lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  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 <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 kiblnd_cksum (void *ptr, int nob)
  50 {
  51         char  *c  = ptr;
  52         __u32  sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return (sum == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 340         peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
 341         peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 342         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 343
 344         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 345         INIT_LIST_HEAD(&peer->ibp_conns);
 346         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer (kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(kiblnd_peer_idle(peer));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 387         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head        *tmp;
 389         kib_peer_t              *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394                 LASSERT(!kiblnd_peer_idle(peer));
 395
 396                 if (peer->ibp_nid != nid)
 397                         continue;
 398
 399                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 400                        peer, libcfs_nid2str(nid),
 401                        atomic_read(&peer->ibp_refcount),
 402                        peer->ibp_version);
 403                 return peer;
 404         }
 405         return NULL;
 406 }
 407
 408 void
 409 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 410 {
 411         LASSERT(list_empty(&peer->ibp_conns));
 412
 413         LASSERT (kiblnd_peer_active(peer));
 414         list_del_init(&peer->ibp_list);
 415         /* lose peerlist's ref */
 416         kiblnd_peer_decref(peer);
 417 }
 418
 419 static int
 420 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 421                      lnet_nid_t *nidp, int *count)
 422 {
 423         kib_peer_t              *peer;
 424         struct list_head        *ptmp;
 425         int                      i;
 426         unsigned long            flags;
 427
 428         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 429
 430         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 431
 432                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 433
 434                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 435                         LASSERT(!kiblnd_peer_idle(peer));
 436
 437                         if (peer->ibp_ni != ni)
 438                                 continue;
 439
 440                         if (index-- > 0)
 441                                 continue;
 442
 443                         *nidp = peer->ibp_nid;
 444                         *count = atomic_read(&peer->ibp_refcount);
 445
 446                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 447                                                flags);
 448                         return 0;
 449                 }
 450         }
 451
 452         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 453         return -ENOENT;
 454 }
 455
 456 static void
 457 kiblnd_del_peer_locked (kib_peer_t *peer)
 458 {
 459         struct list_head        *ctmp;
 460         struct list_head        *cnxt;
 461         kib_conn_t              *conn;
 462
 463         if (list_empty(&peer->ibp_conns)) {
 464                 kiblnd_unlink_peer_locked(peer);
 465         } else {
 466                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 467                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 468
 469                         kiblnd_close_conn_locked(conn, 0);
 470                 }
 471                 /* NB closing peer's last conn unlinked it. */
 472         }
 473         /* NB peer now unlinked; might even be freed if the peer table had the
 474          * last ref on it. */
 475 }
 476
 477 static int
 478 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 479 {
 480         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 481         struct list_head        *ptmp;
 482         struct list_head        *pnxt;
 483         kib_peer_t              *peer;
 484         int                     lo;
 485         int                     hi;
 486         int                     i;
 487         unsigned long           flags;
 488         int                     rc = -ENOENT;
 489
 490         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 491
 492         if (nid != LNET_NID_ANY) {
 493                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 494         } else {
 495                 lo = 0;
 496                 hi = kiblnd_data.kib_peer_hash_size - 1;
 497         }
 498
 499         for (i = lo; i <= hi; i++) {
 500                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 501                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 502                         LASSERT(!kiblnd_peer_idle(peer));
 503
 504                         if (peer->ibp_ni != ni)
 505                                 continue;
 506
 507                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 508                                 continue;
 509
 510                         if (!list_empty(&peer->ibp_tx_queue)) {
 511                                 LASSERT(list_empty(&peer->ibp_conns));
 512
 513                                 list_splice_init(&peer->ibp_tx_queue,
 514                                                  &zombies);
 515                         }
 516
 517                         kiblnd_del_peer_locked(peer);
 518                         rc = 0;         /* matched something */
 519                 }
 520         }
 521
 522         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 523
 524         kiblnd_txlist_done(ni, &zombies, -EIO);
 525
 526         return rc;
 527 }
 528
 529 static kib_conn_t *
 530 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 531 {
 532         kib_peer_t              *peer;
 533         struct list_head        *ptmp;
 534         kib_conn_t              *conn;
 535         struct list_head        *ctmp;
 536         int                     i;
 537         unsigned long           flags;
 538
 539         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 540
 541         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 542                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 543
 544                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 545                         LASSERT(!kiblnd_peer_idle(peer));
 546
 547                         if (peer->ibp_ni != ni)
 548                                 continue;
 549
 550                         list_for_each(ctmp, &peer->ibp_conns) {
 551                                 if (index-- > 0)
 552                                         continue;
 553
 554                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 555                                 kiblnd_conn_addref(conn);
 556                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 557                                                        flags);
 558                                 return conn;
 559                         }
 560                 }
 561         }
 562
 563         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 564         return NULL;
 565 }
 566
 567 static void
 568 kiblnd_debug_rx (kib_rx_t *rx)
 569 {
 570         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 571                rx, rx->rx_status, rx->rx_msg->ibm_type,
 572                rx->rx_msg->ibm_credits);
 573 }
 574
 575 static void
 576 kiblnd_debug_tx (kib_tx_t *tx)
 577 {
 578         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 579                "cookie "LPX64" msg %s%s type %x cred %d\n",
 580                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 581                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 582                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 583                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 584                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 585 }
 586
 587 void
 588 kiblnd_debug_conn (kib_conn_t *conn)
 589 {
 590         struct list_head        *tmp;
 591         int                     i;
 592
 593         spin_lock(&conn->ibc_lock);
 594
 595         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 596                atomic_read(&conn->ibc_refcount), conn,
 597                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 598         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 599                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 600                conn->ibc_nsends_posted, conn->ibc_credits,
 601                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 602         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 603
 604         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 605         list_for_each(tmp, &conn->ibc_early_rxs)
 606                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 607
 608         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 609         list_for_each(tmp, &conn->ibc_tx_noops)
 610                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 611
 612         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 613         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 614                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 615
 616         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 617         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 618                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 621         list_for_each(tmp, &conn->ibc_tx_queue)
 622                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 623
 624         CDEBUG(D_CONSOLE, "   active_txs:\n");
 625         list_for_each(tmp, &conn->ibc_active_txs)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   rxs:\n");
 629         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 630                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 631
 632         spin_unlock(&conn->ibc_lock);
 633 }
 634
 635 int
 636 kiblnd_translate_mtu(int value)
 637 {
 638         switch (value) {
 639         default:
 640                 return -1;
 641         case 0:
 642                 return 0;
 643         case 256:
 644                 return IB_MTU_256;
 645         case 512:
 646                 return IB_MTU_512;
 647         case 1024:
 648                 return IB_MTU_1024;
 649         case 2048:
 650                 return IB_MTU_2048;
