT (!cfs_in_interrupt());
 712
 713         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 714         if (init_qp_attr == NULL) {
 715                 CERROR("Can't allocate qp_attr for %s\n",
 716                        libcfs_nid2str(peer->ibp_nid));
 717                 goto failed_0;
 718         }
 719
 720         LIBCFS_ALLOC(conn, sizeof(*conn));
 721         if (conn == NULL) {
 722                 CERROR("Can't allocate connection for %s\n",
 723                        libcfs_nid2str(peer->ibp_nid));
 724                 goto failed_1;
 725         }
 726
 727         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 728
 729         conn->ibc_state = IBLND_CONN_INIT;
 730         conn->ibc_version = version;
 731         conn->ibc_peer = peer;                  /* I take the caller's ref */
 732         cmid->context = conn;                   /* for future CM callbacks */
 733         conn->ibc_cmid = cmid;
 734
 735         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 736         CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops);
 737         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 738         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 739         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 740         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 741         cfs_spin_lock_init(&conn->ibc_lock);
 742
 743         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 744         if (conn->ibc_connvars == NULL) {
 745                 CERROR("Can't allocate in-progress connection state\n");
 746                 goto failed_2;
 747         }
 748         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 749
 750         cfs_write_lock_irqsave(glock, flags);
 751         if (dev->ibd_failover) {
 752                 cfs_write_unlock_irqrestore(glock, flags);
 753                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 754                 goto failed_2;
 755         }
 756
 757         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 758                 /* wakeup failover thread and teardown connection */
 759                 if (kiblnd_dev_can_failover(dev)) {
 760                         cfs_list_add_tail(&dev->ibd_fail_list,
 761                                       &kiblnd_data.kib_failed_devs);
 762                         cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
 763                 }
 764
 765                 cfs_write_unlock_irqrestore(glock, flags);
 766                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 767                        cmid->device->name, dev->ibd_ifname);
 768                 goto failed_2;
 769         }
 770
 771         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 772         conn->ibc_hdev = dev->ibd_hdev;
 773
 774         kiblnd_setup_mtu_locked(cmid);
 775
 776         cfs_write_unlock_irqrestore(glock, flags);
 777
 778         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 779         if (conn->ibc_rxs == NULL) {
 780                 CERROR("Cannot allocate RX buffers\n");
 781                 goto failed_2;
 782         }
 783         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 784
 785         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
 786                                 IBLND_RX_MSG_PAGES(version));
 787         if (rc != 0)
 788                 goto failed_2;
 789
 790         kiblnd_map_rx_descs(conn);
 791
 792 #ifdef HAVE_OFED_IB_COMP_VECTOR
 793         cq = ib_create_cq(cmid->device,
 794                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 795                           IBLND_CQ_ENTRIES(version), 0);
 796 #else
 797         cq = ib_create_cq(cmid->device,
 798                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 799                           IBLND_CQ_ENTRIES(version));
 800 #endif
 801         if (IS_ERR(cq)) {
 802                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 803                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 804                 goto failed_2;
 805         }
 806
 807         conn->ibc_cq = cq;
 808
 809         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 810         if (rc != 0) {
 811                 CERROR("Can't request completion notificiation: %d\n", rc);
 812                 goto failed_2;
 813         }
 814
 815         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 816         init_qp_attr->event_handler = kiblnd_qp_event;
 817         init_qp_attr->qp_context = conn;
 818         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 819         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 820         init_qp_attr->cap.max_send_sge = 1;
 821         init_qp_attr->cap.max_recv_sge = 1;
 822         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 823         init_qp_attr->qp_type = IB_QPT_RC;
 824         init_qp_attr->send_cq = cq;
 825         init_qp_attr->recv_cq = cq;
 826
 827         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 828         if (rc != 0) {
 829                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 830                        rc, init_qp_attr->cap.max_send_wr,
 831                        init_qp_attr->cap.max_recv_wr);
 832                 goto failed_2;
 833         }
 834
 835         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 836
 837         /* 1 ref for caller and each rxmsg */
 838         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 839         conn->ibc_nrx = IBLND_RX_MSGS(version);
 840
 841         /* post receives */
 842         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 843                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 844                                     IBLND_POSTRX_NO_CREDIT);
 845                 if (rc != 0) {
 846                         CERROR("Can't post rxmsg: %d\n", rc);
 847
 848                         /* Make posted receives complete */
 849                         kiblnd_abort_receives(conn);
 850
 851                         /* correct # of posted buffers
 852                          * NB locking needed now I'm racing with completion */
 853                         cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
 854                                               flags);
 855                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 856                         cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 857                                                    flags);
 858
 859                         /* cmid will be destroyed by CM(ofed) after cm_callback
 860                          * returned, so we can't refer it anymore
 861                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 862                         rdma_destroy_qp(conn->ibc_cmid);
 863                         conn->ibc_cmid = NULL;
 864
 865                         /* Drop my own and unused rxbuffer refcounts */
 866                         while (i++ <= IBLND_RX_MSGS(version))
 867                                 kiblnd_conn_decref(conn);
 868
 869                         return NULL;
 870                 }
 871         }
 872
 873         /* Init successful! */
 874         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 875                  state == IBLND_CONN_PASSIVE_WAIT);
 876         conn->ibc_state = state;
 877
 878         /* 1 more conn */
 879         cfs_atomic_inc(&net->ibn_nconns);
 880         return conn;
 881
 882  failed_2:
 883         kiblnd_destroy_conn(conn);
 884  failed_1:
 885         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 886  failed_0:
 887         return NULL;
 888 }
 889
 890 void
 891 kiblnd_destroy_conn (kib_conn_t *conn)
 892 {
 893         struct rdma_cm_id *cmid = conn->ibc_cmid;
 894         kib_peer_t        *peer = conn->ibc_peer;
 895         int                rc;
 896
 897         LASSERT (!cfs_in_interrupt());
 898         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 899         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 900         LASSERT (cfs_list_empty(&conn->ibc_tx_noops));
 901         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 902         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 903         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 904         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 905         LASSERT (conn->ibc_noops_posted == 0);
 906         LASSERT (conn->ibc_nsends_posted == 0);
 907
 908         switch (conn->ibc_state) {
 909         default:
 910                 /* conn must be completely disengaged from the network */
 911                 LBUG();
 912
 913         case IBLND_CONN_DISCONNECTED:
 914                 /* connvars should have been freed already */
 915                 LASSERT (conn->ibc_connvars == NULL);
 916                 break;
 917
 918         case IBLND_CONN_INIT:
 919                 break;
 920         }
 921
 922         /* conn->ibc_cmid might be destroyed by CM already */
 923         if (cmid != NULL && cmid->qp != NULL)
 924                 rdma_destroy_qp(cmid);
 925
 926         if (conn->ibc_cq != NULL) {
 927                 rc = ib_destroy_cq(conn->ibc_cq);
 928                 if (rc != 0)
 929                         CWARN("Error destroying CQ: %d\n", rc);
 930         }
 931
 932         if (conn->ibc_rx_pages != NULL)
 933                 kiblnd_unmap_rx_descs(conn);
 934
 935         if (conn->ibc_rxs != NULL) {
 936                 LIBCFS_FREE(conn->ibc_rxs,
 937                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 938         }
 939
 940         if (conn->ibc_connvars != NULL)
 941                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 942
 943         if (conn->ibc_hdev != NULL)
 944                 kiblnd_hdev_decref(conn->ibc_hdev);
 945
 946         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 947         if (conn->ibc_state != IBLND_CONN_INIT) {
 948                 kib_net_t *net = peer->ibp_ni->ni_data;
 949
 950                 kiblnd_peer_decref(peer);
 951                 rdma_destroy_id(cmid);
 952                 cfs_atomic_dec(&net->ibn_nconns);
 953         }
 954
 955         LIBCFS_FREE(conn, sizeof(*conn));
 956 }
 957
 958 int
 959 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 960 {
 961         kib_conn_t             *conn;
 962         cfs_list_t             *ctmp;
