lnet/klnds/o2iblnd/o2iblnd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 lnd_t the_o2iblnd = {
  44         .lnd_type       = O2IBLND,
  45         .lnd_startup    = kiblnd_startup,
  46         .lnd_shutdown   = kiblnd_shutdown,
  47         .lnd_ctl        = kiblnd_ctl,
  48         .lnd_query      = kiblnd_query,
  49         .lnd_send       = kiblnd_send,
  50         .lnd_recv       = kiblnd_recv,
  51 };
  52
  53 kib_data_t              kiblnd_data;
  54
  55 __u32
  56 kiblnd_cksum (void *ptr, int nob)
  57 {
  58         char  *c  = ptr;
  59         __u32  sum = 0;
  60
  61         while (nob-- > 0)
  62                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  63
  64         /* ensure I don't return 0 (== no checksum) */
  65         return (sum == 0) ? 1 : sum;
  66 }
  67
  68 static char *
  69 kiblnd_msgtype2str(int type)
  70 {
  71         switch (type) {
  72         case IBLND_MSG_CONNREQ:
  73                 return "CONNREQ";
  74
  75         case IBLND_MSG_CONNACK:
  76                 return "CONNACK";
  77
  78         case IBLND_MSG_NOOP:
  79                 return "NOOP";
  80
  81         case IBLND_MSG_IMMEDIATE:
  82                 return "IMMEDIATE";
  83
  84         case IBLND_MSG_PUT_REQ:
  85                 return "PUT_REQ";
  86
  87         case IBLND_MSG_PUT_NAK:
  88                 return "PUT_NAK";
  89
  90         case IBLND_MSG_PUT_ACK:
  91                 return "PUT_ACK";
  92
  93         case IBLND_MSG_PUT_DONE:
  94                 return "PUT_DONE";
  95
  96         case IBLND_MSG_GET_REQ:
  97                 return "GET_REQ";
  98
  99         case IBLND_MSG_GET_DONE:
 100                 return "GET_DONE";
 101
 102         default:
 103                 return "???";
 104         }
 105 }
 106
 107 static int
 108 kiblnd_msgtype2size(int type)
 109 {
 110         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 111
 112         switch (type) {
 113         case IBLND_MSG_CONNREQ:
 114         case IBLND_MSG_CONNACK:
 115                 return hdr_size + sizeof(kib_connparams_t);
 116
 117         case IBLND_MSG_NOOP:
 118                 return hdr_size;
 119
 120         case IBLND_MSG_IMMEDIATE:
 121                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 122
 123         case IBLND_MSG_PUT_REQ:
 124                 return hdr_size + sizeof(kib_putreq_msg_t);
 125
 126         case IBLND_MSG_PUT_ACK:
 127                 return hdr_size + sizeof(kib_putack_msg_t);
 128
 129         case IBLND_MSG_GET_REQ:
 130                 return hdr_size + sizeof(kib_get_msg_t);
 131
 132         case IBLND_MSG_PUT_NAK:
 133         case IBLND_MSG_PUT_DONE:
 134         case IBLND_MSG_GET_DONE:
 135                 return hdr_size + sizeof(kib_completion_msg_t);
 136         default:
 137                 return -1;
 138         }
 139 }
 140
 141 static int
 142 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 143 {
 144         kib_rdma_desc_t   *rd;
 145         int                nob;
 146         int                n;
 147         int                i;
 148
 149         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 150                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 151
 152         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 153                               &msg->ibm_u.get.ibgm_rd :
 154                               &msg->ibm_u.putack.ibpam_rd;
 155
 156         if (flip) {
 157                 __swab32s(&rd->rd_key);
 158                 __swab32s(&rd->rd_nfrags);
 159         }
 160
 161         n = rd->rd_nfrags;
 162
 163         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 164                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 165                        n, IBLND_MAX_RDMA_FRAGS);
 166                 return 1;
 167         }
 168
 169         nob = offsetof (kib_msg_t, ibm_u) +
 170               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 171
 172         if (msg->ibm_nob < nob) {
 173                 CERROR("Short %s: %d(%d)\n",
 174                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 175                 return 1;
 176         }
 177
 178         if (!flip)
 179                 return 0;
 180
 181         for (i = 0; i < n; i++) {
 182                 __swab32s(&rd->rd_frags[i].rf_nob);
 183                 __swab64s(&rd->rd_frags[i].rf_addr);
 184         }
 185
 186         return 0;
 187 }
 188
 189 void
 190 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 191                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 192 {
 193         kib_net_t *net = ni->ni_data;
 194
 195         /* CAVEAT EMPTOR! all message fields not set here should have been
 196          * initialised previously. */
 197         msg->ibm_magic    = IBLND_MSG_MAGIC;
 198         msg->ibm_version  = version;
 199         /*   ibm_type */
 200         msg->ibm_credits  = credits;
 201         /*   ibm_nob */
 202         msg->ibm_cksum    = 0;
 203         msg->ibm_srcnid   = ni->ni_nid;
 204         msg->ibm_srcstamp = net->ibn_incarnation;
 205         msg->ibm_dstnid   = dstnid;
 206         msg->ibm_dststamp = dststamp;
 207
 208         if (*kiblnd_tunables.kib_cksum) {
 209                 /* NB ibm_cksum zero while computing cksum */
 210                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 211         }
 212 }
 213
 214 int
 215 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 216 {
 217         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 218         __u32     msg_cksum;
 219         __u16     version;
 220         int       msg_nob;
 221         int       flip;
 222
 223         /* 6 bytes are enough to have received magic + version */
 224         if (nob < 6) {
 225                 CERROR("Short message: %d\n", nob);
 226                 return -EPROTO;
 227         }
 228
 229         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 230                 flip = 0;
 231         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 232                 flip = 1;
 233         } else {
 234                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 235                 return -EPROTO;
 236         }
 237
 238         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 239         if (version != IBLND_MSG_VERSION &&
 240             version != IBLND_MSG_VERSION_1) {
 241                 CERROR("Bad version: %x\n", version);
 242                 return -EPROTO;
 243         }
 244
 245         if (nob < hdr_size) {
 246                 CERROR("Short message: %d\n", nob);
 247                 return -EPROTO;
 248         }
 249
 250         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 251         if (msg_nob > nob) {
 252                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 253                 return -EPROTO;
 254         }
 255
 256         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 257          * gets flipped */
 258         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 259         msg->ibm_cksum = 0;
 260         if (msg_cksum != 0 &&
 261             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 262                 CERROR("Bad checksum\n");
 263                 return -EPROTO;
 264         }
 265
 266         msg->ibm_cksum = msg_cksum;
 267
 268         if (flip) {
 269                 /* leave magic unflipped as a clue to peer endianness */
 270                 msg->ibm_version = version;
 271                 CLASSERT (sizeof(msg->ibm_type) == 1);
 272                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 273                 msg->ibm_nob     = msg_nob;
 274                 __swab64s(&msg->ibm_srcnid);
 275                 __swab64s(&msg->ibm_srcstamp);
 276                 __swab64s(&msg->ibm_dstnid);
 277                 __swab64s(&msg->ibm_dststamp);
 278         }
 279
 280         if (msg->ibm_srcnid == LNET_NID_ANY) {
 281                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 282                 return -EPROTO;
 283         }
 284
 285         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 286                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 287                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 288                 return -EPROTO;
 289         }
 290
 291         switch (msg->ibm_type) {
 292         default:
 293                 CERROR("Unknown message type %x\n", msg->ibm_type);
 294                 return -EPROTO;
 295
 296         case IBLND_MSG_NOOP:
 297         case IBLND_MSG_IMMEDIATE:
 298         case IBLND_MSG_PUT_REQ:
 299                 break;
 300
 301         case IBLND_MSG_PUT_ACK:
 302         case IBLND_MSG_GET_REQ:
 303                 if (kiblnd_unpack_rd(msg, flip))
 304                         return -EPROTO;
 305                 break;
 306
 307         case IBLND_MSG_PUT_NAK:
 308         case IBLND_MSG_PUT_DONE:
 309         case IBLND_MSG_GET_DONE:
 310                 if (flip)
 311                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 312                 break;
 313
 314         case IBLND_MSG_CONNREQ:
 315         case IBLND_MSG_CONNACK:
 316                 if (flip) {
 317                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 318                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 319                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 320                 }
 321                 break;
 322         }
 323         return 0;
 324 }
 325
 326 int
 327 kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 328 {
 329         kib_peer_t     *peer;
 330         kib_net_t      *net = ni->ni_data;
 331         unsigned long   flags;
 332
 333         LASSERT (net != NULL);
 334         LASSERT (nid != LNET_NID_ANY);
 335
 336         LIBCFS_ALLOC(peer, sizeof(*peer));
 337         if (peer == NULL) {
 338                 CERROR("Cannot allocate peer\n");
 339                 return -ENOMEM;
 340         }
 341
 342         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 343
 344         peer->ibp_ni = ni;
 345         peer->ibp_nid = nid;
 346         peer->ibp_error = 0;
 347         peer->ibp_last_alive = 0;
 348         cfs_atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 349
 350         CFS_INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 351         CFS_INIT_LIST_HEAD(&peer->ibp_conns);
 352         CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue);
 353
 354         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 355
 356         /* always called with a ref on ni, which prevents ni being shutdown */
 357         LASSERT (net->ibn_shutdown == 0);
 358
 359         /* npeers only grows with the global lock held */
 360         cfs_atomic_inc(&net->ibn_npeers);
 361
 362         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 363
 364         *peerp = peer;
 365         return 0;
 366 }
 367
 368 void
 369 kiblnd_destroy_peer (kib_peer_t *peer)
 370 {
 371         kib_net_t *net = peer->ibp_ni->ni_data;
 372
 373         LASSERT (net != NULL);
 374         LASSERT (cfs_atomic_read(&peer->ibp_refcount) == 0);
 375         LASSERT (!kiblnd_peer_active(peer));
 376         LASSERT (peer->ibp_connecting == 0);
 377         LASSERT (peer->ibp_accepting == 0);
 378         LASSERT (cfs_list_empty(&peer->ibp_conns));
 379         LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
 380
 381         LIBCFS_FREE(peer, sizeof(*peer));
 382
 383         /* NB a peer's connections keep a reference on their peer until
 384          * they are destroyed, so we can be assured that _all_ state to do
 385          * with this peer has been cleaned up when its refcount drops to
 386          * zero. */
 387         cfs_atomic_dec(&net->ibn_npeers);
 388 }
 389
 390 kib_peer_t *
 391 kiblnd_find_peer_locked (lnet_nid_t nid)
 392 {