 651         case 4096:
 652                 return IB_MTU_4096;
 653         }
 654 }
 655
 656 static void
 657 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 658 {
 659         int           mtu;
 660
 661         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 662         if (cmid->route.path_rec == NULL)
 663                 return;
 664
 665         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 666         LASSERT (mtu >= 0);
 667         if (mtu != 0)
 668                 cmid->route.path_rec->mtu = mtu;
 669 }
 670
 671 static int
 672 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 673 {
 674         cpumask_t       *mask;
 675         int             vectors;
 676         int             off;
 677         int             i;
 678         lnet_nid_t      ibp_nid;
 679
 680         vectors = conn->ibc_cmid->device->num_comp_vectors;
 681         if (vectors <= 1)
 682                 return 0;
 683
 684         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 685
 686         /* hash NID to CPU id in this partition... */
 687         ibp_nid = conn->ibc_peer->ibp_nid;
 688         off = do_div(ibp_nid, cpumask_weight(mask));
 689         for_each_cpu(i, mask) {
 690                 if (off-- == 0)
 691                         return i % vectors;
 692         }
 693
 694         LBUG();
 695         return 1;
 696 }
 697
 698 kib_conn_t *
 699 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 700                    int state, int version)
 701 {
 702         /* CAVEAT EMPTOR:
 703          * If the new conn is created successfully it takes over the caller's
 704          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 705          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 706          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 707          * to destroy 'cmid' here since I'm called from the CM which still has
 708          * its ref on 'cmid'). */
 709         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 710         kib_net_t              *net = peer->ibp_ni->ni_data;
 711         kib_dev_t              *dev;
 712         struct ib_qp_init_attr *init_qp_attr;
 713         struct kib_sched_info   *sched;
 714 #ifdef HAVE_IB_CQ_INIT_ATTR
 715         struct ib_cq_init_attr  cq_attr = {};
 716 #endif
 717         kib_conn_t              *conn;
 718         struct ib_cq            *cq;
 719         unsigned long           flags;
 720         int                     cpt;
 721         int                     rc;
 722         int                     i;
 723
 724         LASSERT(net != NULL);
 725         LASSERT(!in_interrupt());
 726
 727         dev = net->ibn_dev;
 728
 729         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 730         sched = kiblnd_data.kib_scheds[cpt];
 731
 732         LASSERT(sched->ibs_nthreads > 0);
 733
 734         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 735                          sizeof(*init_qp_attr));
 736         if (init_qp_attr == NULL) {
 737                 CERROR("Can't allocate qp_attr for %s\n",
 738                        libcfs_nid2str(peer->ibp_nid));
 739                 goto failed_0;
 740         }
 741
 742         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 743         if (conn == NULL) {
 744                 CERROR("Can't allocate connection for %s\n",
 745                        libcfs_nid2str(peer->ibp_nid));
 746                 goto failed_1;
 747         }
 748
 749         conn->ibc_state = IBLND_CONN_INIT;
 750         conn->ibc_version = version;
 751         conn->ibc_peer = peer;                  /* I take the caller's ref */
 752         cmid->context = conn;                   /* for future CM callbacks */
 753         conn->ibc_cmid = cmid;
 754         conn->ibc_max_frags = peer->ibp_max_frags;
 755         conn->ibc_queue_depth = peer->ibp_queue_depth;
 756
 757         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 758         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 759         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 760         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 761         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 762         INIT_LIST_HEAD(&conn->ibc_active_txs);
 763         spin_lock_init(&conn->ibc_lock);
 764
 765         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 766                          sizeof(*conn->ibc_connvars));
 767         if (conn->ibc_connvars == NULL) {
 768                 CERROR("Can't allocate in-progress connection state\n");
 769                 goto failed_2;
 770         }
 771
 772         write_lock_irqsave(glock, flags);
 773         if (dev->ibd_failover) {
 774                 write_unlock_irqrestore(glock, flags);
 775                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 776                 goto failed_2;
 777         }
 778
 779         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 780                 /* wakeup failover thread and teardown connection */
 781                 if (kiblnd_dev_can_failover(dev)) {
 782                         list_add_tail(&dev->ibd_fail_list,
 783                                       &kiblnd_data.kib_failed_devs);
 784                         wake_up(&kiblnd_data.kib_failover_waitq);
 785                 }
 786
 787                 write_unlock_irqrestore(glock, flags);
 788                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 789                        cmid->device->name, dev->ibd_ifname);
 790                 goto failed_2;
 791         }
 792
 793         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 794         conn->ibc_hdev = dev->ibd_hdev;
 795
 796         kiblnd_setup_mtu_locked(cmid);
 797
 798         write_unlock_irqrestore(glock, flags);
 799
 800         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 801                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 802         if (conn->ibc_rxs == NULL) {
 803                 CERROR("Cannot allocate RX buffers\n");
 804                 goto failed_2;
 805         }
 806
 807         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 808                                 IBLND_RX_MSG_PAGES(conn));
 809         if (rc != 0)
 810                 goto failed_2;
 811
 812         kiblnd_map_rx_descs(conn);
 813
 814 #ifdef HAVE_IB_CQ_INIT_ATTR
 815         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 816         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 817         cq = ib_create_cq(cmid->device,
 818                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 819                           &cq_attr);
 820 #else
 821         cq = ib_create_cq(cmid->device,
 822                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 823                           IBLND_CQ_ENTRIES(conn),
 824                           kiblnd_get_completion_vector(conn, cpt));
 825 #endif
 826         if (IS_ERR(cq)) {
 827                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 828                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 829                 goto failed_2;
 830         }
 831
 832         conn->ibc_cq = cq;
 833
 834         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 835         if (rc != 0) {
 836                 CERROR("Can't request completion notification: %d\n", rc);
 837                 goto failed_2;
 838         }
 839
 840         init_qp_attr->event_handler = kiblnd_qp_event;
 841         init_qp_attr->qp_context = conn;
 842         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 843         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 844         init_qp_attr->cap.max_send_sge = 1;
 845         init_qp_attr->cap.max_recv_sge = 1;
 846         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 847         init_qp_attr->qp_type = IB_QPT_RC;
 848         init_qp_attr->send_cq = cq;
 849         init_qp_attr->recv_cq = cq;
 850
 851         conn->ibc_sched = sched;
 852
 853         do {
 854                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 855                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 856                         break;
 857
 858                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 859         } while (rc);
 860
 861         if (rc) {
 862                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 863                         rc, init_qp_attr->cap.max_send_wr,
 864                         init_qp_attr->cap.max_recv_wr);
 865                 goto failed_2;
 866         }
 867
 868         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 869                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 870                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 871
 872         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 873
 874         /* 1 ref for caller and each rxmsg */
 875         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 876         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 877
 878         /* post receives */
 879         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 880                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 881                 if (rc != 0) {
 882                         CERROR("Can't post rxmsg: %d\n", rc);
 883
 884                         /* Make posted receives complete */
 885                         kiblnd_abort_receives(conn);
 886
 887                         /* correct # of posted buffers
 888                          * NB locking needed now I'm racing with completion */
 889                         spin_lock_irqsave(&sched->ibs_lock, flags);
 890                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 891                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 892
 893                         /* cmid will be destroyed by CM(ofed) after cm_callback
 894                          * returned, so we can't refer it anymore
 895                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 896                         rdma_destroy_qp(conn->ibc_cmid);
 897                         conn->ibc_cmid = NULL;
 898
 899                         /* Drop my own and unused rxbuffer refcounts */
 900                         while (i++ <= IBLND_RX_MSGS(conn))
 901                                 kiblnd_conn_decref(conn);
 902
 903                         return NULL;
 904                 }
 905         }
 906
 907         /* Init successful! */
 908         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 909                  state == IBLND_CONN_PASSIVE_WAIT);
 910         conn->ibc_state = state;
 911
 912         /* 1 more conn */