 963         cfs_list_t             *cnxt;
 964         int                     count = 0;
 965
 966         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 967                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 968
 969                 CDEBUG(D_NET, "Closing conn -> %s, "
 970                               "version: %x, reason: %d\n",
 971                        libcfs_nid2str(peer->ibp_nid),
 972                        conn->ibc_version, why);
 973
 974                 kiblnd_close_conn_locked(conn, why);
 975                 count++;
 976         }
 977
 978         return count;
 979 }
 980
 981 int
 982 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 983                                  int version, __u64 incarnation)
 984 {
 985         kib_conn_t             *conn;
 986         cfs_list_t             *ctmp;
 987         cfs_list_t             *cnxt;
 988         int                     count = 0;
 989
 990         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 991                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 992
 993                 if (conn->ibc_version     == version &&
 994                     conn->ibc_incarnation == incarnation)
 995                         continue;
 996
 997                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
 998                               "incarnation:"LPX64"(%x, "LPX64")\n",
 999                        libcfs_nid2str(peer->ibp_nid),
1000                        conn->ibc_version, conn->ibc_incarnation,
1001                        version, incarnation);
1002
1003                 kiblnd_close_conn_locked(conn, -ESTALE);
1004                 count++;
1005         }
1006
1007         return count;
1008 }
1009
1010 int
1011 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1012 {
1013         kib_peer_t             *peer;
1014         cfs_list_t             *ptmp;
1015         cfs_list_t             *pnxt;
1016         int                     lo;
1017         int                     hi;
1018         int                     i;
1019         unsigned long           flags;
1020         int                     count = 0;
1021
1022         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1023
1024         if (nid != LNET_NID_ANY)
1025                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1026         else {
1027                 lo = 0;
1028                 hi = kiblnd_data.kib_peer_hash_size - 1;
1029         }
1030
1031         for (i = lo; i <= hi; i++) {
1032                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1033
1034                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1035                         LASSERT (peer->ibp_connecting > 0 ||
1036                                  peer->ibp_accepting > 0 ||
1037                                  !cfs_list_empty(&peer->ibp_conns));
1038
1039                         if (peer->ibp_ni != ni)
1040                                 continue;
1041
1042                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1043                                 continue;
1044
1045                         count += kiblnd_close_peer_conns_locked(peer, 0);
1046                 }
1047         }
1048
1049         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1050
1051         /* wildcards always succeed */
1052         if (nid == LNET_NID_ANY)
1053                 return 0;
1054
1055         return (count == 0) ? -ENOENT : 0;
1056 }
1057
1058 int
1059 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1060 {
1061         struct libcfs_ioctl_data *data = arg;
1062         int                       rc = -EINVAL;
1063
1064         switch(cmd) {
1065         case IOC_LIBCFS_GET_PEER: {
1066                 lnet_nid_t   nid = 0;
1067                 int          count = 0;
1068
1069                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1070                                           &nid, &count);
1071                 data->ioc_nid    = nid;
1072                 data->ioc_count  = count;
1073                 break;
1074         }
1075
1076         case IOC_LIBCFS_DEL_PEER: {
1077                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1078                 break;
1079         }
1080         case IOC_LIBCFS_GET_CONN: {
1081                 kib_conn_t *conn;
1082
1083                 rc = 0;
1084                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1085                 if (conn == NULL) {
1086                         rc = -ENOENT;
1087                         break;
1088                 }
1089
1090                 LASSERT (conn->ibc_cmid != NULL);
1091                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1092                 if (conn->ibc_cmid->route.path_rec == NULL)
1093                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1094                 else
1095                         data->ioc_u32[0] =
1096                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1097                 kiblnd_conn_decref(conn);
1098                 break;
1099         }
1100         case IOC_LIBCFS_CLOSE_CONNECTION: {
1101                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1102                 break;
1103         }
1104
1105         default:
1106                 break;
1107         }
1108
1109         return rc;
1110 }
1111
1112 void
1113 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1114 {
1115         cfs_time_t         last_alive = 0;
1116         cfs_time_t         now = cfs_time_current();
1117         cfs_rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1118         kib_peer_t        *peer;
1119         unsigned long      flags;
1120
1121         cfs_read_lock_irqsave(glock, flags);
1122
1123         peer = kiblnd_find_peer_locked(nid);
1124         if (peer != NULL) {
1125                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1126                          peer->ibp_accepting > 0 ||
1127                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1128                 last_alive = peer->ibp_last_alive;
1129         }
1130
1131         cfs_read_unlock_irqrestore(glock, flags);
1132
1133         if (last_alive != 0)
1134                 *when = last_alive;
1135
1136         /* peer is not persistent in hash, trigger peer creation
1137          * and connection establishment with a NULL tx */
1138         if (peer == NULL)
1139                 kiblnd_launch_tx(ni, NULL, nid);
1140
1141         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1142                libcfs_nid2str(nid), peer,
1143                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1144         return;
1145 }
1146
1147 void
1148 kiblnd_free_pages (kib_pages_t *p)
1149 {
1150         int         npages = p->ibp_npages;
1151         int         i;
1152
1153         for (i = 0; i < npages; i++)
1154                 if (p->ibp_pages[i] != NULL)
1155                         __free_page(p->ibp_pages[i]);
1156
1157         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1158 }
1159
1160 int
1161 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1162 {
1163         kib_pages_t   *p;
1164         int            i;
1165
1166         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1167         if (p == NULL) {
1168                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1169                 return -ENOMEM;
1170         }
1171
1172         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1173         p->ibp_npages = npages;
1174
1175         for (i = 0; i < npages; i++) {
1176                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1177                 if (p->ibp_pages[i] == NULL) {
1178                         CERROR("Can't allocate page %d of %d\n", i, npages);
1179                         kiblnd_free_pages(p);
1180                         return -ENOMEM;
1181                 }
1182         }
1183
1184         *pp = p;
1185         return 0;
1186 }
1187
1188 void
1189 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1190 {
1191         kib_rx_t *rx;
1192         int       i;
1193
1194         LASSERT (conn->ibc_rxs != NULL);
1195         LASSERT (conn->ibc_hdev != NULL);
1196
1197         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1198                 rx = &conn->ibc_rxs[i];
1199
1200                 LASSERT (rx->rx_nob >= 0); /* not posted */
1201
1202                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1203                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1204                                                           rx->rx_msgaddr),
1205                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1206         }
1207
1208         kiblnd_free_pages(conn->ibc_rx_pages);
1209
1210         conn->ibc_rx_pages = NULL;
1211 }
1212
1213 void
1214 kiblnd_map_rx_descs(kib_conn_t *conn)
1215 {
1216         kib_rx_t       *rx;
1217         struct page    *pg;
1218         int             pg_off;
1219         int             ipg;
1220         int             i;
1221
1222         for (pg_off = ipg = i = 0;
1223              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1224                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1225                 rx = &conn->ibc_rxs[i];
1226
1227                 rx->rx_conn = conn;
1228                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1229
1230                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1231                                                        rx->rx_msg, IBLND_MSG_SIZE,
1232                                                        DMA_FROM_DEVICE);
1233                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1234                                                    rx->rx_msgaddr));
1235                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1236
1237                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1238                        i, rx->rx_msg, rx->rx_msgaddr,
1239                        lnet_page2phys(pg) + pg_off);
1240
1241                 pg_off += IBLND_MSG_SIZE;
1242                 LASSERT (pg_off <= PAGE_SIZE);
1243
1244                 if (pg_off == PAGE_SIZE) {
1245                         pg_off = 0;
1246                         ipg++;
1247                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1248                 }
1249         }
1250 }
1251
1252 static void