 393         /* the caller is responsible for accounting the additional reference
 394          * that this creates */
 395         cfs_list_t       *peer_list = kiblnd_nid2peerlist(nid);
 396         cfs_list_t       *tmp;
 397         kib_peer_t       *peer;
 398
 399         cfs_list_for_each (tmp, peer_list) {
 400
 401                 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
 402
 403                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 404                          peer->ibp_accepting > 0 ||
 405                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
 406
 407                 if (peer->ibp_nid != nid)
 408                         continue;
 409
 410                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 411                        peer, libcfs_nid2str(nid),
 412                        cfs_atomic_read(&peer->ibp_refcount),
 413                        peer->ibp_version);
 414                 return peer;
 415         }
 416         return NULL;
 417 }
 418
 419 void
 420 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 421 {
 422         LASSERT (cfs_list_empty(&peer->ibp_conns));
 423
 424         LASSERT (kiblnd_peer_active(peer));
 425         cfs_list_del_init(&peer->ibp_list);
 426         /* lose peerlist's ref */
 427         kiblnd_peer_decref(peer);
 428 }
 429
 430 int
 431 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 432                       lnet_nid_t *nidp, int *count)
 433 {
 434         kib_peer_t            *peer;
 435         cfs_list_t            *ptmp;
 436         int                    i;
 437         unsigned long          flags;
 438
 439         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 440
 441         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 442
 443                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 444
 445                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 446                         LASSERT (peer->ibp_connecting > 0 ||
 447                                  peer->ibp_accepting > 0 ||
 448                                  !cfs_list_empty(&peer->ibp_conns));
 449
 450                         if (peer->ibp_ni != ni)
 451                                 continue;
 452
 453                         if (index-- > 0)
 454                                 continue;
 455
 456                         *nidp = peer->ibp_nid;
 457                         *count = cfs_atomic_read(&peer->ibp_refcount);
 458
 459                         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 460                                                    flags);
 461                         return 0;
 462                 }
 463         }
 464
 465         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 466         return -ENOENT;
 467 }
 468
 469 void
 470 kiblnd_del_peer_locked (kib_peer_t *peer)
 471 {
 472         cfs_list_t           *ctmp;
 473         cfs_list_t           *cnxt;
 474         kib_conn_t           *conn;
 475
 476         if (cfs_list_empty(&peer->ibp_conns)) {
 477                 kiblnd_unlink_peer_locked(peer);
 478         } else {
 479                 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 480                         conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 481
 482                         kiblnd_close_conn_locked(conn, 0);
 483                 }
 484                 /* NB closing peer's last conn unlinked it. */
 485         }
 486         /* NB peer now unlinked; might even be freed if the peer table had the
 487          * last ref on it. */
 488 }
 489
 490 int
 491 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 492 {
 493         CFS_LIST_HEAD         (zombies);
 494         cfs_list_t            *ptmp;
 495         cfs_list_t            *pnxt;
 496         kib_peer_t            *peer;
 497         int                    lo;
 498         int                    hi;
 499         int                    i;
 500         unsigned long          flags;
 501         int                    rc = -ENOENT;
 502
 503         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 504
 505         if (nid != LNET_NID_ANY) {
 506                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 507         } else {
 508                 lo = 0;
 509                 hi = kiblnd_data.kib_peer_hash_size - 1;
 510         }
 511
 512         for (i = lo; i <= hi; i++) {
 513                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 514                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 515                         LASSERT (peer->ibp_connecting > 0 ||
 516                                  peer->ibp_accepting > 0 ||
 517                                  !cfs_list_empty(&peer->ibp_conns));
 518
 519                         if (peer->ibp_ni != ni)
 520                                 continue;
 521
 522                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 523                                 continue;
 524
 525                         if (!cfs_list_empty(&peer->ibp_tx_queue)) {
 526                                 LASSERT (cfs_list_empty(&peer->ibp_conns));
 527
 528                                 cfs_list_splice_init(&peer->ibp_tx_queue,
 529                                                      &zombies);
 530                         }
 531
 532                         kiblnd_del_peer_locked(peer);
 533                         rc = 0;         /* matched something */
 534                 }
 535         }
 536
 537         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 538
 539         kiblnd_txlist_done(ni, &zombies, -EIO);
 540
 541         return rc;
 542 }
 543
 544 kib_conn_t *
 545 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 546 {
 547         kib_peer_t            *peer;
 548         cfs_list_t            *ptmp;
 549         kib_conn_t            *conn;
 550         cfs_list_t            *ctmp;
 551         int                    i;
 552         unsigned long          flags;
 553
 554         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 555
 556         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 557                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 558
 559                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 560                         LASSERT (peer->ibp_connecting > 0 ||
 561                                  peer->ibp_accepting > 0 ||
 562                                  !cfs_list_empty(&peer->ibp_conns));
 563
 564                         if (peer->ibp_ni != ni)
 565                                 continue;
 566
 567                         cfs_list_for_each (ctmp, &peer->ibp_conns) {
 568                                 if (index-- > 0)
 569                                         continue;
 570
 571                                 conn = cfs_list_entry(ctmp, kib_conn_t,
 572                                                       ibc_list);
 573                                 kiblnd_conn_addref(conn);
 574                                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 575                                                            flags);
 576                                 return conn;
 577                         }
 578                 }
 579         }
 580
 581         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 582         return NULL;
 583 }
 584
 585 void
 586 kiblnd_debug_rx (kib_rx_t *rx)
 587 {
 588         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 589                rx, rx->rx_status, rx->rx_msg->ibm_type,
 590                rx->rx_msg->ibm_credits);
 591 }
 592
 593 void
 594 kiblnd_debug_tx (kib_tx_t *tx)
 595 {
 596         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 597                "cookie "LPX64" msg %s%s type %x cred %d\n",
 598                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 599                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 600                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 601                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 602                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 603 }
 604
 605 void
 606 kiblnd_debug_conn (kib_conn_t *conn)
 607 {
 608         cfs_list_t           *tmp;
 609         int                   i;
 610
 611         cfs_spin_lock(&conn->ibc_lock);
 612
 613         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 614                cfs_atomic_read(&conn->ibc_refcount), conn,
 615                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 616         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 617                conn->ibc_state, conn->ibc_noops_posted,
 618                conn->ibc_nsends_posted, conn->ibc_credits,
 619                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 620         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 621
 622         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 623         cfs_list_for_each(tmp, &conn->ibc_early_rxs)
 624                 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
 625
 626         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 627         cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 628                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 629
 630         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 631         cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 632                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 633
 634         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 635         cfs_list_for_each(tmp, &conn->ibc_tx_queue)
 636                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 637
 638         CDEBUG(D_CONSOLE, "   active_txs:\n");
 639         cfs_list_for_each(tmp, &conn->ibc_active_txs)
 640                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 641
 642         CDEBUG(D_CONSOLE, "   rxs:\n");
 643         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 644                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 645
 646         cfs_spin_unlock(&conn->ibc_lock);
 647 }
 648
 649 int
 650 kiblnd_translate_mtu(int value)
 651 {
 652         switch (value) {
 653         default:
 654                 return -1;
 655         case 0:
 656                 return 0;
 657         case 256:
 658                 return IB_MTU_256;
 659         case 512:
 660                 return IB_MTU_512;
 661         case 1024:
 662                 return IB_MTU_1024;
 663         case 2048:
 664                 return IB_MTU_2048;
 665         case 4096:
 666                 return IB_MTU_4096;
 667         }
 668 }
 669
 670 static void
 671 kiblnd_setup_mtu(struct rdma_cm_id *cmid)
 672 {
 673         unsigned long flags;
 674         int           mtu;
 675
 676         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 677         if (cmid->route.path_rec == NULL)
 678                 return;
 679
 680         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 681
 682         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 683         LASSERT (mtu >= 0);
 684         if (mtu != 0)
 685                 cmid->route.path_rec->mtu = mtu;
 686
 687         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 688 }
 689
 690 kib_conn_t *
 691 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 692                    int state, int version)
 693 {
 694         /* CAVEAT EMPTOR:
 695          * If the new conn is created successfully it takes over the caller's
 696          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 697          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 698          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 699          * to destroy 'cmid' here since I'm called from the CM which still has
 700          * its ref on 'cmid'). */