 913         atomic_inc(&net->ibn_nconns);
 914         return conn;
 915
 916  failed_2:
 917         kiblnd_destroy_conn(conn, true);
 918  failed_1:
 919         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 920  failed_0:
 921         return NULL;
 922 }
 923
 924 void
 925 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 926 {
 927         struct rdma_cm_id *cmid = conn->ibc_cmid;
 928         kib_peer_t        *peer = conn->ibc_peer;
 929         int                rc;
 930
 931         LASSERT (!in_interrupt());
 932         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 933         LASSERT(list_empty(&conn->ibc_early_rxs));
 934         LASSERT(list_empty(&conn->ibc_tx_noops));
 935         LASSERT(list_empty(&conn->ibc_tx_queue));
 936         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 937         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 938         LASSERT(list_empty(&conn->ibc_active_txs));
 939         LASSERT (conn->ibc_noops_posted == 0);
 940         LASSERT (conn->ibc_nsends_posted == 0);
 941
 942         switch (conn->ibc_state) {
 943         default:
 944                 /* conn must be completely disengaged from the network */
 945                 LBUG();
 946
 947         case IBLND_CONN_DISCONNECTED:
 948                 /* connvars should have been freed already */
 949                 LASSERT (conn->ibc_connvars == NULL);
 950                 break;
 951
 952         case IBLND_CONN_INIT:
 953                 break;
 954         }
 955
 956         /* conn->ibc_cmid might be destroyed by CM already */
 957         if (cmid != NULL && cmid->qp != NULL)
 958                 rdma_destroy_qp(cmid);
 959
 960         if (conn->ibc_cq != NULL) {
 961                 rc = ib_destroy_cq(conn->ibc_cq);
 962                 if (rc != 0)
 963                         CWARN("Error destroying CQ: %d\n", rc);
 964         }
 965
 966         if (conn->ibc_rx_pages != NULL)
 967                 kiblnd_unmap_rx_descs(conn);
 968
 969         if (conn->ibc_rxs != NULL) {
 970                 LIBCFS_FREE(conn->ibc_rxs,
 971                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 972         }
 973
 974         if (conn->ibc_connvars != NULL)
 975                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 976
 977         if (conn->ibc_hdev != NULL)
 978                 kiblnd_hdev_decref(conn->ibc_hdev);
 979
 980         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 981         if (conn->ibc_state != IBLND_CONN_INIT) {
 982                 kib_net_t *net = peer->ibp_ni->ni_data;
 983
 984                 kiblnd_peer_decref(peer);
 985                 rdma_destroy_id(cmid);
 986                 atomic_dec(&net->ibn_nconns);
 987         }
 988
 989         if (free_conn)
 990                 LIBCFS_FREE(conn, sizeof(*conn));
 991 }
 992
 993 int
 994 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 995 {
 996         kib_conn_t              *conn;
 997         struct list_head        *ctmp;
 998         struct list_head        *cnxt;
 999         int                     count = 0;
1000
1001         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1002                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1003
1004                 CDEBUG(D_NET, "Closing conn -> %s, "
1005                               "version: %x, reason: %d\n",
1006                        libcfs_nid2str(peer->ibp_nid),
1007                        conn->ibc_version, why);
1008
1009                 kiblnd_close_conn_locked(conn, why);
1010                 count++;
1011         }
1012
1013         return count;
1014 }
1015
1016 int
1017 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1018                                 int version, __u64 incarnation)
1019 {
1020         kib_conn_t              *conn;
1021         struct list_head        *ctmp;
1022         struct list_head        *cnxt;
1023         int                     count = 0;
1024
1025         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1026                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1027
1028                 if (conn->ibc_version     == version &&
1029                     conn->ibc_incarnation == incarnation)
1030                         continue;
1031
1032                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1033                               "incarnation:"LPX64"(%x, "LPX64")\n",
1034                        libcfs_nid2str(peer->ibp_nid),
1035                        conn->ibc_version, conn->ibc_incarnation,
1036                        version, incarnation);
1037
1038                 kiblnd_close_conn_locked(conn, -ESTALE);
1039                 count++;
1040         }
1041
1042         return count;
1043 }
1044
1045 static int
1046 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1047 {
1048         kib_peer_t              *peer;
1049         struct list_head        *ptmp;
1050         struct list_head        *pnxt;
1051         int                     lo;
1052         int                     hi;
1053         int                     i;
1054         unsigned long           flags;
1055         int                     count = 0;
1056
1057         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1058
1059         if (nid != LNET_NID_ANY)
1060                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1061         else {
1062                 lo = 0;
1063                 hi = kiblnd_data.kib_peer_hash_size - 1;
1064         }
1065
1066         for (i = lo; i <= hi; i++) {
1067                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1068
1069                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1070                         LASSERT(!kiblnd_peer_idle(peer));
1071
1072                         if (peer->ibp_ni != ni)
1073                                 continue;
1074
1075                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1076                                 continue;
1077
1078                         count += kiblnd_close_peer_conns_locked(peer, 0);
1079                 }
1080         }
1081
1082         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1083
1084         /* wildcards always succeed */
1085         if (nid == LNET_NID_ANY)
1086                 return 0;
1087
1088         return (count == 0) ? -ENOENT : 0;
1089 }
1090
1091 static int
1092 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1093 {
1094         struct libcfs_ioctl_data *data = arg;
1095         int                       rc = -EINVAL;
1096
1097         switch(cmd) {
1098         case IOC_LIBCFS_GET_PEER: {
1099                 lnet_nid_t   nid = 0;
1100                 int          count = 0;
1101
1102                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1103                                           &nid, &count);
1104                 data->ioc_nid    = nid;
1105                 data->ioc_count  = count;
1106                 break;
1107         }
1108
1109         case IOC_LIBCFS_DEL_PEER: {
1110                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1111                 break;
1112         }
1113         case IOC_LIBCFS_GET_CONN: {
1114                 kib_conn_t *conn;
1115
1116                 rc = 0;
1117                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1118                 if (conn == NULL) {
1119                         rc = -ENOENT;
1120                         break;
1121                 }
1122
1123                 LASSERT (conn->ibc_cmid != NULL);
1124                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1125                 if (conn->ibc_cmid->route.path_rec == NULL)
1126                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1127                 else
1128                         data->ioc_u32[0] =
1129                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1130                 kiblnd_conn_decref(conn);
1131                 break;
1132         }
1133         case IOC_LIBCFS_CLOSE_CONNECTION: {
1134                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1135                 break;
1136         }
1137
1138         default:
1139                 break;
1140         }
1141
1142         return rc;
1143 }
1144
1145 static void
1146 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1147 {
1148         cfs_time_t      last_alive = 0;
1149         cfs_time_t      now = cfs_time_current();
1150         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1151         kib_peer_t      *peer;
1152         unsigned long   flags;
1153
1154         read_lock_irqsave(glock, flags);
1155
1156         peer = kiblnd_find_peer_locked(nid);
1157         if (peer != NULL)
1158                 last_alive = peer->ibp_last_alive;
1159
1160         read_unlock_irqrestore(glock, flags);
1161
1162         if (last_alive != 0)
1163                 *when = last_alive;
1164
1165         /* peer is not persistent in hash, trigger peer creation
1166          * and connection establishment with a NULL tx */
1167         if (peer == NULL)
1168                 kiblnd_launch_tx(ni, NULL, nid);
1169
1170         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1171                libcfs_nid2str(nid), peer,
1172                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1173         return;
1174 }
1175
1176 static void
1177 kiblnd_free_pages(kib_pages_t *p)
1178 {
1179         int     npages = p->ibp_npages;
1180         int     i;
1181
1182         for (i = 0; i < npages; i++) {
1183                 if (p->ibp_pages[i] != NULL)
1184                         __free_page(p->ibp_pages[i]);
1185         }
1186
1187         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1188 }
1189
1190 int
1191 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1192 {
1193         kib_pages_t     *p;
1194         int             i;
1195
1196         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1197                          offsetof(kib_pages_t, ibp_pages[npages]));
1198         if (p == NULL) {
1199                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1200                 return -ENOMEM;
1201         }
1202
1203         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1204         p->ibp_npages = npages;
1205
1206         for (i = 0; i < npages; i++) {
1207                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1208                                                      GFP_NOFS);
1209                 if (p->ibp_pages[i] == NULL) {
1210                         CERROR("Can't allocate page %d of %d\n", i, npages);
1211                         kiblnd_free_pages(p);
1212                         return -ENOMEM;