1253 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1254 {
1255         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1256         kib_tx_t       *tx;
1257         int             i;
1258
1259         LASSERT (tpo->tpo_pool.po_allocated == 0);
1260
1261         if (hdev == NULL)
1262                 return;
1263
1264         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1265                 tx = &tpo->tpo_tx_descs[i];
1266                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1267                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1268                                                           tx->tx_msgaddr),
1269                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1270         }
1271
1272         kiblnd_hdev_decref(hdev);
1273         tpo->tpo_hdev = NULL;
1274 }
1275
1276 static kib_hca_dev_t *
1277 kiblnd_current_hdev(kib_dev_t *dev)
1278 {
1279         kib_hca_dev_t *hdev;
1280         unsigned long  flags;
1281         int            i = 0;
1282
1283         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1284         while (dev->ibd_failover) {
1285                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1286                 if (i++ % 50 == 0)
1287                         CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
1288                 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1289
1290                 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1291         }
1292
1293         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1294         hdev = dev->ibd_hdev;
1295
1296         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1297
1298         return hdev;
1299 }
1300
1301 static void
1302 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1303 {
1304         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1305         kib_pool_t     *pool  = &tpo->tpo_pool;
1306         kib_net_t      *net   = pool->po_owner->ps_net;
1307         kib_dev_t      *dev   = net->ibn_dev;
1308         struct page    *page;
1309         kib_tx_t       *tx;
1310         int             page_offset;
1311         int             ipage;
1312         int             i;
1313
1314         LASSERT (net != NULL);
1315
1316         /* pre-mapped messages are not bigger than 1 page */
1317         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1318
1319         /* No fancy arithmetic when we do the buffer calculations */
1320         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1321
1322         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1323
1324         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1325                 page = txpgs->ibp_pages[ipage];
1326                 tx = &tpo->tpo_tx_descs[i];
1327
1328                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1329                                            page_offset);
1330
1331                 tx->tx_msgaddr = kiblnd_dma_map_single(
1332                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1333                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1334                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1335                                                    tx->tx_msgaddr));
1336                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1337
1338                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1339
1340                 page_offset += IBLND_MSG_SIZE;
1341                 LASSERT (page_offset <= PAGE_SIZE);
1342
1343                 if (page_offset == PAGE_SIZE) {
1344                         page_offset = 0;
1345                         ipage++;
1346                         LASSERT (ipage <= txpgs->ibp_npages);
1347                 }
1348         }
1349 }
1350
1351 struct ib_mr *
1352 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1353 {
1354         __u64   index;
1355
1356         LASSERT (hdev->ibh_mrs[0] != NULL);
1357
1358         if (hdev->ibh_nmrs == 1)
1359                 return hdev->ibh_mrs[0];
1360
1361         index = addr >> hdev->ibh_mr_shift;
1362
1363         if (index <  hdev->ibh_nmrs &&
1364             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1365                 return hdev->ibh_mrs[index];
1366
1367         return NULL;
1368 }
1369
1370 struct ib_mr *
1371 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1372 {
1373         struct ib_mr *prev_mr;
1374         struct ib_mr *mr;
1375         int           i;
1376
1377         LASSERT (hdev->ibh_mrs[0] != NULL);
1378
1379         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1380             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1381                 return NULL;
1382
1383         if (hdev->ibh_nmrs == 1)
1384                 return hdev->ibh_mrs[0];
1385
1386         for (i = 0, mr = prev_mr = NULL;
1387              i < rd->rd_nfrags; i++) {
1388                 mr = kiblnd_find_dma_mr(hdev,
1389                                         rd->rd_frags[i].rf_addr,
1390                                         rd->rd_frags[i].rf_nob);
1391                 if (prev_mr == NULL)
1392                         prev_mr = mr;
1393
1394                 if (mr == NULL || prev_mr != mr) {
1395                         /* Can't covered by one single MR */
1396                         mr = NULL;
1397                         break;
1398                 }
1399         }
1400
1401         return mr;
1402 }
1403
1404 void
1405 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1406 {
1407         LASSERT (pool->fpo_map_count == 0);
1408
1409         if (pool->fpo_fmr_pool != NULL)
1410                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1411
1412         if (pool->fpo_hdev != NULL)
1413                 kiblnd_hdev_decref(pool->fpo_hdev);
1414
1415         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1416 }
1417
1418 void
1419 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1420 {
1421         kib_fmr_pool_t *pool;
1422
1423         while (!cfs_list_empty(head)) {
1424                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1425                 cfs_list_del(&pool->fpo_list);
1426                 kiblnd_destroy_fmr_pool(pool);
1427         }
1428 }
1429
1430 int
1431 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1432 {
1433         /* FMR pool for RDMA */
1434         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1435         kib_fmr_pool_t          *fpo;
1436         struct ib_fmr_pool_param param = {
1437                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1438                 .page_shift        = PAGE_SHIFT,
1439                 .access            = (IB_ACCESS_LOCAL_WRITE |
1440                                       IB_ACCESS_REMOTE_WRITE),
1441                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1442                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1443                 .flush_function    = NULL,
1444                 .flush_arg         = NULL,
1445                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1446         int rc;
1447
1448         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1449         if (fpo == NULL)
1450                 return -ENOMEM;
1451
1452         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1453         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1454         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1455         if (IS_ERR(fpo->fpo_fmr_pool)) {
1456                 CERROR("Failed to create FMR pool: %ld\n",
1457                        PTR_ERR(fpo->fpo_fmr_pool));
1458                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1459                 kiblnd_hdev_decref(fpo->fpo_hdev);
1460                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1461                 return rc;
1462         }
1463
1464         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1465         fpo->fpo_owner    = fps;
1466         *pp_fpo = fpo;
1467
1468         return 0;
1469 }
1470
1471 static void
1472 kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1473 {
1474         cfs_spin_lock(&fps->fps_lock);
1475
1476         while (!cfs_list_empty(&fps->fps_pool_list)) {
1477                 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1478                                                  kib_fmr_pool_t, fpo_list);
1479                 fpo->fpo_failed = 1;
1480                 cfs_list_del(&fpo->fpo_list);
1481                 if (fpo->fpo_map_count == 0)
1482                         cfs_list_add(&fpo->fpo_list, zombies);
1483                 else
1484                         cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1485         }
1486
1487         cfs_spin_unlock(&fps->fps_lock);
1488 }
1489
1490 static void
1491 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1492 {
1493         kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1494         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1495 }
1496
1497 static int
1498 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1499 {
1500         kib_fmr_pool_t *fpo;
1501         int             rc;
1502
1503         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1504
1505         fps->fps_net = net;
1506         cfs_spin_lock_init(&fps->fps_lock);
1507         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1508         CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1509         rc = kiblnd_create_fmr_pool(fps, &fpo);
1510         if (rc == 0)
1511                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1512
1513         return rc;
1514 }
1515
1516 static int
1517 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1518 {
1519         if (fpo->fpo_map_count != 0) /* still in use */
1520                 return 0;
1521         if (fpo->fpo_failed)
1522                 return 1;
1523         return cfs_time_aftereq(now, fpo->fpo_deadline);
1524 }
1525
1526 void
1527 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1528 {
1529         CFS_LIST_HEAD     (zombies);
1530         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1531         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1532         cfs_time_t         now = cfs_time_current();
1533         kib_fmr_pool_t    *tmp;
1534         int                rc;
1535
1536         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1537         LASSERT (rc == 0);
1538