 701         kib_net_t              *net = peer->ibp_ni->ni_data;
 702         struct ib_qp_init_attr *init_qp_attr;
 703         kib_conn_t             *conn;
 704         struct ib_cq           *cq;
 705         unsigned long           flags;
 706         int                     rc;
 707         int                     i;
 708
 709         LASSERT (net != NULL);
 710         LASSERT (!cfs_in_interrupt());
 711
 712         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 713         if (init_qp_attr == NULL) {
 714                 CERROR("Can't allocate qp_attr for %s\n",
 715                        libcfs_nid2str(peer->ibp_nid));
 716                 goto failed_0;
 717         }
 718
 719         LIBCFS_ALLOC(conn, sizeof(*conn));
 720         if (conn == NULL) {
 721                 CERROR("Can't allocate connection for %s\n",
 722                        libcfs_nid2str(peer->ibp_nid));
 723                 goto failed_1;
 724         }
 725
 726         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 727
 728         conn->ibc_state = IBLND_CONN_INIT;
 729         conn->ibc_version = version;
 730         conn->ibc_peer = peer;                  /* I take the caller's ref */
 731         cmid->context = conn;                   /* for future CM callbacks */
 732         conn->ibc_cmid = cmid;
 733
 734         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 735         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 736         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 737         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 738         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 739         cfs_spin_lock_init(&conn->ibc_lock);
 740
 741         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 742         if (conn->ibc_connvars == NULL) {
 743                 CERROR("Can't allocate in-progress connection state\n");
 744                 goto failed_2;
 745         }
 746         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 747
 748         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 749         if (conn->ibc_rxs == NULL) {
 750                 CERROR("Cannot allocate RX buffers\n");
 751                 goto failed_2;
 752         }
 753         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 754
 755         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
 756                                 IBLND_RX_MSG_PAGES(version));
 757         if (rc != 0)
 758                 goto failed_2;
 759
 760         kiblnd_map_rx_descs(conn);
 761
 762 #ifdef HAVE_OFED_IB_COMP_VECTOR
 763         cq = ib_create_cq(cmid->device,
 764                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 765                           IBLND_CQ_ENTRIES(version), 0);
 766 #else
 767         cq = ib_create_cq(cmid->device,
 768                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 769                           IBLND_CQ_ENTRIES(version));
 770 #endif
 771         if (IS_ERR(cq)) {
 772                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 773                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 774                 goto failed_2;
 775         }
 776
 777         conn->ibc_cq = cq;
 778
 779         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 780         if (rc != 0) {
 781                 CERROR("Can't request completion notificiation: %d\n", rc);
 782                 goto failed_2;
 783         }
 784
 785         kiblnd_setup_mtu(cmid);
 786
 787         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 788         init_qp_attr->event_handler = kiblnd_qp_event;
 789         init_qp_attr->qp_context = conn;
 790         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 791         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 792         init_qp_attr->cap.max_send_sge = 1;
 793         init_qp_attr->cap.max_recv_sge = 1;
 794         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 795         init_qp_attr->qp_type = IB_QPT_RC;
 796         init_qp_attr->send_cq = cq;
 797         init_qp_attr->recv_cq = cq;
 798
 799         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 800         if (rc != 0) {
 801                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 802                        rc, init_qp_attr->cap.max_send_wr,
 803                        init_qp_attr->cap.max_recv_wr);
 804                 goto failed_2;
 805         }
 806
 807         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 808
 809         /* 1 ref for caller and each rxmsg */
 810         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 811         conn->ibc_nrx = IBLND_RX_MSGS(version);
 812
 813         /* post receives */
 814         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 815                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 816                                     IBLND_POSTRX_NO_CREDIT);
 817                 if (rc != 0) {
 818                         CERROR("Can't post rxmsg: %d\n", rc);
 819
 820                         /* Make posted receives complete */
 821                         kiblnd_abort_receives(conn);
 822
 823                         /* correct # of posted buffers
 824                          * NB locking needed now I'm racing with completion */
 825                         cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
 826                                               flags);
 827                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 828                         cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 829                                                    flags);
 830
 831                         /* cmid will be destroyed by CM(ofed) after cm_callback
 832                          * returned, so we can't refer it anymore
 833                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 834                         rdma_destroy_qp(conn->ibc_cmid);
 835                         conn->ibc_cmid = NULL;
 836
 837                         /* Drop my own and unused rxbuffer refcounts */
 838                         while (i++ <= IBLND_RX_MSGS(version))
 839                                 kiblnd_conn_decref(conn);
 840
 841                         return NULL;
 842                 }
 843         }
 844
 845         /* Init successful! */
 846         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 847                  state == IBLND_CONN_PASSIVE_WAIT);
 848         conn->ibc_state = state;
 849
 850         /* 1 more conn */
 851         cfs_atomic_inc(&net->ibn_nconns);
 852         return conn;
 853
 854  failed_2:
 855         kiblnd_destroy_conn(conn);
 856  failed_1:
 857         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 858  failed_0:
 859         return NULL;
 860 }
 861
 862 void
 863 kiblnd_destroy_conn (kib_conn_t *conn)
 864 {
 865         struct rdma_cm_id *cmid = conn->ibc_cmid;
 866         kib_peer_t        *peer = conn->ibc_peer;
 867         int                rc;
 868
 869         LASSERT (!cfs_in_interrupt());
 870         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 871         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 872         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 873         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 874         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 875         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 876         LASSERT (conn->ibc_noops_posted == 0);
 877         LASSERT (conn->ibc_nsends_posted == 0);
 878
 879         switch (conn->ibc_state) {
 880         default:
 881                 /* conn must be completely disengaged from the network */
 882                 LBUG();
 883
 884         case IBLND_CONN_DISCONNECTED:
 885                 /* connvars should have been freed already */
 886                 LASSERT (conn->ibc_connvars == NULL);
 887                 break;
 888
 889         case IBLND_CONN_INIT:
 890                 break;
 891         }
 892
 893         /* conn->ibc_cmid might be destroyed by CM already */
 894         if (cmid != NULL && cmid->qp != NULL)
 895                 rdma_destroy_qp(cmid);
 896
 897         if (conn->ibc_cq != NULL) {
 898                 rc = ib_destroy_cq(conn->ibc_cq);
 899                 if (rc != 0)
 900                         CWARN("Error destroying CQ: %d\n", rc);
 901         }
 902
 903         if (conn->ibc_rx_pages != NULL)
 904                 kiblnd_unmap_rx_descs(conn);
 905
 906         if (conn->ibc_rxs != NULL) {
 907                 LIBCFS_FREE(conn->ibc_rxs,
 908                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 909         }
 910
 911         if (conn->ibc_connvars != NULL)
 912                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 913
 914         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 915         if (conn->ibc_state != IBLND_CONN_INIT) {
 916                 kib_net_t *net = peer->ibp_ni->ni_data;
 917
 918                 kiblnd_peer_decref(peer);
 919                 rdma_destroy_id(cmid);
 920                 cfs_atomic_dec(&net->ibn_nconns);
 921         }
 922
 923         LIBCFS_FREE(conn, sizeof(*conn));
 924 }
 925
 926 int
 927 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 928 {
 929         kib_conn_t             *conn;
 930         cfs_list_t             *ctmp;
 931         cfs_list_t             *cnxt;
 932         int                     count = 0;
 933
 934         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 935                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 936
 937                 CDEBUG(D_NET, "Closing conn -> %s, "
 938                               "version: %x, reason: %d\n",
 939                        libcfs_nid2str(peer->ibp_nid),
 940                        conn->ibc_version, why);
 941
 942                 kiblnd_close_conn_locked(conn, why);
 943                 count++;
 944         }
 945
 946         return count;
 947 }
 948
 949 int
 950 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 951                                  int version, __u64 incarnation)
 952 {
 953         kib_conn_t             *conn;
 954         cfs_list_t             *ctmp;
 955         cfs_list_t             *cnxt;
 956         int                     count = 0;
 957
 958         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 959                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 960
 961                 if (conn->ibc_version     == version &&
 962                     conn->ibc_incarnation == incarnation)
 963                         continue;
 964
 965                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
 966                               "incarnation:"LPX64"(%x, "LPX64")\n",
 967                        libcfs_nid2str(peer->ibp_nid),
 968                        conn->ibc_version, conn->ibc_incarnation,
 969                        version, incarnation);
 970
 971                 kiblnd_close_conn_locked(conn, -ESTALE);
 972                 count++;
 973         }
 974
 975         return count;
 976 }
 977
 978 int
 979 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 980 {
 981         kib_peer_t             *peer;
 982         cfs_list_t             *ptmp;
 983         cfs_list_t             *pnxt;
 984         int                     lo;
 985         int                     hi;
 986         int                     i;
 987         unsigned long           flags;
 988         int                     count = 0;
 989
 990         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 991
 992         if (nid != LNET_NID_ANY)
 993                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 994         else {
 995                 lo = 0;
 996                 hi = kiblnd_data.kib_peer_hash_size - 1;
 997         }
 998
 999         for (i = lo; i <= hi; i++) {