1213                 }
1214         }
1215
1216         *pp = p;
1217         return 0;
1218 }
1219
1220 void
1221 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1222 {
1223         kib_rx_t *rx;
1224         int       i;
1225
1226         LASSERT (conn->ibc_rxs != NULL);
1227         LASSERT (conn->ibc_hdev != NULL);
1228
1229         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1230                 rx = &conn->ibc_rxs[i];
1231
1232                 LASSERT(rx->rx_nob >= 0); /* not posted */
1233
1234                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1235                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1236                                                           rx->rx_msgaddr),
1237                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1238         }
1239
1240         kiblnd_free_pages(conn->ibc_rx_pages);
1241
1242         conn->ibc_rx_pages = NULL;
1243 }
1244
1245 void
1246 kiblnd_map_rx_descs(kib_conn_t *conn)
1247 {
1248         kib_rx_t       *rx;
1249         struct page    *pg;
1250         int             pg_off;
1251         int             ipg;
1252         int             i;
1253
1254         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1255                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1256                 rx = &conn->ibc_rxs[i];
1257
1258                 rx->rx_conn = conn;
1259                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1260
1261                 rx->rx_msgaddr =
1262                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1263                                               rx->rx_msg, IBLND_MSG_SIZE,
1264                                               DMA_FROM_DEVICE);
1265                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1266                                                   rx->rx_msgaddr));
1267                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1268
1269                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1270                        i, rx->rx_msg, rx->rx_msgaddr,
1271                        (__u64)(page_to_phys(pg) + pg_off));
1272
1273                 pg_off += IBLND_MSG_SIZE;
1274                 LASSERT(pg_off <= PAGE_SIZE);
1275
1276                 if (pg_off == PAGE_SIZE) {
1277                         pg_off = 0;
1278                         ipg++;
1279                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1280                 }
1281         }
1282 }
1283
1284 static void
1285 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1286 {
1287         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1288         kib_tx_t       *tx;
1289         int             i;
1290
1291         LASSERT (tpo->tpo_pool.po_allocated == 0);
1292
1293         if (hdev == NULL)
1294                 return;
1295
1296         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1297                 tx = &tpo->tpo_tx_descs[i];
1298                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1299                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1300                                                           tx->tx_msgaddr),
1301                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1302         }
1303
1304         kiblnd_hdev_decref(hdev);
1305         tpo->tpo_hdev = NULL;
1306 }
1307
1308 static kib_hca_dev_t *
1309 kiblnd_current_hdev(kib_dev_t *dev)
1310 {
1311         kib_hca_dev_t *hdev;
1312         unsigned long  flags;
1313         int            i = 0;
1314
1315         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1316         while (dev->ibd_failover) {
1317                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1318                 if (i++ % 50 == 0)
1319                         CDEBUG(D_NET, "%s: Wait for failover\n",
1320                                dev->ibd_ifname);
1321                 set_current_state(TASK_INTERRUPTIBLE);
1322                 schedule_timeout(cfs_time_seconds(1) / 100);
1323
1324                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1325         }
1326
1327         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1328         hdev = dev->ibd_hdev;
1329
1330         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1331
1332         return hdev;
1333 }
1334
1335 static void
1336 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1337 {
1338         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1339         kib_pool_t     *pool  = &tpo->tpo_pool;
1340         kib_net_t      *net   = pool->po_owner->ps_net;
1341         kib_dev_t      *dev;
1342         struct page    *page;
1343         kib_tx_t       *tx;
1344         int             page_offset;
1345         int             ipage;
1346         int             i;
1347
1348         LASSERT (net != NULL);
1349
1350         dev = net->ibn_dev;
1351
1352         /* pre-mapped messages are not bigger than 1 page */
1353         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1354
1355         /* No fancy arithmetic when we do the buffer calculations */
1356         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1357
1358         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1359
1360         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1361                 page = txpgs->ibp_pages[ipage];
1362                 tx = &tpo->tpo_tx_descs[i];
1363
1364                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1365                                            page_offset);
1366
1367                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1368                                                        tx->tx_msg,
1369                                                        IBLND_MSG_SIZE,
1370                                                        DMA_TO_DEVICE);
1371                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1372                                                   tx->tx_msgaddr));
1373                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1374
1375                 list_add(&tx->tx_list, &pool->po_free_list);
1376
1377                 page_offset += IBLND_MSG_SIZE;
1378                 LASSERT(page_offset <= PAGE_SIZE);
1379
1380                 if (page_offset == PAGE_SIZE) {
1381                         page_offset = 0;
1382                         ipage++;
1383                         LASSERT(ipage <= txpgs->ibp_npages);
1384                 }
1385         }
1386 }
1387
1388 struct ib_mr *
1389 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1390                       int negotiated_nfrags)
1391 {
1392         __u16   nfrags = (negotiated_nfrags != -1) ?
1393           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1394
1395         LASSERT(hdev->ibh_mrs != NULL);
1396
1397         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1398             nfrags <= rd->rd_nfrags)
1399                 return NULL;
1400
1401         return hdev->ibh_mrs;
1402 }
1403
1404 static void
1405 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1406 {
1407         LASSERT(fpo->fpo_map_count == 0);
1408
1409         if (fpo->fpo_is_fmr) {
1410                 if (fpo->fmr.fpo_fmr_pool)
1411                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1412         } else {
1413                 struct kib_fast_reg_descriptor *frd, *tmp;
1414                 int i = 0;
1415
1416                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1417                                          frd_list) {
1418                         list_del(&frd->frd_list);
1419                         ib_free_fast_reg_page_list(frd->frd_frpl);
1420                         ib_dereg_mr(frd->frd_mr);
1421                         LIBCFS_FREE(frd, sizeof(*frd));
1422                         i++;
1423                 }
1424                 if (i < fpo->fast_reg.fpo_pool_size)
1425                         CERROR("FastReg pool still has %d regions registered\n",
1426                                 fpo->fast_reg.fpo_pool_size - i);
1427         }
1428
1429         if (fpo->fpo_hdev)
1430                 kiblnd_hdev_decref(fpo->fpo_hdev);
1431
1432         LIBCFS_FREE(fpo, sizeof(*fpo));
1433 }
1434
1435 static void
1436 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1437 {
1438         kib_fmr_pool_t *fpo, *tmp;
1439
1440         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1441                 list_del(&fpo->fpo_list);
1442                 kiblnd_destroy_fmr_pool(fpo);
1443         }
1444 }
1445
1446 static int kiblnd_fmr_pool_size(int ncpts)
1447 {
1448         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1449
1450         return max(IBLND_FMR_POOL, size);
1451 }
1452
1453 static int kiblnd_fmr_flush_trigger(int ncpts)
1454 {
1455         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1456
1457         return max(IBLND_FMR_POOL_FLUSH, size);
1458 }
1459
1460 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1461 {
1462         struct ib_fmr_pool_param param = {
1463                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1464                 .page_shift        = PAGE_SHIFT,
1465                 .access            = (IB_ACCESS_LOCAL_WRITE |
1466                                       IB_ACCESS_REMOTE_WRITE),
1467                 .pool_size         = fps->fps_pool_size,
1468                 .dirty_watermark   = fps->fps_flush_trigger,
1469                 .flush_function    = NULL,
1470                 .flush_arg         = NULL,
1471                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1472         int rc = 0;
1473
1474         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1475                                                    &param);
1476         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1477                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1478                 if (rc != -ENOSYS)
1479                         CERROR("Failed to create FMR pool: %d\n", rc);
1480                 else
1481                         CERROR("FMRs are not supported\n");
1482         }
1483
1484         return rc;
1485 }
1486
1487 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1488 {
1489         struct kib_fast_reg_descriptor *frd, *tmp;
1490         int i, rc;
1491
1492         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1493         fpo->fast_reg.fpo_pool_size = 0;
1494         for (i = 0; i < fps->fps_pool_size; i++) {
1495                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1496                                  sizeof(*frd));
1497                 if (!frd) {
1498                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1499                         rc = -ENOMEM;