1539         if (status != 0) {
1540                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1541                 LASSERT (rc == 0);
1542         }
1543
1544         fmr->fmr_pool = NULL;
1545         fmr->fmr_pfmr = NULL;
1546
1547         cfs_spin_lock(&fps->fps_lock);
1548         fpo->fpo_map_count --;  /* decref the pool */
1549
1550         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1551                 /* the first pool is persistent */
1552                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1553                         continue;
1554
1555                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1556                         cfs_list_move(&fpo->fpo_list, &zombies);
1557                         fps->fps_version ++;
1558                 }
1559         }
1560         cfs_spin_unlock(&fps->fps_lock);
1561
1562         if (!cfs_list_empty(&zombies))
1563                 kiblnd_destroy_fmr_pool_list(&zombies);
1564 }
1565
1566 int
1567 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1568                     __u64 iov, kib_fmr_t *fmr)
1569 {
1570         struct ib_pool_fmr *pfmr;
1571         kib_fmr_pool_t     *fpo;
1572         __u64               version;
1573         int                 rc;
1574
1575         LASSERT (fps->fps_net->ibn_with_fmr);
1576  again:
1577         cfs_spin_lock(&fps->fps_lock);
1578         version = fps->fps_version;
1579         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1580                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1581                 fpo->fpo_map_count ++;
1582                 cfs_spin_unlock(&fps->fps_lock);
1583
1584                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1585                                             pages, npages, iov);
1586                 if (likely(!IS_ERR(pfmr))) {
1587                         fmr->fmr_pool = fpo;
1588                         fmr->fmr_pfmr = pfmr;
1589                         return 0;
1590                 }
1591
1592                 cfs_spin_lock(&fps->fps_lock);
1593                 fpo->fpo_map_count --;
1594                 if (PTR_ERR(pfmr) != -EAGAIN) {
1595                         cfs_spin_unlock(&fps->fps_lock);
1596                         return PTR_ERR(pfmr);
1597                 }
1598
1599                 /* EAGAIN and ... */
1600                 if (version != fps->fps_version) {
1601                         cfs_spin_unlock(&fps->fps_lock);
1602                         goto again;
1603                 }
1604         }
1605
1606         if (fps->fps_increasing) {
1607                 cfs_spin_unlock(&fps->fps_lock);
1608                 CDEBUG(D_NET, "Another thread is allocating new "
1609                               "FMR pool, waiting for her to complete\n");
1610                 cfs_schedule();
1611                 goto again;
1612
1613         }
1614
1615         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1616                 /* someone failed recently */
1617                 cfs_spin_unlock(&fps->fps_lock);
1618                 return -EAGAIN;
1619         }
1620
1621         fps->fps_increasing = 1;
1622         cfs_spin_unlock(&fps->fps_lock);
1623
1624         CDEBUG(D_NET, "Allocate new FMR pool\n");
1625         rc = kiblnd_create_fmr_pool(fps, &fpo);
1626         cfs_spin_lock(&fps->fps_lock);
1627         fps->fps_increasing = 0;
1628         if (rc == 0) {
1629                 fps->fps_version ++;
1630                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1631         } else {
1632                 fps->fps_next_retry = cfs_time_shift(10);
1633         }
1634         cfs_spin_unlock(&fps->fps_lock);
1635
1636         goto again;
1637 }
1638
1639 static void
1640 kiblnd_fini_pool(kib_pool_t *pool)
1641 {
1642         LASSERT (cfs_list_empty(&pool->po_free_list));
1643         LASSERT (pool->po_allocated == 0);
1644
1645         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1646 }
1647
1648 static void
1649 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1650 {
1651         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1652
1653         memset(pool, 0, sizeof(kib_pool_t));
1654         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1655         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1656         pool->po_owner    = ps;
1657         pool->po_size     = size;
1658 }
1659
1660 void
1661 kiblnd_destroy_pool_list(cfs_list_t *head)
1662 {
1663         kib_pool_t *pool;
1664
1665         while (!cfs_list_empty(head)) {
1666                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1667                 cfs_list_del(&pool->po_list);
1668
1669                 LASSERT (pool->po_owner != NULL);
1670                 pool->po_owner->ps_pool_destroy(pool);
1671         }
1672 }
1673
1674 static void
1675 kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies)
1676 {
1677         cfs_spin_lock(&ps->ps_lock);
1678         while (!cfs_list_empty(&ps->ps_pool_list)) {
1679                 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1680                                             kib_pool_t, po_list);
1681                 po->po_failed = 1;
1682                 cfs_list_del(&po->po_list);
1683                 if (po->po_allocated == 0)
1684                         cfs_list_add(&po->po_list, zombies);
1685                 else
1686                         cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1687         }
1688         cfs_spin_unlock(&ps->ps_lock);
1689 }
1690
1691 static void
1692 kiblnd_fini_pool_set(kib_poolset_t *ps)
1693 {
1694         kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1695         kiblnd_destroy_pool_list(&ps->ps_pool_list);
1696 }
1697
1698 static int
1699 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1700                      char *name, int size,
1701                      kib_ps_pool_create_t po_create,
1702                      kib_ps_pool_destroy_t po_destroy,
1703                      kib_ps_node_init_t nd_init,
1704                      kib_ps_node_fini_t nd_fini)
1705 {
1706         kib_pool_t    *pool;
1707         int            rc;
1708
1709         memset(ps, 0, sizeof(kib_poolset_t));
1710
1711         ps->ps_net          = net;
1712         ps->ps_pool_create  = po_create;
1713         ps->ps_pool_destroy = po_destroy;
1714         ps->ps_node_init    = nd_init;
1715         ps->ps_node_fini    = nd_fini;
1716         ps->ps_pool_size    = size;
1717         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1718         cfs_spin_lock_init(&ps->ps_lock);
1719         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1720         CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1721
1722         rc = ps->ps_pool_create(ps, size, &pool);
1723         if (rc == 0)
1724                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1725         else
1726                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1727
1728         return rc;
1729 }
1730
1731 static int
1732 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1733 {
1734         if (pool->po_allocated != 0) /* still in use */
1735                 return 0;
1736         if (pool->po_failed)
1737                 return 1;
1738         return cfs_time_aftereq(now, pool->po_deadline);
1739 }
1740
1741 void
1742 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1743 {
1744         CFS_LIST_HEAD  (zombies);
1745         kib_poolset_t  *ps = pool->po_owner;
1746         kib_pool_t     *tmp;
1747         cfs_time_t      now = cfs_time_current();
1748
1749         cfs_spin_lock(&ps->ps_lock);
1750
1751         if (ps->ps_node_fini != NULL)
1752                 ps->ps_node_fini(pool, node);
1753
1754         LASSERT (pool->po_allocated > 0);
1755         cfs_list_add(node, &pool->po_free_list);
1756         pool->po_allocated --;
1757
1758         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1759                 /* the first pool is persistent */
1760                 if (ps->ps_pool_list.next == &pool->po_list)
1761                         continue;
1762
1763                 if (kiblnd_pool_is_idle(pool, now))
1764                         cfs_list_move(&pool->po_list, &zombies);
1765         }
1766         cfs_spin_unlock(&ps->ps_lock);
1767
1768         if (!cfs_list_empty(&zombies))
1769                 kiblnd_destroy_pool_list(&zombies);
1770 }
1771
1772 cfs_list_t *
1773 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1774 {
1775         cfs_list_t            *node;
1776         kib_pool_t            *pool;
1777         int                    rc;
1778
1779  again:
1780         cfs_spin_lock(&ps->ps_lock);
1781         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1782                 if (cfs_list_empty(&pool->po_free_list))
1783                         continue;
1784
1785                 pool->po_allocated ++;
1786                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1787                 node = pool->po_free_list.next;
1788                 cfs_list_del(node);
1789
1790                 if (ps->ps_node_init != NULL) {
1791                         /* still hold the lock */
1792                         ps->ps_node_init(pool, node);
1793                 }
1794                 cfs_spin_unlock(&ps->ps_lock);
1795                 return node;
1796         }
1797
1798         /* no available tx pool and ... */
1799         if (ps->ps_increasing) {
1800                 /* another thread is allocating a new pool */
1801                 cfs_spin_unlock(&ps->ps_lock);
1802                 CDEBUG(D_NET, "Another thread is allocating new "
1803                        "%s pool, waiting for her to complete\n",
1804                        ps->ps_name);
1805                 cfs_schedule();
1806                 goto again;
1807         }
1808
1809         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1810                 /* someone failed recently */
1811                 cfs_spin_unlock(&ps->ps_lock);
1812                 return NULL;
1813         }
1814
1815         ps->ps_increasing = 1;
1816         cfs_spin_unlock(&ps->ps_lock);
1817
1818         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1819
1820         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1821
1822         cfs_spin_lock(&ps->ps_lock);
1823         ps->ps_increasing = 0;
1824         if (rc == 0) {
1825                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1826         } else {