1000                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1001
1002                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1003                         LASSERT (peer->ibp_connecting > 0 ||
1004                                  peer->ibp_accepting > 0 ||
1005                                  !cfs_list_empty(&peer->ibp_conns));
1006
1007                         if (peer->ibp_ni != ni)
1008                                 continue;
1009
1010                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1011                                 continue;
1012
1013                         count += kiblnd_close_peer_conns_locked(peer, 0);
1014                 }
1015         }
1016
1017         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1018
1019         /* wildcards always succeed */
1020         if (nid == LNET_NID_ANY)
1021                 return 0;
1022
1023         return (count == 0) ? -ENOENT : 0;
1024 }
1025
1026 int
1027 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1028 {
1029         struct libcfs_ioctl_data *data = arg;
1030         int                       rc = -EINVAL;
1031
1032         switch(cmd) {
1033         case IOC_LIBCFS_GET_PEER: {
1034                 lnet_nid_t   nid = 0;
1035                 int          count = 0;
1036
1037                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1038                                           &nid, &count);
1039                 data->ioc_nid    = nid;
1040                 data->ioc_count  = count;
1041                 break;
1042         }
1043
1044         case IOC_LIBCFS_DEL_PEER: {
1045                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1046                 break;
1047         }
1048         case IOC_LIBCFS_GET_CONN: {
1049                 kib_conn_t *conn;
1050
1051                 rc = 0;
1052                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1053                 if (conn == NULL) {
1054                         rc = -ENOENT;
1055                         break;
1056                 }
1057
1058                 LASSERT (conn->ibc_cmid != NULL);
1059                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1060                 if (conn->ibc_cmid->route.path_rec == NULL)
1061                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1062                 else
1063                         data->ioc_u32[0] =
1064                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1065                 kiblnd_conn_decref(conn);
1066                 break;
1067         }
1068         case IOC_LIBCFS_CLOSE_CONNECTION: {
1069                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1070                 break;
1071         }
1072
1073         default:
1074                 break;
1075         }
1076
1077         return rc;
1078 }
1079
1080 void
1081 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1082 {
1083         cfs_time_t         last_alive = 0;
1084         cfs_time_t         now = cfs_time_current();
1085         cfs_rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1086         kib_peer_t        *peer;
1087         unsigned long      flags;
1088
1089         cfs_read_lock_irqsave(glock, flags);
1090
1091         peer = kiblnd_find_peer_locked(nid);
1092         if (peer != NULL) {
1093                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1094                          peer->ibp_accepting > 0 ||
1095                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1096                 last_alive = peer->ibp_last_alive;
1097         }
1098
1099         cfs_read_unlock_irqrestore(glock, flags);
1100
1101         if (last_alive != 0)
1102                 *when = last_alive;
1103
1104         /* peer is not persistent in hash, trigger peer creation
1105          * and connection establishment with a NULL tx */
1106         if (peer == NULL)
1107                 kiblnd_launch_tx(ni, NULL, nid);
1108
1109         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1110                libcfs_nid2str(nid), peer,
1111                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1112         return;
1113 }
1114
1115 void
1116 kiblnd_free_pages (kib_pages_t *p)
1117 {
1118         int         npages = p->ibp_npages;
1119         int         i;
1120
1121         LASSERT (p->ibp_device == NULL);
1122
1123         for (i = 0; i < npages; i++)
1124                 if (p->ibp_pages[i] != NULL)
1125                         __free_page(p->ibp_pages[i]);
1126
1127         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1128 }
1129
1130 int
1131 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1132 {
1133         kib_pages_t   *p;
1134         int            i;
1135
1136         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1137         if (p == NULL) {
1138                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1139                 return -ENOMEM;
1140         }
1141
1142         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1143         p->ibp_npages = npages;
1144
1145         for (i = 0; i < npages; i++) {
1146                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1147                 if (p->ibp_pages[i] == NULL) {
1148                         CERROR("Can't allocate page %d of %d\n", i, npages);
1149                         kiblnd_free_pages(p);
1150                         return -ENOMEM;
1151                 }
1152         }
1153
1154         *pp = p;
1155         return 0;
1156 }
1157
1158 void
1159 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1160 {
1161         kib_rx_t *rx;
1162         int       i;
1163
1164         LASSERT (conn->ibc_rxs != NULL);
1165         LASSERT (conn->ibc_rx_pages->ibp_device != NULL);
1166
1167         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1168                 rx = &conn->ibc_rxs[i];
1169
1170                 LASSERT (rx->rx_nob >= 0); /* not posted */
1171
1172                 kiblnd_dma_unmap_single(conn->ibc_rx_pages->ibp_device,
1173                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1174                                                           rx->rx_msgaddr),
1175                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1176         }
1177
1178         conn->ibc_rx_pages->ibp_device = NULL;
1179
1180         kiblnd_free_pages(conn->ibc_rx_pages);
1181
1182         conn->ibc_rx_pages = NULL;
1183 }
1184
1185 void
1186 kiblnd_map_rx_descs(kib_conn_t *conn)
1187 {
1188         kib_rx_t       *rx;
1189         struct page    *pg;
1190         int             pg_off;
1191         int             ipg;
1192         int             i;
1193
1194         for (pg_off = ipg = i = 0;
1195              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1196                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1197                 rx = &conn->ibc_rxs[i];
1198
1199                 rx->rx_conn = conn;
1200                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1201
1202                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_cmid->device,
1203                                                        rx->rx_msg, IBLND_MSG_SIZE,
1204                                                        DMA_FROM_DEVICE);
1205                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_cmid->device,
1206                                                    rx->rx_msgaddr));
1207                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1208
1209                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1210                        i, rx->rx_msg, rx->rx_msgaddr,
1211                        lnet_page2phys(pg) + pg_off);
1212
1213                 pg_off += IBLND_MSG_SIZE;
1214                 LASSERT (pg_off <= PAGE_SIZE);
1215
1216                 if (pg_off == PAGE_SIZE) {
1217                         pg_off = 0;
1218                         ipg++;
1219                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1220                 }
1221         }
1222
1223         conn->ibc_rx_pages->ibp_device = conn->ibc_cmid->device;
1224 }
1225
1226 static void
1227 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1228 {
1229         kib_tx_t       *tx;
1230         int             i;
1231
1232         LASSERT (tpo->tpo_pool.po_allocated == 0);
1233         LASSERT (tpo->tpo_tx_pages->ibp_device != NULL);
1234
1235         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1236                 tx = &tpo->tpo_tx_descs[i];
1237                 kiblnd_dma_unmap_single(tpo->tpo_tx_pages->ibp_device,
1238                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1239                                                           tx->tx_msgaddr),
1240                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1241         }
1242         tpo->tpo_tx_pages->ibp_device = NULL;
1243 }
1244
1245 static void
1246 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1247 {
1248         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1249         kib_pool_t     *pool  = &tpo->tpo_pool;
1250         kib_net_t      *net   = pool->po_owner->ps_net;
1251         struct page    *page;
1252         kib_tx_t       *tx;
1253         int             page_offset;
1254         int             ipage;
1255         int             i;
1256
1257         LASSERT (net != NULL);
1258         LASSERT (net->ibn_dev->ibd_cmid != NULL &&
1259                  net->ibn_dev->ibd_cmid->device != NULL);
1260
1261         /* pre-mapped messages are not bigger than 1 page */
1262         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1263
1264         /* No fancy arithmetic when we do the buffer calculations */
1265         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1266
1267         txpgs->ibp_device = net->ibn_dev->ibd_cmid->device;
1268
1269         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1270                 page = txpgs->ibp_pages[ipage];
1271                 tx = &tpo->tpo_tx_descs[i];
1272
1273                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1274                                            page_offset);
1275
1276                 tx->tx_msgaddr = kiblnd_dma_map_single(
1277                         txpgs->ibp_device, tx->tx_msg,
1278                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1279                 LASSERT (!kiblnd_dma_mapping_error(txpgs->ibp_device,
1280                                                    tx->tx_msgaddr));
1281                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1282
1283                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1284
1285                 page_offset += IBLND_MSG_SIZE;
1286                 LASSERT (page_offset <= PAGE_SIZE);
1287
1288                 if (page_offset == PAGE_SIZE) {
1289                         page_offset = 0;
1290                         ipage++;
1291                         LASSERT (ipage <= txpgs->ibp_npages);
1292                 }
1293         }
1294 }
1295
1296 struct ib_mr *
1297 kiblnd_find_dma_mr(kib_net_t *net, __u64 addr, __u64 size)
1298 {
1299         __u64   index;
1300
1301         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1302
1303         if (net->ibn_dev->ibd_nmrs == 1)
1304                 return net->ibn_dev->ibd_mrs[0];
1305
1306         index = addr >> net->ibn_dev->ibd_mr_shift;
1307
1308         if (index <  net->ibn_dev->ibd_nmrs &&
1309             index == ((addr + size - 1) >> net->ibn_dev->ibd_mr_shift))
1310                 return net->ibn_dev->ibd_mrs[index];
1311
1312         return NULL;
1313 }
1314
1315 struct ib_mr *
1316 kiblnd_find_rd_dma_mr(kib_net_t *net, kib_rdma_desc_t *rd)
1317 {
1318         struct ib_mr *prev_mr;
1319         struct ib_mr *mr;
1320         int           i;
1321
1322         LASSERT (net->ibn_dev->ibd_mrs[0] != NULL);
1323
