lnet/klnds/o2iblnd/o2iblnd.c

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