1827                 /* retry 10 seconds later */
1828                 ps->ps_next_retry = cfs_time_shift(10);
1829                 CERROR("Can't allocate new %s pool because out of memory\n",
1830                        ps->ps_name);
1831         }
1832         cfs_spin_unlock(&ps->ps_lock);
1833
1834         goto again;
1835 }
1836
1837 void
1838 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1839 {
1840         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1841         struct ib_mr        *mr  = pmr->pmr_mr;
1842
1843         pmr->pmr_mr = NULL;
1844         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1845         if (mr != NULL)
1846                 ib_dereg_mr(mr);
1847 }
1848
1849 int
1850 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1851                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1852 {
1853         kib_phys_mr_t *pmr;
1854         cfs_list_t    *node;
1855         int            rc;
1856         int            i;
1857
1858         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1859         if (node == NULL) {
1860                 CERROR("Failed to allocate PMR descriptor\n");
1861                 return -ENOMEM;
1862         }
1863
1864         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1865         if (pmr->pmr_pool->ppo_hdev != hdev) {
1866                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1867                 return -EAGAIN;
1868         }
1869
1870         for (i = 0; i < rd->rd_nfrags; i ++) {
1871                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1872                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1873         }
1874
1875         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1876                                      pmr->pmr_ipb, rd->rd_nfrags,
1877                                      IB_ACCESS_LOCAL_WRITE |
1878                                      IB_ACCESS_REMOTE_WRITE,
1879                                      iova);
1880         if (!IS_ERR(pmr->pmr_mr)) {
1881                 pmr->pmr_iova = *iova;
1882                 *pp_pmr = pmr;
1883                 return 0;
1884         }
1885
1886         rc = PTR_ERR(pmr->pmr_mr);
1887         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1888
1889         pmr->pmr_mr = NULL;
1890         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1891
1892         return rc;
1893 }
1894
1895 static void
1896 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1897 {
1898         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1899         kib_phys_mr_t  *pmr;
1900
1901         LASSERT (pool->po_allocated == 0);
1902
1903         while (!cfs_list_empty(&pool->po_free_list)) {
1904                 pmr = cfs_list_entry(pool->po_free_list.next,
1905                                      kib_phys_mr_t, pmr_list);
1906
1907                 LASSERT (pmr->pmr_mr == NULL);
1908                 cfs_list_del(&pmr->pmr_list);
1909
1910                 if (pmr->pmr_ipb != NULL) {
1911                         LIBCFS_FREE(pmr->pmr_ipb,
1912                                     IBLND_MAX_RDMA_FRAGS *
1913                                     sizeof(struct ib_phys_buf));
1914                 }
1915
1916                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1917         }
1918
1919         kiblnd_fini_pool(pool);
1920         if (ppo->ppo_hdev != NULL)
1921                 kiblnd_hdev_decref(ppo->ppo_hdev);
1922
1923         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1924 }
1925
1926 static int
1927 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1928 {
1929         kib_pmr_pool_t      *ppo;
1930         kib_pool_t          *pool;
1931         kib_phys_mr_t       *pmr;
1932         int                  i;
1933
1934         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1935         if (ppo == NULL) {
1936                 CERROR("Failed to allocate PMR pool\n");
1937                 return -ENOMEM;
1938         }
1939
1940         pool = &ppo->ppo_pool;
1941         kiblnd_init_pool(ps, pool, size);
1942
1943         for (i = 0; i < size; i++) {
1944                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1945                 if (pmr == NULL)
1946                         break;
1947
1948                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1949                 pmr->pmr_pool = ppo;
1950                 LIBCFS_ALLOC(pmr->pmr_ipb,
1951                              IBLND_MAX_RDMA_FRAGS *
1952                              sizeof(struct ib_phys_buf));
1953                 if (pmr->pmr_ipb == NULL)
1954                         break;
1955
1956                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1957         }
1958
1959         if (i < size) {
1960                 ps->ps_pool_destroy(pool);
1961                 return -ENOMEM;
1962         }
1963
1964         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1965         *pp_po = pool;
1966         return 0;
1967 }
1968
1969 static void
1970 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1971 {
1972         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1973         int             i;
1974
1975         LASSERT (pool->po_allocated == 0);
1976
1977         if (tpo->tpo_tx_pages != NULL) {
1978                 kiblnd_unmap_tx_pool(tpo);
1979                 kiblnd_free_pages(tpo->tpo_tx_pages);
1980         }
1981
1982         if (tpo->tpo_tx_descs == NULL)
1983                 goto out;
1984
1985         for (i = 0; i < pool->po_size; i++) {
1986                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1987
1988                 cfs_list_del(&tx->tx_list);
1989                 if (tx->tx_pages != NULL)
1990                         LIBCFS_FREE(tx->tx_pages,
1991                                     LNET_MAX_IOV *
1992                                     sizeof(*tx->tx_pages));
1993                 if (tx->tx_frags != NULL)
1994                         LIBCFS_FREE(tx->tx_frags,
1995                                     IBLND_MAX_RDMA_FRAGS *
1996                                             sizeof(*tx->tx_frags));
1997                 if (tx->tx_wrq != NULL)
1998                         LIBCFS_FREE(tx->tx_wrq,
1999                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2000                                     sizeof(*tx->tx_wrq));
2001                 if (tx->tx_sge != NULL)
2002                         LIBCFS_FREE(tx->tx_sge,
2003                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2004                                     sizeof(*tx->tx_sge));
2005                 if (tx->tx_rd != NULL)
2006                         LIBCFS_FREE(tx->tx_rd,
2007                                     offsetof(kib_rdma_desc_t,
2008                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2009         }
2010
2011         LIBCFS_FREE(tpo->tpo_tx_descs,
2012                     pool->po_size * sizeof(kib_tx_t));
2013 out:
2014         kiblnd_fini_pool(pool);
2015         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2016 }
2017
2018 static int
2019 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2020 {
2021         int            i;
2022         int            npg;
2023         kib_pool_t    *pool;
2024         kib_tx_pool_t *tpo;
2025
2026         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
2027         if (tpo == NULL) {
2028                 CERROR("Failed to allocate TX pool\n");
2029                 return -ENOMEM;
2030         }
2031
2032         pool = &tpo->tpo_pool;
2033         kiblnd_init_pool(ps, pool, size);
2034         tpo->tpo_tx_descs = NULL;
2035         tpo->tpo_tx_pages = NULL;
2036
2037         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2038         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
2039                 CERROR("Can't allocate tx pages: %d\n", npg);
2040                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2041                 return -ENOMEM;
2042         }
2043
2044         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
2045         if (tpo->tpo_tx_descs == NULL) {
2046                 CERROR("Can't allocate %d tx descriptors\n", size);
2047                 ps->ps_pool_destroy(pool);
2048                 return -ENOMEM;
2049         }
2050
2051         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2052
2053         for (i = 0; i < size; i++) {
2054                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2055
2056                 tx->tx_pool = tpo;
2057                 if (ps->ps_net->ibn_with_fmr){
2058                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
2059                                      sizeof(*tx->tx_pages));
2060                         if (tx->tx_pages == NULL)
2061                                 break;
2062                 }
2063
2064                 LIBCFS_ALLOC(tx->tx_frags,
2065                              IBLND_MAX_RDMA_FRAGS *
2066                              sizeof(*tx->tx_frags));
2067                 if (tx->tx_frags == NULL)
2068                         break;
2069
2070                 LIBCFS_ALLOC(tx->tx_wrq,
2071                              (1 + IBLND_MAX_RDMA_FRAGS) *
2072                              sizeof(*tx->tx_wrq));
2073                 if (tx->tx_wrq == NULL)
2074                         break;
2075
2076                 LIBCFS_ALLOC(tx->tx_sge,
2077                              (1 + IBLND_MAX_RDMA_FRAGS) *
2078                              sizeof(*tx->tx_sge));
2079                 if (tx->tx_sge == NULL)
2080                         break;
2081
2082                 LIBCFS_ALLOC(tx->tx_rd,
2083                              offsetof(kib_rdma_desc_t,
2084                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
2085                 if (tx->tx_rd == NULL)
2086                         break;
2087         }
2088
2089         if (i == size) {
2090                 kiblnd_map_tx_pool(tpo);
2091                 *pp_po = pool;
2092                 return 0;
2093         }
2094
2095         ps->ps_pool_destroy(pool);
2096         return -ENOMEM;
2097 }
2098
2099 static void
2100 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2101 {
2102         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2103                                              tps_poolset);
2104         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
2105