1324         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1325             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1326                 return NULL;
1327
1328         if (net->ibn_dev->ibd_nmrs == 1)
1329                 return net->ibn_dev->ibd_mrs[0];
1330
1331         for (i = 0, mr = prev_mr = NULL;
1332              i < rd->rd_nfrags; i++) {
1333                 mr = kiblnd_find_dma_mr(net,
1334                                         rd->rd_frags[i].rf_addr,
1335                                         rd->rd_frags[i].rf_nob);
1336                 if (prev_mr == NULL)
1337                         prev_mr = mr;
1338
1339                 if (mr == NULL || prev_mr != mr) {
1340                         /* Can't covered by one single MR */
1341                         mr = NULL;
1342                         break;
1343                 }
1344         }
1345
1346         return mr;
1347 }
1348
1349 void
1350 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1351 {
1352         LASSERT (pool->fpo_map_count == 0);
1353
1354         if (pool->fpo_fmr_pool != NULL)
1355                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1356
1357         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1358 }
1359
1360 void
1361 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1362 {
1363         kib_fmr_pool_t *pool;
1364
1365         while (!cfs_list_empty(head)) {
1366                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1367                 cfs_list_del(&pool->fpo_list);
1368                 kiblnd_destroy_fmr_pool(pool);
1369         }
1370 }
1371
1372 int
1373 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1374 {
1375         /* FMR pool for RDMA */
1376         kib_fmr_pool_t          *fpo;
1377         struct ib_fmr_pool_param param = {
1378                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1379                 .page_shift        = PAGE_SHIFT,
1380                 .access            = (IB_ACCESS_LOCAL_WRITE |
1381                                       IB_ACCESS_REMOTE_WRITE),
1382                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1383                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1384                 .flush_function    = NULL,
1385                 .flush_arg         = NULL,
1386                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1387         int rc;
1388
1389         LASSERT (fps->fps_net->ibn_dev != NULL &&
1390                  fps->fps_net->ibn_dev->ibd_pd != NULL);
1391
1392         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1393         if (fpo == NULL)
1394                 return -ENOMEM;
1395
1396         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1397         fpo->fpo_fmr_pool = ib_create_fmr_pool(fps->fps_net->ibn_dev->ibd_pd, &param);
1398         if (IS_ERR(fpo->fpo_fmr_pool)) {
1399                 CERROR("Failed to create FMR pool: %ld\n",
1400                        PTR_ERR(fpo->fpo_fmr_pool));
1401                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1402                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1403                 return rc;
1404         }
1405
1406         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1407         fpo->fpo_owner    = fps;
1408         *pp_fpo = fpo;
1409
1410         return 0;
1411 }
1412
1413 static void
1414 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1415 {
1416         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1417 }
1418
1419 static int
1420 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1421 {
1422         kib_fmr_pool_t *fpo;
1423         int             rc;
1424
1425         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1426
1427         fps->fps_net = net;
1428         cfs_spin_lock_init(&fps->fps_lock);
1429         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1430         rc = kiblnd_create_fmr_pool(fps, &fpo);
1431         if (rc == 0)
1432                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1433
1434         return rc;
1435 }
1436
1437 void
1438 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1439 {
1440         CFS_LIST_HEAD     (zombies);
1441         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1442         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1443         kib_fmr_pool_t    *tmp;
1444         int                rc;
1445
1446         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1447         LASSERT (rc == 0);
1448
1449         if (status != 0) {
1450                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1451                 LASSERT (rc == 0);
1452         }
1453
1454         fmr->fmr_pool = NULL;
1455         fmr->fmr_pfmr = NULL;
1456
1457         cfs_spin_lock(&fps->fps_lock);
1458         fpo->fpo_map_count --;  /* decref the pool */
1459
1460         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1461                 /* the first pool is persistent */
1462                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1463                         continue;
1464
1465                 if (fpo->fpo_map_count == 0 &&  /* no more reference */
1466                     cfs_time_aftereq(cfs_time_current(), fpo->fpo_deadline)) {
1467                         cfs_list_move(&fpo->fpo_list, &zombies);
1468                         fps->fps_version ++;
1469                 }
1470         }
1471         cfs_spin_unlock(&fps->fps_lock);
1472
1473         if (!cfs_list_empty(&zombies))
1474                 kiblnd_destroy_fmr_pool_list(&zombies);
1475 }
1476
1477 int
1478 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1479                     __u64 iov, kib_fmr_t *fmr)
1480 {
1481         struct ib_pool_fmr *pfmr;
1482         kib_fmr_pool_t     *fpo;
1483         __u64               version;
1484         int                 rc;
1485
1486         LASSERT (fps->fps_net->ibn_with_fmr);
1487  again:
1488         cfs_spin_lock(&fps->fps_lock);
1489         version = fps->fps_version;
1490         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1491                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1492                 fpo->fpo_map_count ++;
1493                 cfs_spin_unlock(&fps->fps_lock);
1494
1495                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1496                                             pages, npages, iov);
1497                 if (likely(!IS_ERR(pfmr))) {
1498                         fmr->fmr_pool = fpo;
1499                         fmr->fmr_pfmr = pfmr;
1500                         return 0;
1501                 }
1502
1503                 cfs_spin_lock(&fps->fps_lock);
1504                 fpo->fpo_map_count --;
1505                 if (PTR_ERR(pfmr) != -EAGAIN) {
1506                         cfs_spin_unlock(&fps->fps_lock);
1507                         return PTR_ERR(pfmr);
1508                 }
1509
1510                 /* EAGAIN and ... */
1511                 if (version != fps->fps_version) {
1512                         cfs_spin_unlock(&fps->fps_lock);
1513                         goto again;
1514                 }
1515         }
1516
1517         if (fps->fps_increasing) {
1518                 cfs_spin_unlock(&fps->fps_lock);
1519                 CDEBUG(D_NET, "Another thread is allocating new "
1520                               "FMR pool, waiting for her to complete\n");
1521                 cfs_schedule();
1522                 goto again;
1523
1524         }
1525
1526         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1527                 /* someone failed recently */
1528                 cfs_spin_unlock(&fps->fps_lock);
1529                 return -EAGAIN;
1530         }
1531
1532         fps->fps_increasing = 1;
1533         cfs_spin_unlock(&fps->fps_lock);
1534
1535         CDEBUG(D_NET, "Allocate new FMR pool\n");
1536         rc = kiblnd_create_fmr_pool(fps, &fpo);
1537         cfs_spin_lock(&fps->fps_lock);
1538         fps->fps_increasing = 0;
1539         if (rc == 0) {
1540                 fps->fps_version ++;
1541                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1542         } else {
1543                 fps->fps_next_retry = cfs_time_shift(10);
1544         }
1545         cfs_spin_unlock(&fps->fps_lock);
1546
1547         goto again;
1548 }
1549
1550 static void
1551 kiblnd_fini_pool(kib_pool_t *pool)
1552 {
1553         LASSERT (cfs_list_empty(&pool->po_free_list));
1554         LASSERT (pool->po_allocated == 0);
1555
1556         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1557 }
1558
1559 static void
1560 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1561 {
1562         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1563
1564         memset(pool, 0, sizeof(kib_pool_t));
1565         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1566         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1567         pool->po_owner    = ps;
1568         pool->po_size     = size;
1569 }
1570
1571 void
1572 kiblnd_destroy_pool_list(kib_poolset_t *ps, cfs_list_t *head)
1573 {
1574         kib_pool_t *pool;
1575
1576         while (!cfs_list_empty(head)) {
1577                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1578                 cfs_list_del(&pool->po_list);
1579                 ps->ps_pool_destroy(pool);
1580         }
1581 }
1582
1583 static void
1584 kiblnd_fini_pool_set(kib_poolset_t *ps)
1585 {
1586         kiblnd_destroy_pool_list(ps, &ps->ps_pool_list);
1587 }
1588
1589 static int
1590 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1591                      char *name, int size,
1592                      kib_ps_pool_create_t po_create,
1593                      kib_ps_pool_destroy_t po_destroy,
1594                      kib_ps_node_init_t nd_init,
1595                      kib_ps_node_fini_t nd_fini)
1596 {
1597         kib_pool_t    *pool;
1598         int            rc;
1599
1600         memset(ps, 0, sizeof(kib_poolset_t));
1601
1602         ps->ps_net          = net;
1603         ps->ps_pool_create  = po_create;
1604         ps->ps_pool_destroy = po_destroy;
1605         ps->ps_node_init    = nd_init;
1606         ps->ps_node_fini    = nd_fini;
1607         ps->ps_pool_size    = size;
1608         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1609         cfs_spin_lock_init(&ps->ps_lock);
1610         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1611
1612         rc = ps->ps_pool_create(ps, size, &pool);
1613         if (rc == 0)
1614                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1615         else
1616                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1617
1618         return rc;
1619 }
1620
1621 void
1622 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1623 {
1624         CFS_LIST_HEAD  (zombies);
1625         kib_poolset_t  *ps = pool->po_owner;
1626         kib_pool_t     *tmp;
1627         cfs_time_t      now = cfs_time_current();
1628
1629         cfs_spin_lock(&ps->ps_lock);
1630
1631         if (ps->ps_node_fini != NULL)
1632                 ps->ps_node_fini(pool, node);
1633
1634         LASSERT (pool->po_allocated > 0);
1635         cfs_list_add(node, &pool->po_free_list);
1636         pool->po_allocated --;
1637
1638         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1639                 /* the first pool is persistent */
1640                 if (ps->ps_pool_list.next == &pool->po_list)
1641                         continue;
1642
1643                 if (pool->po_allocated == 0 &&
1644                     cfs_time_aftereq(now, pool->po_deadline))
1645                         cfs_list_move(&pool->po_list, &zombies);
1646         }
1647         cfs_spin_unlock(&ps->ps_lock);
1648
1649         if (!cfs_list_empty(&zombies))
1650                 kiblnd_destroy_pool_list(ps, &zombies);
1651 }
1652
1653 cfs_list_t *
1654 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1655 {