1500                         goto out;
1501                 }
1502                 frd->frd_mr = NULL;
1503
1504                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1505                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1506                 if (IS_ERR(frd->frd_frpl)) {
1507                         rc = PTR_ERR(frd->frd_frpl);
1508                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1509                                 rc);
1510                         goto out_middle;
1511                 }
1512
1513                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1514                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1515                 if (IS_ERR(frd->frd_mr)) {
1516                         rc = PTR_ERR(frd->frd_mr);
1517                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1518                         goto out_middle;
1519                 }
1520
1521                 frd->frd_valid = true;
1522
1523                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1524                 fpo->fast_reg.fpo_pool_size++;
1525         }
1526
1527         return 0;
1528
1529 out_middle:
1530         if (frd->frd_mr)
1531                 ib_dereg_mr(frd->frd_mr);
1532         if (frd->frd_frpl)
1533                 ib_free_fast_reg_page_list(frd->frd_frpl);
1534         LIBCFS_FREE(frd, sizeof(*frd));
1535
1536 out:
1537         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1538                                  frd_list) {
1539                 list_del(&frd->frd_list);
1540                 ib_free_fast_reg_page_list(frd->frd_frpl);
1541                 ib_dereg_mr(frd->frd_mr);
1542                 LIBCFS_FREE(frd, sizeof(*frd));
1543         }
1544
1545         return rc;
1546 }
1547
1548 static int
1549 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1550 {
1551         struct ib_device_attr *dev_attr;
1552         kib_dev_t *dev = fps->fps_net->ibn_dev;
1553         kib_fmr_pool_t *fpo;
1554         int rc;
1555
1556         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1557         if (!dev_attr)
1558                 return -ENOMEM;
1559
1560         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1561         if (!fpo) {
1562                 rc = -ENOMEM;
1563                 goto out_dev_attr;
1564         }
1565
1566         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1567
1568         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1569         if (rc) {
1570                 CERROR("Query device failed for %s: %d\n",
1571                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1572                 goto out_dev_attr;
1573         }
1574
1575         /* Check for FMR or FastReg support */
1576         fpo->fpo_is_fmr = 0;
1577         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1578             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1579             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1580             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1581                 LCONSOLE_INFO("Using FMR for registration\n");
1582                 fpo->fpo_is_fmr = 1;
1583         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1584                 LCONSOLE_INFO("Using FastReg for registration\n");
1585         } else {
1586                 rc = -ENOSYS;
1587                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1588                 goto out_dev_attr;
1589         }
1590
1591         if (fpo->fpo_is_fmr)
1592                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1593         else
1594                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1595         if (rc)
1596                 goto out_fpo;
1597
1598         kfree(dev_attr);
1599         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1600         fpo->fpo_owner    = fps;
1601         *pp_fpo = fpo;
1602
1603         return 0;
1604
1605 out_fpo:
1606         kiblnd_hdev_decref(fpo->fpo_hdev);
1607         LIBCFS_FREE(fpo, sizeof(*fpo));
1608
1609 out_dev_attr:
1610         kfree(dev_attr);
1611
1612         return rc;
1613 }
1614
1615 static void
1616 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1617 {
1618         if (fps->fps_net == NULL) /* intialized? */
1619                 return;
1620
1621         spin_lock(&fps->fps_lock);
1622
1623         while (!list_empty(&fps->fps_pool_list)) {
1624                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1625                                                  kib_fmr_pool_t, fpo_list);
1626                 fpo->fpo_failed = 1;
1627                 list_del(&fpo->fpo_list);
1628                 if (fpo->fpo_map_count == 0)
1629                         list_add(&fpo->fpo_list, zombies);
1630                 else
1631                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1632         }
1633
1634         spin_unlock(&fps->fps_lock);
1635 }
1636
1637 static void
1638 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1639 {
1640         if (fps->fps_net != NULL) { /* initialized? */
1641                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1642                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1643         }
1644 }
1645
1646 static int
1647 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1648                         int pool_size, int flush_trigger)
1649 {
1650         kib_fmr_pool_t *fpo;
1651         int             rc;
1652
1653         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1654
1655         fps->fps_net = net;
1656         fps->fps_cpt = cpt;
1657         fps->fps_pool_size = pool_size;
1658         fps->fps_flush_trigger = flush_trigger;
1659         spin_lock_init(&fps->fps_lock);
1660         INIT_LIST_HEAD(&fps->fps_pool_list);
1661         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1662
1663         rc = kiblnd_create_fmr_pool(fps, &fpo);
1664         if (rc == 0)
1665                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1666
1667         return rc;
1668 }
1669
1670 static int
1671 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1672 {
1673         if (fpo->fpo_map_count != 0) /* still in use */
1674                 return 0;
1675         if (fpo->fpo_failed)
1676                 return 1;
1677         return cfs_time_aftereq(now, fpo->fpo_deadline);
1678 }
1679
1680 void
1681 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1682 {
1683         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1684         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1685         kib_fmr_poolset_t *fps;
1686         cfs_time_t         now = cfs_time_current();
1687         kib_fmr_pool_t    *tmp;
1688         int                rc;
1689
1690         if (!fpo)
1691                 return;
1692
1693         fps = fpo->fpo_owner;
1694         if (fpo->fpo_is_fmr) {
1695                 if (fmr->fmr_pfmr) {
1696                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1697                         LASSERT(!rc);
1698                         fmr->fmr_pfmr = NULL;
1699                 }
1700
1701                 if (status) {
1702                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1703                         LASSERT(!rc);
1704                 }
1705         } else {
1706                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1707
1708                 if (frd) {
1709                         frd->frd_valid = false;
1710                         spin_lock(&fps->fps_lock);
1711                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1712                         spin_unlock(&fps->fps_lock);
1713                         fmr->fmr_frd = NULL;
1714                 }
1715         }
1716         fmr->fmr_pool = NULL;
1717
1718         spin_lock(&fps->fps_lock);
1719         fpo->fpo_map_count--;   /* decref the pool */
1720
1721         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1722                 /* the first pool is persistent */
1723                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1724                         continue;
1725
1726                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1727                         list_move(&fpo->fpo_list, &zombies);
1728                         fps->fps_version++;
1729                 }
1730         }
1731         spin_unlock(&fps->fps_lock);
1732
1733         if (!list_empty(&zombies))
1734                 kiblnd_destroy_fmr_pool_list(&zombies);
1735 }
1736
1737 int
1738 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1739                     __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1740 {
1741         kib_fmr_pool_t *fpo;
1742         __u64 version;
1743         int rc;
1744
1745 again:
1746         spin_lock(&fps->fps_lock);
1747         version = fps->fps_version;
1748         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1749                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1750                 fpo->fpo_map_count++;
1751
1752                 if (fpo->fpo_is_fmr) {
1753                         struct ib_pool_fmr *pfmr;
1754
1755                         spin_unlock(&fps->fps_lock);
1756                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1757                                                     pages, npages, iov);
1758                         if (likely(!IS_ERR(pfmr))) {
1759                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1760                                                       : pfmr->fmr->lkey;
1761                                 fmr->fmr_frd  = NULL;
1762                                 fmr->fmr_pfmr = pfmr;
1763                                 fmr->fmr_pool = fpo;
1764                                 return 0;
1765                         }
1766                         rc = PTR_ERR(pfmr);
1767                 } else {
1768                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1769                                 struct ib_send_wr *wr;
1770                                 struct kib_fast_reg_descriptor *frd;
1771                                 struct ib_fast_reg_page_list *frpl;
1772                                 struct ib_mr *mr;
1773
1774                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1775                                                         struct kib_fast_reg_descriptor,
1776                                                         frd_list);
1777                                 list_del(&frd->frd_list);
1778                                 spin_unlock(&fps->fps_lock);
1779
1780                                 frpl = frd->frd_frpl;