2106         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2107 }
2108
2109 void
2110 kiblnd_ni_fini_pools(kib_net_t *net)
2111 {
2112         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
2113         if (net->ibn_with_fmr)
2114                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
2115         else if (net->ibn_with_pmr)
2116                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
2117 }
2118
2119 int
2120 kiblnd_net_init_pools(kib_net_t *net)
2121 {
2122         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
2123         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
2124         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
2125         unsigned long      flags;
2126         int                rc;
2127
2128         if (*kiblnd_tunables.kib_fmr_pool_size <
2129             *kiblnd_tunables.kib_ntx / 4) {
2130                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2131                        *kiblnd_tunables.kib_fmr_pool_size,
2132                        *kiblnd_tunables.kib_ntx / 4);
2133                 return -EINVAL;
2134         }
2135
2136         if (*kiblnd_tunables.kib_pmr_pool_size <
2137             *kiblnd_tunables.kib_ntx / 4) {
2138                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2139                        *kiblnd_tunables.kib_pmr_pool_size,
2140                        *kiblnd_tunables.kib_ntx / 4);
2141                 return -EINVAL;
2142         }
2143
2144         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2145         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2146             net->ibn_dev->ibd_hdev->ibh_nmrs > 1) {
2147                 /* premapping can fail if ibd_nmr > 1, so we always create
2148                  * FMR/PMR pool and map-on-demand if premapping failed */
2149                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2150                 rc = kiblnd_init_fmr_pool_set(fps, net);
2151                 if (rc == 0) {
2152                         net->ibn_with_fmr = 1;
2153                 } else if (rc == -ENOSYS) {
2154                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2155                                                   *kiblnd_tunables.kib_pmr_pool_size,
2156                                                   kiblnd_create_pmr_pool,
2157                                                   kiblnd_destroy_pmr_pool,
2158                                                   NULL, NULL);
2159                         if (rc == 0)
2160                                 net->ibn_with_pmr = 1;
2161                 }
2162                 if (rc != 0)
2163                         return rc;
2164         } else {
2165                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2166         }
2167
2168         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2169                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2170                                   kiblnd_tx_init, NULL);
2171         if (rc == 0)
2172                 return 0;
2173
2174         if (net->ibn_with_fmr)
2175                 kiblnd_fini_fmr_pool_set(fps);
2176         else if (net->ibn_with_pmr)
2177                 kiblnd_fini_pool_set(&pps->pps_poolset);
2178
2179         return rc;
2180 }
2181
2182 static int
2183 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2184 {
2185         struct ib_device_attr *attr;
2186         int                    rc;
2187
2188         /* It's safe to assume a HCA can handle a page size
2189          * matching that of the native system */
2190         hdev->ibh_page_shift = PAGE_SHIFT;
2191         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2192         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2193
2194         LIBCFS_ALLOC(attr, sizeof(*attr));
2195         if (attr == NULL) {
2196                 CERROR("Out of memory\n");
2197                 return -ENOMEM;
2198         }
2199
2200         rc = ib_query_device(hdev->ibh_ibdev, attr);
2201         if (rc == 0)
2202                 hdev->ibh_mr_size = attr->max_mr_size;
2203
2204         LIBCFS_FREE(attr, sizeof(*attr));
2205
2206         if (rc != 0) {
2207                 CERROR("Failed to query IB device: %d\n", rc);
2208                 return rc;
2209         }
2210
2211 #ifdef HAVE_OFED_TRANSPORT_IWARP
2212         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2213          * and we have reported the bug, remove these in the future when Chelsio
2214          * bug got fixed. */
2215         if (rdma_node_get_transport(hdev->ibh_ibdev->node_type) ==
2216             RDMA_TRANSPORT_IWARP)
2217                 hdev->ibh_mr_size = (1ULL << 32) - 1;
2218 #endif
2219
2220         if (hdev->ibh_mr_size == ~0ULL) {
2221                 hdev->ibh_mr_shift = 64;
2222                 return 0;
2223         }
2224
2225         for (hdev->ibh_mr_shift = 0;
2226              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2227                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2228                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2229                         return 0;
2230         }
2231
2232         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2233         return -EINVAL;
2234 }
2235
2236 void
2237 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2238 {
2239         int     i;
2240
2241         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2242                 return;
2243
2244         for (i = 0; i < hdev->ibh_nmrs; i++) {
2245                 if (hdev->ibh_mrs[i] == NULL)
2246                         break;
2247
2248                 ib_dereg_mr(hdev->ibh_mrs[i]);
2249         }
2250
2251         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2252         hdev->ibh_mrs  = NULL;
2253         hdev->ibh_nmrs = 0;
2254 }
2255
2256 void
2257 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2258 {
2259         kiblnd_hdev_cleanup_mrs(hdev);
2260
2261         if (hdev->ibh_pd != NULL)
2262                 ib_dealloc_pd(hdev->ibh_pd);
2263
2264         if (hdev->ibh_cmid != NULL)
2265                 rdma_destroy_id(hdev->ibh_cmid);
2266
2267         LIBCFS_FREE(hdev, sizeof(*hdev));
2268 }
2269
2270 int
2271 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2272 {
2273         struct ib_mr *mr;
2274         int           i;
2275         int           rc;
2276         __u64         mm_size;
2277         __u64         mr_size;
2278         int           acflags = IB_ACCESS_LOCAL_WRITE |
2279                                 IB_ACCESS_REMOTE_WRITE;
2280
2281         rc = kiblnd_hdev_get_attr(hdev);
2282         if (rc != 0)
2283                 return rc;
2284
2285         if (hdev->ibh_mr_shift == 64) {
2286                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2287                 if (hdev->ibh_mrs == NULL) {
2288                         CERROR("Failed to allocate MRs table\n");
2289                         return -ENOMEM;
2290                 }
2291
2292                 hdev->ibh_mrs[0] = NULL;
2293                 hdev->ibh_nmrs   = 1;
2294
2295                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2296                 if (IS_ERR(mr)) {
2297                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2298                         kiblnd_hdev_cleanup_mrs(hdev);
2299                         return PTR_ERR(mr);
2300                 }
2301
2302                 hdev->ibh_mrs[0] = mr;
2303
2304                 goto out;
2305         }
2306
2307         mr_size = (1ULL << hdev->ibh_mr_shift);
2308         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2309
2310         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2311
2312         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2313                 /* it's 4T..., assume we will re-code at that time */
2314                 CERROR("Can't support memory size: x"LPX64
2315                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2316                 return -EINVAL;
2317         }
2318
2319         /* create an array of MRs to cover all memory */
2320         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2321         if (hdev->ibh_mrs == NULL) {
2322                 CERROR("Failed to allocate MRs' table\n");
2323                 return -ENOMEM;
2324         }
2325
2326         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2327
2328         for (i = 0; i < hdev->ibh_nmrs; i++) {
2329                 struct ib_phys_buf ipb;
2330                 __u64              iova;
2331
2332                 ipb.size = hdev->ibh_mr_size;
2333                 ipb.addr = i * mr_size;
2334                 iova     = ipb.addr;
2335
2336                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2337                 if (IS_ERR(mr)) {
2338                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2339                                " size "LPX64" : %ld\n",
2340                                ipb.addr, ipb.size, PTR_ERR(mr));
2341                         kiblnd_hdev_cleanup_mrs(hdev);
2342                         return PTR_ERR(mr);
2343                 }
2344
2345                 LASSERT (iova == ipb.addr);
2346
2347                 hdev->ibh_mrs[i] = mr;
2348         }
2349
2350 out:
2351         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2352                 LCONSOLE_INFO("Register global MR array, MR size: "
2353                               LPX64", array size: %d\n",
2354                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2355         return 0;
2356 }
2357
2358 static int
2359 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2360 {       /* DUMMY */
2361         return 0;
2362 }
2363
2364 static int
2365 kiblnd_dev_need_failover(kib_dev_t *dev)
2366 {
2367         struct rdma_cm_id  *cmid;
2368         struct sockaddr_in  srcaddr;
2369         struct sockaddr_in  dstaddr;
2370         int                 rc;
2371
2372         if (dev->ibd_hdev == NULL || /* initializing */
2373             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2374             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2375                 return 1;
2376
2377         /* XXX: it's UGLY, but I don't have better way to find
2378          * ib-bonding HCA failover because:
2379          *
2380          * a. no reliable CM event for HCA failover...
2381          * b. no OFED API to get ib_device for current net_device...