1656         cfs_list_t            *node;
1657         kib_pool_t            *pool;
1658         int                    rc;
1659
1660  again:
1661         cfs_spin_lock(&ps->ps_lock);
1662         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1663                 if (cfs_list_empty(&pool->po_free_list))
1664                         continue;
1665
1666                 pool->po_allocated ++;
1667                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1668                 node = pool->po_free_list.next;
1669                 cfs_list_del(node);
1670
1671                 if (ps->ps_node_init != NULL) {
1672                         /* still hold the lock */
1673                         ps->ps_node_init(pool, node);
1674                 }
1675                 cfs_spin_unlock(&ps->ps_lock);
1676                 return node;
1677         }
1678
1679         /* no available tx pool and ... */
1680         if (ps->ps_increasing) {
1681                 /* another thread is allocating a new pool */
1682                 cfs_spin_unlock(&ps->ps_lock);
1683                 CDEBUG(D_NET, "Another thread is allocating new "
1684                        "%s pool, waiting for her to complete\n",
1685                        ps->ps_name);
1686                 cfs_schedule();
1687                 goto again;
1688         }
1689
1690         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1691                 /* someone failed recently */
1692                 cfs_spin_unlock(&ps->ps_lock);
1693                 return NULL;
1694         }
1695
1696         ps->ps_increasing = 1;
1697         cfs_spin_unlock(&ps->ps_lock);
1698
1699         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1700
1701         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1702
1703         cfs_spin_lock(&ps->ps_lock);
1704         ps->ps_increasing = 0;
1705         if (rc == 0) {
1706                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1707         } else {
1708                 /* retry 10 seconds later */
1709                 ps->ps_next_retry = cfs_time_shift(10);
1710                 CERROR("Can't allocate new %s pool because out of memory\n",
1711                        ps->ps_name);
1712         }
1713         cfs_spin_unlock(&ps->ps_lock);
1714
1715         goto again;
1716 }
1717
1718 void
1719 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1720 {
1721         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1722         struct ib_mr        *mr  = pmr->pmr_mr;
1723
1724         pmr->pmr_mr = NULL;
1725         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1726         if (mr != NULL)
1727                 ib_dereg_mr(mr);
1728 }
1729
1730 int
1731 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
1732                     __u64 *iova, kib_phys_mr_t **pp_pmr)
1733 {
1734         kib_phys_mr_t *pmr;
1735         cfs_list_t    *node;
1736         int            rc;
1737         int            i;
1738
1739         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1740         if (node == NULL) {
1741                 CERROR("Failed to allocate PMR descriptor\n");
1742                 return -ENOMEM;
1743         }
1744
1745         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1746         for (i = 0; i < rd->rd_nfrags; i ++) {
1747                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1748                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1749         }
1750
1751         pmr->pmr_mr = ib_reg_phys_mr(pps->pps_poolset.ps_net->ibn_dev->ibd_pd,
1752                                      pmr->pmr_ipb, rd->rd_nfrags,
1753                                      IB_ACCESS_LOCAL_WRITE |
1754                                      IB_ACCESS_REMOTE_WRITE,
1755                                      iova);
1756         if (!IS_ERR(pmr->pmr_mr)) {
1757                 pmr->pmr_iova = *iova;
1758                 *pp_pmr = pmr;
1759                 return 0;
1760         }
1761
1762         rc = PTR_ERR(pmr->pmr_mr);
1763         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1764
1765         pmr->pmr_mr = NULL;
1766         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1767
1768         return rc;
1769 }
1770
1771 static void
1772 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1773 {
1774         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1775         kib_phys_mr_t  *pmr;
1776
1777         LASSERT (pool->po_allocated == 0);
1778
1779         while (!cfs_list_empty(&pool->po_free_list)) {
1780                 pmr = cfs_list_entry(pool->po_free_list.next,
1781                                      kib_phys_mr_t, pmr_list);
1782
1783                 LASSERT (pmr->pmr_mr == NULL);
1784                 cfs_list_del(&pmr->pmr_list);
1785
1786                 if (pmr->pmr_ipb != NULL) {
1787                         LIBCFS_FREE(pmr->pmr_ipb,
1788                                     IBLND_MAX_RDMA_FRAGS *
1789                                     sizeof(struct ib_phys_buf));
1790                 }
1791
1792                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1793         }
1794
1795         kiblnd_fini_pool(pool);
1796         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1797 }
1798
1799 static int
1800 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1801 {
1802         kib_pmr_pool_t      *ppo;
1803         kib_pool_t          *pool;
1804         kib_phys_mr_t       *pmr;
1805         int                  i;
1806
1807         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1808         if (ppo == NULL) {
1809                 CERROR("Failed to allocate PMR pool\n");
1810                 return -ENOMEM;
1811         }
1812
1813         pool = &ppo->ppo_pool;
1814         kiblnd_init_pool(ps, pool, size);
1815
1816         for (i = 0; i < size; i++) {
1817                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1818                 if (pmr == NULL)
1819                         break;
1820
1821                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1822                 pmr->pmr_pool = ppo;
1823                 LIBCFS_ALLOC(pmr->pmr_ipb,
1824                              IBLND_MAX_RDMA_FRAGS *
1825                              sizeof(struct ib_phys_buf));
1826                 if (pmr->pmr_ipb == NULL)
1827                         break;
1828
1829                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1830         }
1831
1832         if (i < size) {
1833                 ps->ps_pool_destroy(pool);
1834                 return -ENOMEM;
1835         }
1836
1837         *pp_po = pool;
1838         return 0;
1839 }
1840
1841 static void
1842 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1843 {
1844         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1845         int             i;
1846
1847         LASSERT (pool->po_allocated == 0);
1848
1849         if (tpo->tpo_tx_pages != NULL) {
1850                 if (tpo->tpo_tx_pages->ibp_device != NULL)
1851                         kiblnd_unmap_tx_pool(tpo);
1852                 kiblnd_free_pages(tpo->tpo_tx_pages);
1853         }
1854
1855         if (tpo->tpo_tx_descs == NULL)
1856                 goto out;
1857
1858         for (i = 0; i < pool->po_size; i++) {
1859                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1860
1861                 cfs_list_del(&tx->tx_list);
1862                 if (tx->tx_pages != NULL)
1863                         LIBCFS_FREE(tx->tx_pages,
1864                                     LNET_MAX_IOV *
1865                                     sizeof(*tx->tx_pages));
1866                 if (tx->tx_frags != NULL)
1867                         LIBCFS_FREE(tx->tx_frags,
1868                                     IBLND_MAX_RDMA_FRAGS *
1869                                             sizeof(*tx->tx_frags));
1870                 if (tx->tx_wrq != NULL)
1871                         LIBCFS_FREE(tx->tx_wrq,
1872                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1873                                     sizeof(*tx->tx_wrq));
1874                 if (tx->tx_sge != NULL)
1875                         LIBCFS_FREE(tx->tx_sge,
1876                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1877                                     sizeof(*tx->tx_sge));
1878                 if (tx->tx_rd != NULL)
1879                         LIBCFS_FREE(tx->tx_rd,
1880                                     offsetof(kib_rdma_desc_t,
1881                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1882         }
1883
1884         LIBCFS_FREE(tpo->tpo_tx_descs,
1885                     pool->po_size * sizeof(kib_tx_t));
1886 out:
1887         kiblnd_fini_pool(pool);
1888         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1889 }
1890
1891 static int
1892 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1893 {
1894         int            i;
1895         int            npg;
1896         kib_pool_t    *pool;
1897         kib_tx_pool_t *tpo;
1898
1899         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
1900         if (tpo == NULL) {
1901                 CERROR("Failed to allocate TX pool\n");
1902                 return -ENOMEM;
1903         }
1904
1905         pool = &tpo->tpo_pool;
1906         kiblnd_init_pool(ps, pool, size);
1907         tpo->tpo_tx_descs = NULL;
1908         tpo->tpo_tx_pages = NULL;
1909
1910         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1911         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
1912                 CERROR("Can't allocate tx pages: %d\n", npg);
1913                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1914                 return -ENOMEM;
1915         }
1916
1917         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
1918         if (tpo->tpo_tx_descs == NULL) {
1919                 CERROR("Can't allocate %d tx descriptors\n", size);
1920                 ps->ps_pool_destroy(pool);
1921                 return -ENOMEM;
1922         }
1923
1924         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1925
1926         for (i = 0; i < size; i++) {
1927                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1928
1929                 tx->tx_pool = tpo;
1930                 if (ps->ps_net->ibn_with_fmr){
1931                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1932                                      sizeof(*tx->tx_pages));
1933                         if (tx->tx_pages == NULL)
1934                                 break;
1935                 }
1936
1937                 LIBCFS_ALLOC(tx->tx_frags,
1938                              IBLND_MAX_RDMA_FRAGS *
1939                              sizeof(*tx->tx_frags));
1940                 if (tx->tx_frags == NULL)
1941                         break;
1942
1943                 LIBCFS_ALLOC(tx->tx_wrq,
1944                              (1 + IBLND_MAX_RDMA_FRAGS) *
1945                              sizeof(*tx->tx_wrq));
1946                 if (tx->tx_wrq == NULL)
1947                         break;
1948
1949                 LIBCFS_ALLOC(tx->tx_sge,
1950                              (1 + IBLND_MAX_RDMA_FRAGS) *
1951                              sizeof(*tx->tx_sge));
1952                 if (tx->tx_sge == NULL)
1953                         break;
1954
1955                 LIBCFS_ALLOC(tx->tx_rd,
1956                              offsetof(kib_rdma_desc_t,
1957                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1958                 if (tx->tx_rd == NULL)
1959                         break;
1960         }
1961
1962         if (i == size) {
1963                 kiblnd_map_tx_pool(tpo);
1964                 *pp_po = pool;
1965                 return 0;
1966         }
1967
1968         ps->ps_pool_destroy(pool);
1969         return -ENOMEM;
1970 }
1971
1972 static void
1973 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
1974 {