1781                                 mr   = frd->frd_mr;
1782
1783                                 if (!frd->frd_valid) {
1784                                         struct ib_send_wr *inv_wr;
1785                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1786
1787                                         inv_wr = &frd->frd_inv_wr;
1788                                         memset(inv_wr, 0, sizeof(*inv_wr));
1789                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1790                                         inv_wr->wr_id = IBLND_WID_MR;
1791                                         inv_wr->ex.invalidate_rkey = key;
1792
1793                                         /* Bump the key */
1794                                         key = ib_inc_rkey(key);
1795                                         ib_update_fast_reg_key(mr, key);
1796                                 }
1797
1798                                 LASSERT(npages <= frpl->max_page_list_len);
1799                                 memcpy(frpl->page_list, pages,
1800                                         sizeof(*pages) * npages);
1801
1802                                 /* Prepare FastReg WR */
1803                                 wr = &frd->frd_fastreg_wr;
1804                                 memset(wr, 0, sizeof(*wr));
1805                                 wr->opcode = IB_WR_FAST_REG_MR;
1806                                 wr->wr_id = IBLND_WID_MR;
1807                                 wr->wr.fast_reg.iova_start = iov;
1808                                 wr->wr.fast_reg.page_list  = frpl;
1809                                 wr->wr.fast_reg.page_list_len = npages;
1810                                 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1811                                 wr->wr.fast_reg.length = nob;
1812                                 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1813                                                              : mr->lkey;
1814                                 wr->wr.fast_reg.access_flags =
1815                                                 (IB_ACCESS_LOCAL_WRITE |
1816                                                  IB_ACCESS_REMOTE_WRITE);
1817
1818                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1819                                 fmr->fmr_frd  = frd;
1820                                 fmr->fmr_pfmr = NULL;
1821                                 fmr->fmr_pool = fpo;
1822                                 return 0;
1823                         }
1824                         spin_unlock(&fps->fps_lock);
1825                         rc = -EBUSY;
1826                 }
1827
1828                 spin_lock(&fps->fps_lock);
1829                 fpo->fpo_map_count--;
1830                 if (rc != -EAGAIN) {
1831                         spin_unlock(&fps->fps_lock);
1832                         return rc;
1833                 }
1834
1835                 /* EAGAIN and ... */
1836                 if (version != fps->fps_version) {
1837                         spin_unlock(&fps->fps_lock);
1838                         goto again;
1839                 }
1840         }
1841
1842         if (fps->fps_increasing) {
1843                 spin_unlock(&fps->fps_lock);
1844                 CDEBUG(D_NET, "Another thread is allocating new "
1845                        "FMR pool, waiting for her to complete\n");
1846                 schedule();
1847                 goto again;
1848
1849         }
1850
1851         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1852                 /* someone failed recently */
1853                 spin_unlock(&fps->fps_lock);
1854                 return -EAGAIN;
1855         }
1856
1857         fps->fps_increasing = 1;
1858         spin_unlock(&fps->fps_lock);
1859
1860         CDEBUG(D_NET, "Allocate new FMR pool\n");
1861         rc = kiblnd_create_fmr_pool(fps, &fpo);
1862         spin_lock(&fps->fps_lock);
1863         fps->fps_increasing = 0;
1864         if (rc == 0) {
1865                 fps->fps_version++;
1866                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1867         } else {
1868                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1869         }
1870         spin_unlock(&fps->fps_lock);
1871
1872         goto again;
1873 }
1874
1875 static void
1876 kiblnd_fini_pool(kib_pool_t *pool)
1877 {
1878         LASSERT(list_empty(&pool->po_free_list));
1879         LASSERT(pool->po_allocated == 0);
1880
1881         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1882 }
1883
1884 static void
1885 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1886 {
1887         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1888
1889         memset(pool, 0, sizeof(kib_pool_t));
1890         INIT_LIST_HEAD(&pool->po_free_list);
1891         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1892         pool->po_owner    = ps;
1893         pool->po_size     = size;
1894 }
1895
1896 static void
1897 kiblnd_destroy_pool_list(struct list_head *head)
1898 {
1899         kib_pool_t *pool;
1900
1901         while (!list_empty(head)) {
1902                 pool = list_entry(head->next, kib_pool_t, po_list);
1903                 list_del(&pool->po_list);
1904
1905                 LASSERT(pool->po_owner != NULL);
1906                 pool->po_owner->ps_pool_destroy(pool);
1907         }
1908 }
1909
1910 static void
1911 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1912 {
1913         if (ps->ps_net == NULL) /* intialized? */
1914                 return;
1915
1916         spin_lock(&ps->ps_lock);
1917         while (!list_empty(&ps->ps_pool_list)) {
1918                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1919                                             kib_pool_t, po_list);
1920                 po->po_failed = 1;
1921                 list_del(&po->po_list);
1922                 if (po->po_allocated == 0)
1923                         list_add(&po->po_list, zombies);
1924                 else
1925                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1926         }
1927         spin_unlock(&ps->ps_lock);
1928 }
1929
1930 static void
1931 kiblnd_fini_poolset(kib_poolset_t *ps)
1932 {
1933         if (ps->ps_net != NULL) { /* initialized? */
1934                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1935                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1936         }
1937 }
1938
1939 static int
1940 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1941                     kib_net_t *net, char *name, int size,
1942                     kib_ps_pool_create_t po_create,
1943                     kib_ps_pool_destroy_t po_destroy,
1944                     kib_ps_node_init_t nd_init,
1945                     kib_ps_node_fini_t nd_fini)
1946 {
1947         kib_pool_t      *pool;
1948         int             rc;
1949
1950         memset(ps, 0, sizeof(kib_poolset_t));
1951
1952         ps->ps_cpt          = cpt;
1953         ps->ps_net          = net;
1954         ps->ps_pool_create  = po_create;
1955         ps->ps_pool_destroy = po_destroy;
1956         ps->ps_node_init    = nd_init;
1957         ps->ps_node_fini    = nd_fini;
1958         ps->ps_pool_size    = size;
1959         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1960             >= sizeof(ps->ps_name))
1961                 return -E2BIG;
1962         spin_lock_init(&ps->ps_lock);
1963         INIT_LIST_HEAD(&ps->ps_pool_list);
1964         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1965
1966         rc = ps->ps_pool_create(ps, size, &pool);
1967         if (rc == 0)
1968                 list_add(&pool->po_list, &ps->ps_pool_list);
1969         else
1970                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1971
1972         return rc;
1973 }
1974
1975 static int
1976 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1977 {
1978         if (pool->po_allocated != 0) /* still in use */
1979                 return 0;
1980         if (pool->po_failed)
1981                 return 1;
1982         return cfs_time_aftereq(now, pool->po_deadline);
1983 }
1984
1985 void
1986 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1987 {
1988         struct list_head zombies = LIST_HEAD_INIT(zombies);
1989         kib_poolset_t   *ps = pool->po_owner;
1990         kib_pool_t      *tmp;
1991         cfs_time_t       now = cfs_time_current();
1992
1993         spin_lock(&ps->ps_lock);
1994
1995         if (ps->ps_node_fini != NULL)
1996                 ps->ps_node_fini(pool, node);
1997
1998         LASSERT(pool->po_allocated > 0);
1999         list_add(node, &pool->po_free_list);
2000         pool->po_allocated--;
2001
2002         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2003                 /* the first pool is persistent */
2004                 if (ps->ps_pool_list.next == &pool->po_list)
2005                         continue;
2006
2007                 if (kiblnd_pool_is_idle(pool, now))
2008                         list_move(&pool->po_list, &zombies);
2009         }
2010         spin_unlock(&ps->ps_lock);
2011
2012         if (!list_empty(&zombies))
2013                 kiblnd_destroy_pool_list(&zombies);
2014 }
2015
2016 struct list_head *
2017 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2018 {
2019         struct list_head        *node;
2020         kib_pool_t              *pool;
2021         int                     rc;
2022         unsigned int            interval = 1;
2023         cfs_time_t              time_before;
2024         unsigned int            trips = 0;
2025
2026 again:
2027         spin_lock(&ps->ps_lock);
2028         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2029                 if (list_empty(&pool->po_free_list))
2030                         continue;
2031
2032                 pool->po_allocated++;
2033                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2034                 node = pool->po_free_list.next;
2035                 list_del(node);
2036
2037                 if (ps->ps_node_init != NULL) {
2038                         /* still hold the lock */
2039                         ps->ps_node_init(pool, node);
2040                 }
2041                 spin_unlock(&ps->ps_lock);
2042                 return node;
2043         }
2044
2045         /* no available tx pool and ... */
2046         if (ps->ps_increasing) {
2047                 /* another thread is allocating a new pool */
2048                 spin_unlock(&ps->ps_lock);
2049                 trips++;
2050                 CDEBUG(D_NET, "Another thread is allocating new "
2051                        "%s pool, waiting %d HZs for her to complete."