2382          *
2383          * We have only two choices at this point:
2384          *
2385          * a. rdma_bind_addr(), it will conflict with listener cmid
2386          * b. rdma_resolve_addr() to zero addr */
2387         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2388                                      IB_QPT_RC);
2389         if (IS_ERR(cmid)) {
2390                 rc = PTR_ERR(cmid);
2391                 CERROR("Failed to create cmid for failover: %d\n", rc);
2392                 return rc;
2393         }
2394
2395         memset(&srcaddr, 0, sizeof(srcaddr));
2396         srcaddr.sin_family      = AF_INET;
2397         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2398
2399         memset(&dstaddr, 0, sizeof(dstaddr));
2400         dstaddr.sin_family = AF_INET;
2401         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2402                                (struct sockaddr *)&dstaddr, 1);
2403         if (rc != 0) {
2404                 CERROR("Failed to bind %s to device: %d\n",
2405                        dev->ibd_ifname, rc);
2406                 rdma_destroy_id(cmid);
2407                 return rc;
2408         }
2409
2410         LASSERT (cmid->device != NULL);
2411         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2412                 /* don't need device failover */
2413                 rdma_destroy_id(cmid);
2414                 return 0;
2415         }
2416
2417         return 1;
2418 }
2419
2420 int
2421 kiblnd_dev_failover(kib_dev_t *dev)
2422 {
2423         CFS_LIST_HEAD      (zombie_tpo);
2424         CFS_LIST_HEAD      (zombie_ppo);
2425         CFS_LIST_HEAD      (zombie_fpo);
2426         struct rdma_cm_id  *cmid  = NULL;
2427         kib_hca_dev_t      *hdev  = NULL;
2428         kib_hca_dev_t      *old;
2429         struct ib_pd       *pd;
2430         kib_net_t          *net;
2431         struct sockaddr_in  addr;
2432         unsigned long       flags;
2433         int                 rc = 0;
2434
2435         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2436                  dev->ibd_can_failover ||
2437                  dev->ibd_hdev == NULL);
2438
2439         rc = kiblnd_dev_need_failover(dev);
2440         if (rc <= 0)
2441                 goto out;
2442
2443         if (dev->ibd_hdev != NULL &&
2444             dev->ibd_hdev->ibh_cmid != NULL) {
2445                 /* XXX it's not good to close old listener at here,
2446                  * because we can fail to create new listener.
2447                  * But we have to close it now, otherwise rdma_bind_addr
2448                  * will return EADDRINUSE... How crap! */
2449                 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2450
2451                 cmid = dev->ibd_hdev->ibh_cmid;
2452                 /* make next schedule of kiblnd_dev_need_failover
2453                  * will return 1 for me */
2454                 dev->ibd_hdev->ibh_cmid  = NULL;
2455                 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2456
2457                 rdma_destroy_id(cmid);
2458         }
2459
2460         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2461                                      IB_QPT_RC);
2462         if (IS_ERR(cmid)) {
2463                 rc = PTR_ERR(cmid);
2464                 CERROR("Failed to create cmid for failover: %d\n", rc);
2465                 goto out;
2466         }
2467
2468         memset(&addr, 0, sizeof(addr));
2469         addr.sin_family      = AF_INET;
2470         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2471         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2472
2473         /* Bind to failover device or port */
2474         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2475         if (rc != 0) {
2476                 CERROR("Failed to bind %s to device: %d\n",
2477                        dev->ibd_ifname, rc);
2478                 rdma_destroy_id(cmid);
2479                 goto out;
2480         }
2481
2482         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2483         if (hdev == NULL) {
2484                 CERROR("Failed to allocate kib_hca_dev\n");
2485                 rdma_destroy_id(cmid);
2486                 rc = -ENOMEM;
2487                 goto out;
2488         }
2489         memset(hdev, 0, sizeof(*hdev));
2490         atomic_set(&hdev->ibh_ref, 1);
2491         hdev->ibh_dev   = dev;
2492         hdev->ibh_cmid  = cmid;
2493         hdev->ibh_ibdev = cmid->device;
2494
2495         pd = ib_alloc_pd(cmid->device);
2496         if (IS_ERR(pd)) {
2497                 rc = PTR_ERR(pd);
2498                 CERROR("Can't allocate PD: %d\n", rc);
2499                 goto out;
2500         }
2501
2502         hdev->ibh_pd = pd;
2503
2504         rc = rdma_listen(cmid, 0);
2505         if (rc != 0) {
2506                 CERROR("Can't start new listener: %d\n", rc);
2507                 goto out;
2508         }
2509
2510         rc = kiblnd_hdev_setup_mrs(hdev);
2511         if (rc != 0) {
2512                 CERROR("Can't setup device: %d\n", rc);
2513                 goto out;
2514         }
2515
2516         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2517
2518         old = dev->ibd_hdev;
2519         dev->ibd_hdev = hdev; /* take over the refcount */
2520         hdev = old;
2521
2522         cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2523                 kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo);
2524                 if (net->ibn_with_pmr)
2525                         kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo);
2526                 if (net->ibn_with_fmr)
2527                         kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo);
2528         }
2529
2530         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2531  out:
2532         if (!cfs_list_empty(&zombie_tpo))
2533                 kiblnd_destroy_pool_list(&zombie_tpo);
2534         if (!cfs_list_empty(&zombie_ppo))
2535                 kiblnd_destroy_pool_list(&zombie_ppo);
2536         if (!cfs_list_empty(&zombie_fpo))
2537                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2538         if (hdev != NULL)
2539                 kiblnd_hdev_decref(hdev);
2540
2541         if (rc != 0)
2542                 dev->ibd_failed_failover++;
2543         else
2544                 dev->ibd_failed_failover = 0;
2545
2546         return rc;
2547 }
2548
2549 void
2550 kiblnd_destroy_dev (kib_dev_t *dev)
2551 {
2552         LASSERT (dev->ibd_nnets == 0);
2553         LASSERT (cfs_list_empty(&dev->ibd_nets));
2554
2555         cfs_list_del(&dev->ibd_fail_list);
2556         cfs_list_del(&dev->ibd_list);
2557
2558         if (dev->ibd_hdev != NULL)
2559                 kiblnd_hdev_decref(dev->ibd_hdev);
2560
2561         LIBCFS_FREE(dev, sizeof(*dev));
2562 }
2563
2564 kib_dev_t *
2565 kiblnd_create_dev(char *ifname)
2566 {
2567         struct net_device *netdev;
2568         kib_dev_t         *dev;
2569         __u32              netmask;
2570         __u32              ip;
2571         int                up;
2572         int                rc;
2573
2574         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2575         if (rc != 0) {
2576                 CERROR("Can't query IPoIB interface %s: %d\n",
2577                        ifname, rc);
2578                 return NULL;
2579         }
2580
2581         if (!up) {
2582                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2583                 return NULL;
2584         }
2585
2586         LIBCFS_ALLOC(dev, sizeof(*dev));
2587         if (dev == NULL)
2588                 return NULL;
2589
2590         memset(dev, 0, sizeof(*dev));
2591 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2592         if ((netdev = dev_get_by_name(&init_net, ifname)) == NULL) {
2593 #else
2594         if ((netdev = dev_get_by_name(ifname)) == NULL) {
2595 #endif
2596                 CERROR("Can't find netdev of IF: %s\n", ifname);
2597                 LIBCFS_FREE(dev, sizeof(*dev));
2598                 return NULL;
2599         }
2600
2601         dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2602         dev_put(netdev);
2603
2604         CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2605         CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2606         CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2607         dev->ibd_ifip = ip;
2608         strcpy(&dev->ibd_ifname[0], ifname);
2609
2610         /* initialize the device */
2611         rc = kiblnd_dev_failover(dev);
2612         if (rc != 0) {
2613                 CERROR("Can't initialize device: %d\n", rc);
2614                 LIBCFS_FREE(dev, sizeof(*dev));
2615                 return NULL;
2616         }
2617
2618         cfs_list_add_tail(&dev->ibd_list,
2619                           &kiblnd_data.kib_devs);
2620         return dev;
2621 }
2622
2623 void
2624 kiblnd_base_shutdown (void)
2625 {
2626         int i;
2627
2628         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2629
2630         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2631                cfs_atomic_read(&libcfs_kmemory));
2632
2633         switch (kiblnd_data.kib_init) {
2634         default:
2635                 LBUG();
2636
2637         case IBLND_INIT_ALL:
2638         case IBLND_INIT_DATA:
2639                 LASSERT (kiblnd_data.kib_peers != NULL);
2640                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2641                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2642                 }
2643                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2644                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2645
2646                 /* flag threads to terminate; wake and wait for them to die */
2647                 kiblnd_data.kib_shutdown = 1;
2648                 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2649                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2650                 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2651
2652                 i = 2;
2653                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2654                         i++;
2655                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2656                                "Waiting for %d threads to terminate\n",
2657                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2658                         cfs_pause(cfs_time_seconds(1));