1975         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1976                                              tps_poolset);
1977         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
1978
1979         tx->tx_cookie = tps->tps_next_tx_cookie ++;
1980 }
1981
1982 void
1983 kiblnd_ni_fini_pools(kib_net_t *net)
1984 {
1985         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
1986         if (net->ibn_with_fmr)
1987                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
1988         else if (net->ibn_with_pmr)
1989                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
1990 }
1991
1992 int
1993 kiblnd_net_init_pools(kib_net_t *net)
1994 {
1995         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
1996         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
1997         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
1998         int                rc;
1999
2000         if (*kiblnd_tunables.kib_fmr_pool_size <
2001             *kiblnd_tunables.kib_ntx / 4) {
2002                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2003                        *kiblnd_tunables.kib_fmr_pool_size,
2004                        *kiblnd_tunables.kib_ntx / 4);
2005                 return -EINVAL;
2006         }
2007
2008         if (*kiblnd_tunables.kib_pmr_pool_size <
2009             *kiblnd_tunables.kib_ntx / 4) {
2010                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2011                        *kiblnd_tunables.kib_pmr_pool_size,
2012                        *kiblnd_tunables.kib_ntx / 4);
2013                 return -EINVAL;
2014         }
2015
2016         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2017             net->ibn_dev->ibd_nmrs > 1) { /* premapping can fail if ibd_nmr > 1,
2018                                            * so we always create FMR/PMR pool and
2019                                            * map-on-demand if premapping failed */
2020                 rc = kiblnd_init_fmr_pool_set(fps, net);
2021                 if (rc == 0) {
2022                         net->ibn_with_fmr = 1;
2023                 } else if (rc == -ENOSYS) {
2024                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2025                                                   *kiblnd_tunables.kib_pmr_pool_size,
2026                                                   kiblnd_create_pmr_pool,
2027                                                   kiblnd_destroy_pmr_pool,
2028                                                   NULL, NULL);
2029                         if (rc == 0)
2030                                 net->ibn_with_pmr = 1;
2031                 }
2032                 if (rc != 0)
2033                         return rc;
2034         }
2035
2036         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2037                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2038                                   kiblnd_tx_init, NULL);
2039         if (rc == 0)
2040                 return 0;
2041
2042         if (net->ibn_with_fmr)
2043                 kiblnd_fini_fmr_pool_set(fps);
2044         else if (net->ibn_with_pmr)
2045                 kiblnd_fini_pool_set(&pps->pps_poolset);
2046
2047         return rc;
2048 }
2049
2050 void
2051 kiblnd_dev_cleanup(kib_dev_t *ibdev)
2052 {
2053         int     i;
2054
2055         if (ibdev->ibd_mrs == NULL)
2056                 return;
2057
2058         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2059                 if (ibdev->ibd_mrs[i] == NULL)
2060                         break;
2061
2062                 ib_dereg_mr(ibdev->ibd_mrs[i]);
2063         }
2064
2065         LIBCFS_FREE(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2066         ibdev->ibd_mrs = NULL;
2067 }
2068
2069 static int
2070 kiblnd_dev_get_attr(kib_dev_t *ibdev)
2071 {
2072         struct ib_device_attr *attr;
2073         int                    rc;
2074
2075         /* It's safe to assume a HCA can handle a page size
2076          * matching that of the native system */
2077         ibdev->ibd_page_shift = PAGE_SHIFT;
2078         ibdev->ibd_page_size  = 1 << PAGE_SHIFT;
2079         ibdev->ibd_page_mask  = ~((__u64)ibdev->ibd_page_size - 1);
2080
2081         LIBCFS_ALLOC(attr, sizeof(*attr));
2082         if (attr == NULL) {
2083                 CERROR("Out of memory\n");
2084                 return -ENOMEM;
2085         }
2086
2087         rc = ib_query_device(ibdev->ibd_cmid->device, attr);
2088         if (rc == 0)
2089                 ibdev->ibd_mr_size = attr->max_mr_size;
2090
2091         LIBCFS_FREE(attr, sizeof(*attr));
2092
2093         if (rc != 0) {
2094                 CERROR("Failed to query IB device: %d\n", rc);
2095                 return rc;
2096         }
2097
2098 #ifdef HAVE_OFED_TRANSPORT_IWARP
2099         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2100          * and we have reported the bug, remove these in the future when Chelsio
2101          * bug got fixed. */
2102         if (rdma_node_get_transport(ibdev->ibd_cmid->device->node_type) ==
2103             RDMA_TRANSPORT_IWARP)
2104                 ibdev->ibd_mr_size = (1ULL << 32) - 1;
2105 #endif
2106
2107         if (ibdev->ibd_mr_size == ~0ULL) {
2108                 ibdev->ibd_mr_shift = 64;
2109                 return 0;
2110         }
2111
2112         for (ibdev->ibd_mr_shift = 0;
2113              ibdev->ibd_mr_shift < 64; ibdev->ibd_mr_shift ++) {
2114                 if (ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) ||
2115                     ibdev->ibd_mr_size == (1ULL << ibdev->ibd_mr_shift) - 1)
2116                         return 0;
2117         }
2118
2119         CERROR("Invalid mr size: "LPX64"\n", ibdev->ibd_mr_size);
2120         return -EINVAL;
2121 }
2122
2123 int
2124 kiblnd_dev_setup(kib_dev_t *ibdev)
2125 {
2126         struct ib_mr *mr;
2127         int           i;
2128         int           rc;
2129         __u64         mm_size;
2130         __u64         mr_size;
2131         int           acflags = IB_ACCESS_LOCAL_WRITE |
2132                                 IB_ACCESS_REMOTE_WRITE;
2133
2134         rc = kiblnd_dev_get_attr(ibdev);
2135         if (rc != 0)
2136                 return rc;
2137
2138         if (ibdev->ibd_mr_shift == 64) {
2139                 LIBCFS_ALLOC(ibdev->ibd_mrs, 1 * sizeof(*ibdev->ibd_mrs));
2140                 if (ibdev->ibd_mrs == NULL) {
2141                         CERROR("Failed to allocate MRs table\n");
2142                         return -ENOMEM;
2143                 }
2144
2145                 ibdev->ibd_mrs[0] = NULL;
2146                 ibdev->ibd_nmrs   = 1;
2147
2148                 mr = ib_get_dma_mr(ibdev->ibd_pd, acflags);
2149                 if (IS_ERR(mr)) {
2150                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2151                         kiblnd_dev_cleanup(ibdev);
2152                         return PTR_ERR(mr);
2153                 }
2154
2155                 ibdev->ibd_mrs[0] = mr;
2156
2157                 goto out;
2158         }
2159
2160         mr_size = (1ULL << ibdev->ibd_mr_shift);
2161         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2162
2163         ibdev->ibd_nmrs = (int)((mm_size + mr_size - 1) >> ibdev->ibd_mr_shift);
2164
2165         if (ibdev->ibd_mr_shift < 32 || ibdev->ibd_nmrs > 1024) {
2166                 /* it's 4T..., assume we will re-code at that time */
2167                 CERROR("Can't support memory size: x"LPX64
2168                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2169                 return -EINVAL;
2170         }
2171
2172         /* create an array of MRs to cover all memory */
2173         LIBCFS_ALLOC(ibdev->ibd_mrs, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2174         if (ibdev->ibd_mrs == NULL) {
2175                 CERROR("Failed to allocate MRs' table\n");
2176                 return -ENOMEM;
2177         }
2178
2179         memset(ibdev->ibd_mrs, 0, sizeof(*ibdev->ibd_mrs) * ibdev->ibd_nmrs);
2180
2181         for (i = 0; i < ibdev->ibd_nmrs; i++) {
2182                 struct ib_phys_buf ipb;
2183                 __u64              iova;
2184
2185                 ipb.size = ibdev->ibd_mr_size;
2186                 ipb.addr = i * mr_size;
2187                 iova     = ipb.addr;
2188
2189                 mr = ib_reg_phys_mr(ibdev->ibd_pd, &ipb, 1, acflags, &iova);
2190                 if (IS_ERR(mr)) {
2191                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2192                                " size "LPX64" : %ld\n",
2193                                ipb.addr, ipb.size, PTR_ERR(mr));
2194                         kiblnd_dev_cleanup(ibdev);
2195                         return PTR_ERR(mr);
2196                 }
2197
2198                 LASSERT (iova == ipb.addr);
2199
2200                 ibdev->ibd_mrs[i] = mr;
2201         }
2202
2203 out:
2204         CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2205                           LPX64", array size: %d\n",
2206                           ibdev->ibd_mr_size, ibdev->ibd_nmrs);
2207
2208         cfs_list_add_tail(&ibdev->ibd_list,
2209                       &kiblnd_data.kib_devs);
2210         return 0;
2211 }
2212
2213 void
2214 kiblnd_destroy_dev (kib_dev_t *dev)
2215 {
2216         LASSERT (dev->ibd_nnets == 0);
2217
2218         if (!cfs_list_empty(&dev->ibd_list)) /* on kib_devs? */
2219                 cfs_list_del_init(&dev->ibd_list);
2220
2221         kiblnd_dev_cleanup(dev);
2222
2223         if (dev->ibd_pd != NULL)
2224                 ib_dealloc_pd(dev->ibd_pd);
2225
2226         if (dev->ibd_cmid != NULL)
2227                 rdma_destroy_id(dev->ibd_cmid);
2228
2229         LIBCFS_FREE(dev, sizeof(*dev));
2230 }
2231
2232 void
2233 kiblnd_base_shutdown (void)
2234 {
2235         int i;
2236
2237         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2238
2239         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2240                cfs_atomic_read(&libcfs_kmemory));
2241
2242         switch (kiblnd_data.kib_init) {
2243         default:
2244                 LBUG();
2245
2246         case IBLND_INIT_ALL:
2247         case IBLND_INIT_DATA:
2248                 LASSERT (kiblnd_data.kib_peers != NULL);
2249                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2250                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2251                 }
2252                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2253                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2254
2255                 /* flag threads to terminate; wake and wait for them to die */
2256                 kiblnd_data.kib_shutdown = 1;
2257                 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2258                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2259
2260                 i = 2;
2261                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2262                         i++;
2263                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2264                                "Waiting for %d threads to terminate\n",
2265                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2266                         cfs_pause(cfs_time_seconds(1));
2267                 }
2268
2269                 /* fall through */
2270
2271         case IBLND_INIT_NOTHING:
2272                 break;
2273         }
2274
2275         if (kiblnd_data.kib_peers != NULL)
2276                 LIBCFS_FREE(kiblnd_data.kib_peers,
2277                             sizeof(cfs_list_t) *
2278                             kiblnd_data.kib_peer_hash_size);
2279
2280         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2281                cfs_atomic_read(&libcfs_kmemory));
2282
2283         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2284         PORTAL_MODULE_UNUSE;
2285 }
2286
2287 void