2052                        "trips = %d\n",
2053                        ps->ps_name, interval, trips);
2054
2055                 set_current_state(TASK_INTERRUPTIBLE);
2056                 schedule_timeout(interval);
2057                 if (interval < cfs_time_seconds(1))
2058                         interval *= 2;
2059
2060                 goto again;
2061         }
2062
2063         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2064                 /* someone failed recently */
2065                 spin_unlock(&ps->ps_lock);
2066                 return NULL;
2067         }
2068
2069         ps->ps_increasing = 1;
2070         spin_unlock(&ps->ps_lock);
2071
2072         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2073         time_before = cfs_time_current();
2074         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2075         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2076                cfs_time_current() - time_before);
2077
2078         spin_lock(&ps->ps_lock);
2079         ps->ps_increasing = 0;
2080         if (rc == 0) {
2081                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2082         } else {
2083                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2084                 CERROR("Can't allocate new %s pool because out of memory\n",
2085                        ps->ps_name);
2086         }
2087         spin_unlock(&ps->ps_lock);
2088
2089         goto again;
2090 }
2091
2092 static void
2093 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2094 {
2095         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2096         int             i;
2097
2098         LASSERT (pool->po_allocated == 0);
2099
2100         if (tpo->tpo_tx_pages != NULL) {
2101                 kiblnd_unmap_tx_pool(tpo);
2102                 kiblnd_free_pages(tpo->tpo_tx_pages);
2103         }
2104
2105         if (tpo->tpo_tx_descs == NULL)
2106                 goto out;
2107
2108         for (i = 0; i < pool->po_size; i++) {
2109                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2110
2111                 list_del(&tx->tx_list);
2112                 if (tx->tx_pages != NULL)
2113                         LIBCFS_FREE(tx->tx_pages,
2114                                     LNET_MAX_IOV *
2115                                     sizeof(*tx->tx_pages));
2116                 if (tx->tx_frags != NULL)
2117                         LIBCFS_FREE(tx->tx_frags,
2118                                     IBLND_MAX_RDMA_FRAGS *
2119                                             sizeof(*tx->tx_frags));
2120                 if (tx->tx_wrq != NULL)
2121                         LIBCFS_FREE(tx->tx_wrq,
2122                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2123                                     sizeof(*tx->tx_wrq));
2124                 if (tx->tx_sge != NULL)
2125                         LIBCFS_FREE(tx->tx_sge,
2126                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2127                                     sizeof(*tx->tx_sge));
2128                 if (tx->tx_rd != NULL)
2129                         LIBCFS_FREE(tx->tx_rd,
2130                                     offsetof(kib_rdma_desc_t,
2131                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2132         }
2133
2134         LIBCFS_FREE(tpo->tpo_tx_descs,
2135                     pool->po_size * sizeof(kib_tx_t));
2136 out:
2137         kiblnd_fini_pool(pool);
2138         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2139 }
2140
2141 static int kiblnd_tx_pool_size(int ncpts)
2142 {
2143         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2144
2145         return max(IBLND_TX_POOL, ntx);
2146 }
2147
2148 static int
2149 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2150 {
2151         int            i;
2152         int            npg;
2153         kib_pool_t    *pool;
2154         kib_tx_pool_t *tpo;
2155
2156         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2157         if (tpo == NULL) {
2158                 CERROR("Failed to allocate TX pool\n");
2159                 return -ENOMEM;
2160         }
2161
2162         pool = &tpo->tpo_pool;
2163         kiblnd_init_pool(ps, pool, size);
2164         tpo->tpo_tx_descs = NULL;
2165         tpo->tpo_tx_pages = NULL;
2166
2167         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2168         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2169                 CERROR("Can't allocate tx pages: %d\n", npg);
2170                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2171                 return -ENOMEM;
2172         }
2173
2174         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2175                          size * sizeof(kib_tx_t));
2176         if (tpo->tpo_tx_descs == NULL) {
2177                 CERROR("Can't allocate %d tx descriptors\n", size);
2178                 ps->ps_pool_destroy(pool);
2179                 return -ENOMEM;
2180         }
2181
2182         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2183
2184         for (i = 0; i < size; i++) {
2185                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2186
2187                 tx->tx_pool = tpo;
2188                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2189                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2190                                          lnet_cpt_table(), ps->ps_cpt,
2191                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2192                         if (tx->tx_pages == NULL)
2193                                 break;
2194                 }
2195
2196                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2197                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2198                 if (tx->tx_frags == NULL)
2199                         break;
2200
2201                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2202
2203                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2204                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2205                                  sizeof(*tx->tx_wrq));
2206                 if (tx->tx_wrq == NULL)
2207                         break;
2208
2209                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2210                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2211                                  sizeof(*tx->tx_sge));
2212                 if (tx->tx_sge == NULL)
2213                         break;
2214
2215                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2216                                  offsetof(kib_rdma_desc_t,
2217                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2218                 if (tx->tx_rd == NULL)
2219                         break;
2220         }
2221
2222         if (i == size) {
2223                 kiblnd_map_tx_pool(tpo);
2224                 *pp_po = pool;
2225                 return 0;
2226         }
2227
2228         ps->ps_pool_destroy(pool);
2229         return -ENOMEM;
2230 }
2231
2232 static void
2233 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2234 {
2235         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2236                                              tps_poolset);
2237         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2238
2239         tx->tx_cookie = tps->tps_next_tx_cookie++;
2240 }
2241
2242 static void
2243 kiblnd_net_fini_pools(kib_net_t *net)
2244 {
2245         int     i;
2246
2247         cfs_cpt_for_each(i, lnet_cpt_table()) {
2248                 kib_tx_poolset_t        *tps;
2249                 kib_fmr_poolset_t       *fps;
2250
2251                 if (net->ibn_tx_ps != NULL) {
2252                         tps = net->ibn_tx_ps[i];
2253                         kiblnd_fini_poolset(&tps->tps_poolset);
2254                 }
2255
2256                 if (net->ibn_fmr_ps != NULL) {
2257                         fps = net->ibn_fmr_ps[i];
2258                         kiblnd_fini_fmr_poolset(fps);
2259                 }
2260         }
2261
2262         if (net->ibn_tx_ps != NULL) {
2263                 cfs_percpt_free(net->ibn_tx_ps);
2264                 net->ibn_tx_ps = NULL;
2265         }
2266
2267         if (net->ibn_fmr_ps != NULL) {
2268                 cfs_percpt_free(net->ibn_fmr_ps);
2269                 net->ibn_fmr_ps = NULL;
2270         }
2271 }
2272
2273 static int
2274 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2275 {
2276         unsigned long   flags;
2277         int             cpt;
2278         int             rc;
2279         int             i;
2280
2281         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2282         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2283                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2284                                            flags);
2285                 goto create_tx_pool;
2286         }
2287
2288         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2289
2290         if (*kiblnd_tunables.kib_fmr_pool_size <
2291             *kiblnd_tunables.kib_ntx / 4) {
2292                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2293                        *kiblnd_tunables.kib_fmr_pool_size,
2294                        *kiblnd_tunables.kib_ntx / 4);
2295                 rc = -EINVAL;
2296                 goto failed;
2297         }
2298
2299         /* TX pool must be created later than FMR, see LU-2268
2300          * for details */
2301         LASSERT(net->ibn_tx_ps == NULL);
2302
2303         /* premapping can fail if ibd_nmr > 1, so we always create
2304          * FMR pool and map-on-demand if premapping failed */
2305
2306         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2307                                            sizeof(kib_fmr_poolset_t));
2308         if (net->ibn_fmr_ps == NULL) {
2309                 CERROR("Failed to allocate FMR pool array\n");
2310                 rc = -ENOMEM;
2311                 goto failed;
2312         }
2313
2314         for (i = 0; i < ncpts; i++) {
2315                 cpt = (cpts == NULL) ? i : cpts[i];
2316                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2317                                              kiblnd_fmr_pool_size(ncpts),
2318                                              kiblnd_fmr_flush_trigger(ncpts));
2319                 if (rc != 0) {
2320                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2321                                cpt, rc);
2322                         goto failed;
2323                 }
2324         }
2325
2326         if (i > 0)
2327                 LASSERT(i == ncpts);
2328
2329  create_tx_pool:
2330         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2331                                           sizeof(kib_tx_poolset_t));
2332         if (net->ibn_tx_ps == NULL) {
2333                 CERROR("Failed to allocate tx pool array\n");
2334                 rc = -ENOMEM;
2335                 goto failed;