2659                 }
2660
2661                 /* fall through */
2662
2663         case IBLND_INIT_NOTHING:
2664                 break;
2665         }
2666
2667         if (kiblnd_data.kib_peers != NULL)
2668                 LIBCFS_FREE(kiblnd_data.kib_peers,
2669                             sizeof(cfs_list_t) *
2670                             kiblnd_data.kib_peer_hash_size);
2671
2672         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2673                cfs_atomic_read(&libcfs_kmemory));
2674
2675         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2676         PORTAL_MODULE_UNUSE;
2677 }
2678
2679 void
2680 kiblnd_shutdown (lnet_ni_t *ni)
2681 {
2682         kib_net_t        *net = ni->ni_data;
2683         cfs_rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2684         int               i;
2685         unsigned long     flags;
2686
2687         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2688
2689         if (net == NULL)
2690                 goto out;
2691
2692         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2693                cfs_atomic_read(&libcfs_kmemory));
2694
2695         cfs_write_lock_irqsave(g_lock, flags);
2696         net->ibn_shutdown = 1;
2697         cfs_write_unlock_irqrestore(g_lock, flags);
2698
2699         switch (net->ibn_init) {
2700         default:
2701                 LBUG();
2702
2703         case IBLND_INIT_ALL:
2704                 /* nuke all existing peers within this net */
2705                 kiblnd_del_peer(ni, LNET_NID_ANY);
2706
2707                 /* Wait for all peer state to clean up */
2708                 i = 2;
2709                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2710                         i++;
2711                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2712                                "%s: waiting for %d peers to disconnect\n",
2713                                libcfs_nid2str(ni->ni_nid),
2714                                cfs_atomic_read(&net->ibn_npeers));
2715                         cfs_pause(cfs_time_seconds(1));
2716                 }
2717
2718                 kiblnd_ni_fini_pools(net);
2719
2720                 cfs_write_lock_irqsave(g_lock, flags);
2721                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2722                 net->ibn_dev->ibd_nnets--;
2723                 cfs_list_del(&net->ibn_list);
2724                 cfs_write_unlock_irqrestore(g_lock, flags);
2725
2726                 /* fall through */
2727
2728         case IBLND_INIT_NOTHING:
2729                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2730
2731                 if (net->ibn_dev != NULL &&
2732                     net->ibn_dev->ibd_nnets == 0)
2733                         kiblnd_destroy_dev(net->ibn_dev);
2734
2735                 break;
2736         }
2737
2738         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2739                cfs_atomic_read(&libcfs_kmemory));
2740
2741         net->ibn_init = IBLND_INIT_NOTHING;
2742         ni->ni_data = NULL;
2743
2744         LIBCFS_FREE(net, sizeof(*net));
2745
2746 out:
2747         if (cfs_list_empty(&kiblnd_data.kib_devs))
2748                 kiblnd_base_shutdown();
2749         return;
2750 }
2751
2752 int
2753 kiblnd_base_startup (void)
2754 {
2755         int i;
2756         int rc;
2757
2758         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2759
2760         PORTAL_MODULE_USE;
2761         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2762
2763         cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2764
2765         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2766         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2767
2768         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2769         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2770                      sizeof(cfs_list_t) *
2771                             kiblnd_data.kib_peer_hash_size);
2772         if (kiblnd_data.kib_peers == NULL) {
2773                 goto failed;
2774         }
2775         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2776                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2777
2778         cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2779         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2780         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2781         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2782
2783         cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2784         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2785         cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
2786         cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2787
2788         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2789
2790         /* lists/ptrs/locks initialised */
2791         kiblnd_data.kib_init = IBLND_INIT_DATA;
2792         /*****************************************************/
2793
2794         for (i = 0; i < IBLND_N_SCHED; i++) {
2795                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2796                 if (rc != 0) {
2797                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2798                                i, rc);
2799                         goto failed;
2800                 }
2801         }
2802
2803         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2804         if (rc != 0) {
2805                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2806                 goto failed;
2807         }
2808
2809         if (*kiblnd_tunables.kib_dev_failover != 0)
2810                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
2811
2812         if (rc != 0) {
2813                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2814                 goto failed;
2815         }
2816
2817         /* flag everything initialised */
2818         kiblnd_data.kib_init = IBLND_INIT_ALL;
2819         /*****************************************************/
2820
2821         return 0;
2822
2823  failed:
2824         kiblnd_base_shutdown();
2825         return -ENETDOWN;
2826 }
2827
2828 int
2829 kiblnd_startup (lnet_ni_t *ni)
2830 {
2831         char                     *ifname;
2832         kib_dev_t                *ibdev = NULL;
2833         kib_net_t                *net;
2834         cfs_list_t               *tmp;
2835         struct timeval            tv;
2836         unsigned long             flags;
2837         int                       rc;
2838
2839         LASSERT (ni->ni_lnd == &the_o2iblnd);
2840
2841         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2842                 rc = kiblnd_base_startup();
2843                 if (rc != 0)
2844                         return rc;
2845         }
2846
2847         LIBCFS_ALLOC(net, sizeof(*net));
2848         ni->ni_data = net;
2849         if (net == NULL)
2850                 goto failed;
2851
2852         memset(net, 0, sizeof(*net));
2853
2854         cfs_gettimeofday(&tv);
2855         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2856
2857         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2858         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2859         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2860         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2861
2862         if (ni->ni_interfaces[0] != NULL) {
2863                 /* Use the IPoIB interface specified in 'networks=' */
2864
2865                 CLASSERT (LNET_MAX_INTERFACES > 1);
2866                 if (ni->ni_interfaces[1] != NULL) {
2867                         CERROR("Multiple interfaces not supported\n");
2868                         goto failed;
2869                 }
2870
2871                 ifname = ni->ni_interfaces[0];
2872         } else {
2873                 ifname = *kiblnd_tunables.kib_default_ipif;
2874         }
2875
2876         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2877                 CERROR("IPoIB interface name too long: %s\n", ifname);
2878                 goto failed;
2879         }
2880
2881         cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2882                 ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
2883
2884                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2885                         break;
2886
2887                 ibdev = NULL;
2888         }
2889
2890         if (ibdev == NULL)
2891                 ibdev = kiblnd_create_dev(ifname);
2892
2893         if (ibdev == NULL)
2894                 goto failed;
2895
2896         net->ibn_dev = ibdev;
2897         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2898
2899         rc = kiblnd_net_init_pools(net);
2900         if (rc != 0) {
2901                 CERROR("Failed to initialize NI pools: %d\n", rc);
2902                 goto failed;
2903         }
2904
2905         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2906         ibdev->ibd_nnets++;
2907         cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2908         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2909
2910         net->ibn_init = IBLND_INIT_ALL;
2911
2912         return 0;
2913
2914 failed:
2915         if (net->ibn_dev == NULL && ibdev != NULL)
2916                 kiblnd_destroy_dev(ibdev);
2917
2918         kiblnd_shutdown(ni);
2919
2920         CDEBUG(D_NET, "kiblnd_startup failed\n");
2921         return -ENETDOWN;
2922 }
2923
2924 void __exit
2925 kiblnd_module_fini (void)
2926 {
2927         lnet_unregister_lnd(&the_o2iblnd);
2928         kiblnd_tunables_fini();
2929 }
2930
2931 int __init
2932 kiblnd_module_init (void)
2933 {
2934         int    rc;
2935
2936         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2937         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2938                   <= IBLND_MSG_SIZE);
2939         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2940                   <= IBLND_MSG_SIZE);
2941
2942         rc = kiblnd_tunables_init();
2943         if (rc != 0)
2944                 return rc;
2945
2946         lnet_register_lnd(&the_o2iblnd);
2947
2948         return 0;
2949 }
2950
2951 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2952 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2953 MODULE_LICENSE("GPL");
2954
2955 module_init(kiblnd_module_init);
2956 module_exit(kiblnd_module_fini);