2288 kiblnd_shutdown (lnet_ni_t *ni)
2289 {
2290         kib_net_t        *net = ni->ni_data;
2291         cfs_rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2292         int               i;
2293         unsigned long     flags;
2294
2295         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2296
2297         if (net == NULL)
2298                 goto out;
2299
2300         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2301                cfs_atomic_read(&libcfs_kmemory));
2302
2303         cfs_write_lock_irqsave(g_lock, flags);
2304         net->ibn_shutdown = 1;
2305         cfs_write_unlock_irqrestore(g_lock, flags);
2306
2307         switch (net->ibn_init) {
2308         default:
2309                 LBUG();
2310
2311         case IBLND_INIT_ALL:
2312                 /* nuke all existing peers within this net */
2313                 kiblnd_del_peer(ni, LNET_NID_ANY);
2314
2315                 /* Wait for all peer state to clean up */
2316                 i = 2;
2317                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2318                         i++;
2319                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2320                                "%s: waiting for %d peers to disconnect\n",
2321                                libcfs_nid2str(ni->ni_nid),
2322                                cfs_atomic_read(&net->ibn_npeers));
2323                         cfs_pause(cfs_time_seconds(1));
2324                 }
2325
2326                 kiblnd_ni_fini_pools(net);
2327
2328                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2329                 net->ibn_dev->ibd_nnets--;
2330
2331                 /* fall through */
2332
2333         case IBLND_INIT_NOTHING:
2334                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2335
2336                 if (net->ibn_dev != NULL &&
2337                     net->ibn_dev->ibd_nnets == 0)
2338                         kiblnd_destroy_dev(net->ibn_dev);
2339
2340                 break;
2341         }
2342
2343         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2344                cfs_atomic_read(&libcfs_kmemory));
2345
2346         net->ibn_init = IBLND_INIT_NOTHING;
2347         ni->ni_data = NULL;
2348
2349         LIBCFS_FREE(net, sizeof(*net));
2350
2351 out:
2352         if (cfs_list_empty(&kiblnd_data.kib_devs))
2353                 kiblnd_base_shutdown();
2354         return;
2355 }
2356
2357 int
2358 kiblnd_base_startup (void)
2359 {
2360         int i;
2361         int rc;
2362
2363         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2364
2365         PORTAL_MODULE_USE;
2366         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2367
2368         cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2369
2370         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2371
2372         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2373         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2374                      sizeof(cfs_list_t) *
2375                             kiblnd_data.kib_peer_hash_size);
2376         if (kiblnd_data.kib_peers == NULL) {
2377                 goto failed;
2378         }
2379         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2380                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2381
2382         cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2383         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2384         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2385         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2386
2387         cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2388         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2389         cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
2390
2391         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2392
2393         /* lists/ptrs/locks initialised */
2394         kiblnd_data.kib_init = IBLND_INIT_DATA;
2395         /*****************************************************/
2396
2397         for (i = 0; i < IBLND_N_SCHED; i++) {
2398                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2399                 if (rc != 0) {
2400                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2401                                i, rc);
2402                         goto failed;
2403                 }
2404         }
2405
2406         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2407         if (rc != 0) {
2408                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2409                 goto failed;
2410         }
2411
2412         /* flag everything initialised */
2413         kiblnd_data.kib_init = IBLND_INIT_ALL;
2414         /*****************************************************/
2415
2416         return 0;
2417
2418  failed:
2419         kiblnd_base_shutdown();
2420         return -ENETDOWN;
2421 }
2422
2423 int
2424 kiblnd_startup (lnet_ni_t *ni)
2425 {
2426         char                     *ifname;
2427         kib_dev_t                *ibdev = NULL;
2428         kib_net_t                *net;
2429         cfs_list_t               *tmp;
2430         struct timeval            tv;
2431         int                       rc;
2432
2433         LASSERT (ni->ni_lnd == &the_o2iblnd);
2434
2435         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2436                 rc = kiblnd_base_startup();
2437                 if (rc != 0)
2438                         return rc;
2439         }
2440
2441         LIBCFS_ALLOC(net, sizeof(*net));
2442         ni->ni_data = net;
2443         if (net == NULL)
2444                 goto failed;
2445
2446         memset(net, 0, sizeof(*net));
2447
2448         cfs_gettimeofday(&tv);
2449         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2450
2451         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2452         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2453         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2454         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2455
2456         if (ni->ni_interfaces[0] != NULL) {
2457                 /* Use the IPoIB interface specified in 'networks=' */
2458
2459                 CLASSERT (LNET_MAX_INTERFACES > 1);
2460                 if (ni->ni_interfaces[1] != NULL) {
2461                         CERROR("Multiple interfaces not supported\n");
2462                         goto failed;
2463                 }
2464
2465                 ifname = ni->ni_interfaces[0];
2466         } else {
2467                 ifname = *kiblnd_tunables.kib_default_ipif;
2468         }
2469
2470         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2471                 CERROR("IPoIB interface name too long: %s\n", ifname);
2472                 goto failed;
2473         }
2474
2475         cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2476                 ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
2477
2478                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2479                         break;
2480
2481                 ibdev = NULL;
2482         }
2483
2484         if (ibdev == NULL) {
2485                 __u32                     ip;
2486                 __u32                     netmask;
2487                 int                       up;
2488                 struct rdma_cm_id        *id;
2489                 struct ib_pd             *pd;
2490                 struct sockaddr_in        addr;
2491
2492                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2493                 if (rc != 0) {
2494                         CERROR("Can't query IPoIB interface %s: %d\n",
2495                                ifname, rc);
2496                         goto failed;
2497                 }
2498
2499                 if (!up) {
2500                         CERROR("Can't query IPoIB interface %s: it's down\n",
2501                                ifname);
2502                         goto failed;
2503                 }
2504
2505                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
2506                 if (ibdev == NULL)
2507                         goto failed;
2508
2509                 memset(ibdev, 0, sizeof(*ibdev));
2510
2511                 CFS_INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
2512                 ibdev->ibd_ifip = ip;
2513                 strcpy(&ibdev->ibd_ifname[0], ifname);
2514
2515                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
2516                 if (IS_ERR(id)) {
2517                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
2518                         goto failed;
2519                 }
2520
2521                 ibdev->ibd_cmid = id;
2522
2523                 memset(&addr, 0, sizeof(addr));
2524                 addr.sin_family      = AF_INET;
2525                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2526                 addr.sin_addr.s_addr = htonl(ip);
2527
2528                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
2529                 if (rc != 0) {
2530                         CERROR("Can't bind to %s: %d\n", ifname, rc);
2531                         goto failed;
2532                 }
2533
2534                 /* Binding should have assigned me an IB device */
2535                 LASSERT (id->device != NULL);
2536                 CDEBUG(D_CONSOLE, "Listener bound to %s:%u.%u.%u.%u:%d:%s\n",
2537                        ifname, HIPQUAD(ip), *kiblnd_tunables.kib_service,
2538                        id->device->name);
2539
2540                 pd = ib_alloc_pd(id->device);
2541                 if (IS_ERR(pd)) {
2542                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
2543                         goto failed;
2544                 }
2545
2546                 ibdev->ibd_pd = pd;
2547
2548                 rc = rdma_listen(id, 256);
2549                 if (rc != 0) {
2550                         CERROR("Can't start listener: %d\n", rc);
2551                         goto failed;
2552                 }
2553
2554                 rc = kiblnd_dev_setup(ibdev);
2555                 if (rc != 0) {
2556                         CERROR("Can't setup device: %d\n", rc);
2557                         goto failed;
2558                 }
2559         }
2560
2561         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2562         net->ibn_dev = ibdev;
2563
2564         rc = kiblnd_net_init_pools(net);
2565         if (rc != 0) {
2566                 CERROR("Failed to initialize NI pools: %d\n", rc);
2567                 goto failed;
2568         }
2569         ibdev->ibd_nnets++;
2570         net->ibn_init = IBLND_INIT_ALL;
2571
2572         return 0;
2573
2574 failed:
2575         if (net->ibn_dev == NULL && ibdev != NULL)
2576                 kiblnd_destroy_dev(ibdev);
2577
2578         kiblnd_shutdown(ni);
2579
2580         CDEBUG(D_NET, "kiblnd_startup failed\n");
2581         return -ENETDOWN;
2582 }
2583
2584 void __exit
2585 kiblnd_module_fini (void)
2586 {
2587         lnet_unregister_lnd(&the_o2iblnd);
2588         kiblnd_tunables_fini();
2589 }
2590
2591 int __init
2592 kiblnd_module_init (void)
2593 {
2594         int    rc;
2595
2596         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2597         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2598                   <= IBLND_MSG_SIZE);
2599         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2600                   <= IBLND_MSG_SIZE);
2601
2602         rc = kiblnd_tunables_init();
2603         if (rc != 0)
2604                 return rc;
2605
2606         lnet_register_lnd(&the_o2iblnd);
2607
2608         return 0;
2609 }
2610
2611 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2612 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2613 MODULE_LICENSE("GPL");
2614
2615 module_init(kiblnd_module_init);
2616 module_exit(kiblnd_module_fini);