2336         }
2337
2338         for (i = 0; i < ncpts; i++) {
2339                 cpt = (cpts == NULL) ? i : cpts[i];
2340                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2341                                          cpt, net, "TX",
2342                                          kiblnd_tx_pool_size(ncpts),
2343                                          kiblnd_create_tx_pool,
2344                                          kiblnd_destroy_tx_pool,
2345                                          kiblnd_tx_init, NULL);
2346                 if (rc != 0) {
2347                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2348                                cpt, rc);
2349                         goto failed;
2350                 }
2351         }
2352
2353         return 0;
2354  failed:
2355         kiblnd_net_fini_pools(net);
2356         LASSERT(rc != 0);
2357         return rc;
2358 }
2359
2360 static int
2361 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2362 {
2363         struct ib_device_attr *attr;
2364         int                    rc;
2365
2366         /* It's safe to assume a HCA can handle a page size
2367          * matching that of the native system */
2368         hdev->ibh_page_shift = PAGE_SHIFT;
2369         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2370         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2371
2372         LIBCFS_ALLOC(attr, sizeof(*attr));
2373         if (attr == NULL) {
2374                 CERROR("Out of memory\n");
2375                 return -ENOMEM;
2376         }
2377
2378         rc = ib_query_device(hdev->ibh_ibdev, attr);
2379         if (rc == 0)
2380                 hdev->ibh_mr_size = attr->max_mr_size;
2381
2382         LIBCFS_FREE(attr, sizeof(*attr));
2383
2384         if (rc != 0) {
2385                 CERROR("Failed to query IB device: %d\n", rc);
2386                 return rc;
2387         }
2388
2389         if (hdev->ibh_mr_size == ~0ULL) {
2390                 hdev->ibh_mr_shift = 64;
2391                 return 0;
2392         }
2393
2394         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2395         return -EINVAL;
2396 }
2397
2398 static void
2399 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2400 {
2401         if (hdev->ibh_mrs == NULL)
2402                 return;
2403
2404         ib_dereg_mr(hdev->ibh_mrs);
2405
2406         hdev->ibh_mrs = NULL;
2407 }
2408
2409 void
2410 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2411 {
2412         kiblnd_hdev_cleanup_mrs(hdev);
2413
2414         if (hdev->ibh_pd != NULL)
2415                 ib_dealloc_pd(hdev->ibh_pd);
2416
2417         if (hdev->ibh_cmid != NULL)
2418                 rdma_destroy_id(hdev->ibh_cmid);
2419
2420         LIBCFS_FREE(hdev, sizeof(*hdev));
2421 }
2422
2423 static int
2424 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2425 {
2426         struct ib_mr *mr;
2427         int           rc;
2428         int           acflags = IB_ACCESS_LOCAL_WRITE |
2429                                 IB_ACCESS_REMOTE_WRITE;
2430
2431         rc = kiblnd_hdev_get_attr(hdev);
2432         if (rc != 0)
2433                 return rc;
2434
2435         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2436         if (IS_ERR(mr)) {
2437                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2438                 kiblnd_hdev_cleanup_mrs(hdev);
2439                 return PTR_ERR(mr);
2440         }
2441
2442         hdev->ibh_mrs = mr;
2443
2444         return 0;
2445 }
2446
2447 static int
2448 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2449 {       /* DUMMY */
2450         return 0;
2451 }
2452
2453 static int
2454 kiblnd_dev_need_failover(kib_dev_t *dev)
2455 {
2456         struct rdma_cm_id  *cmid;
2457         struct sockaddr_in  srcaddr;
2458         struct sockaddr_in  dstaddr;
2459         int                 rc;
2460
2461         if (dev->ibd_hdev == NULL || /* initializing */
2462             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2463             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2464                 return 1;
2465
2466         /* XXX: it's UGLY, but I don't have better way to find
2467          * ib-bonding HCA failover because:
2468          *
2469          * a. no reliable CM event for HCA failover...
2470          * b. no OFED API to get ib_device for current net_device...
2471          *
2472          * We have only two choices at this point:
2473          *
2474          * a. rdma_bind_addr(), it will conflict with listener cmid
2475          * b. rdma_resolve_addr() to zero addr */
2476         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2477                                      IB_QPT_RC);
2478         if (IS_ERR(cmid)) {
2479                 rc = PTR_ERR(cmid);
2480                 CERROR("Failed to create cmid for failover: %d\n", rc);
2481                 return rc;
2482         }
2483
2484         memset(&srcaddr, 0, sizeof(srcaddr));
2485         srcaddr.sin_family      = AF_INET;
2486         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2487
2488         memset(&dstaddr, 0, sizeof(dstaddr));
2489         dstaddr.sin_family = AF_INET;
2490         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2491                                (struct sockaddr *)&dstaddr, 1);
2492         if (rc != 0 || cmid->device == NULL) {
2493                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2494                        dev->ibd_ifname, &dev->ibd_ifip,
2495                        cmid->device, rc);
2496                 rdma_destroy_id(cmid);
2497                 return rc;
2498         }
2499
2500         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2501         rdma_destroy_id(cmid);
2502         return rc;
2503 }
2504
2505 int
2506 kiblnd_dev_failover(kib_dev_t *dev)
2507 {
2508         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2509         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2510         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2511         struct rdma_cm_id  *cmid  = NULL;
2512         kib_hca_dev_t      *hdev  = NULL;
2513         kib_hca_dev_t      *old;
2514         struct ib_pd       *pd;
2515         kib_net_t          *net;
2516         struct sockaddr_in  addr;
2517         unsigned long       flags;
2518         int                 rc = 0;
2519         int                 i;
2520
2521         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2522                  dev->ibd_can_failover ||
2523                  dev->ibd_hdev == NULL);
2524
2525         rc = kiblnd_dev_need_failover(dev);
2526         if (rc <= 0)
2527                 goto out;
2528
2529         if (dev->ibd_hdev != NULL &&
2530             dev->ibd_hdev->ibh_cmid != NULL) {
2531                 /* XXX it's not good to close old listener at here,
2532                  * because we can fail to create new listener.
2533                  * But we have to close it now, otherwise rdma_bind_addr
2534                  * will return EADDRINUSE... How crap! */
2535                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2536
2537                 cmid = dev->ibd_hdev->ibh_cmid;
2538                 /* make next schedule of kiblnd_dev_need_failover()
2539                  * return 1 for me */
2540                 dev->ibd_hdev->ibh_cmid  = NULL;
2541                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2542
2543                 rdma_destroy_id(cmid);
2544         }
2545
2546         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2547                                      IB_QPT_RC);
2548         if (IS_ERR(cmid)) {
2549                 rc = PTR_ERR(cmid);
2550                 CERROR("Failed to create cmid for failover: %d\n", rc);
2551                 goto out;
2552         }
2553
2554         memset(&addr, 0, sizeof(addr));
2555         addr.sin_family      = AF_INET;
2556         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2557         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2558
2559         /* Bind to failover device or port */
2560         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2561         if (rc != 0 || cmid->device == NULL) {
2562                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2563                        dev->ibd_ifname, &dev->ibd_ifip,
2564                        cmid->device, rc);
2565                 rdma_destroy_id(cmid);
2566                 goto out;
2567         }
2568
2569         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2570         if (hdev == NULL) {
2571                 CERROR("Failed to allocate kib_hca_dev\n");
2572                 rdma_destroy_id(cmid);
2573                 rc = -ENOMEM;
2574                 goto out;
2575         }
2576
2577         atomic_set(&hdev->ibh_ref, 1);
2578         hdev->ibh_dev   = dev;
2579         hdev->ibh_cmid  = cmid;
2580         hdev->ibh_ibdev = cmid->device;
2581
2582         pd = ib_alloc_pd(cmid->device);
2583         if (IS_ERR(pd)) {
2584                 rc = PTR_ERR(pd);
2585                 CERROR("Can't allocate PD: %d\n", rc);
2586                 goto out;
2587         }
2588
2589         hdev->ibh_pd = pd;
2590
2591         rc = rdma_listen(cmid, 0);
2592         if (rc != 0) {
2593                 CERROR("Can't start new listener: %d\n", rc);
2594                 goto out;
2595         }
2596
2597         rc = kiblnd_hdev_setup_mrs(hdev);
2598         if (rc != 0) {
2599                 CERROR("Can't setup device: %d\n", rc);
2600                 goto out;
2601         }
2602
2603         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2604
2605         old = dev->ibd_hdev;
2606         dev->ibd_hdev = hdev;   /* take over the refcount */
2607         hdev = old;
2608
2609         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2610                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2611                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2612                                             &zombie_tpo);
2613
2614                         if (net->ibn_fmr_ps != NULL)
2615                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2616                                                         &zombie_fpo);
2617                 }
2618         }
2619
2620         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2621  out:
2622         if (!list_empty(&zombie_tpo))
2623                 kiblnd_destroy_pool_list(&zombie_tpo);
2624         if (!list_empty(&zombie_ppo))
2625                 kiblnd_destroy_pool_list(&zombie_ppo);
2626         if (!list_empty(&zombie_fpo))
2627                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2628         if (hdev != NULL)
2629                 kiblnd_hdev_decref(hdev);
2630
2631         if (rc != 0)
2632                 dev->ibd_failed_failover++;
2633         else
2634                 dev->ibd_failed_failover = 0;
2635
2636         return rc;
2637 }
2638
2639 void
2640 kiblnd_destroy_dev (kib_dev_t *dev)
2641 {
2642         LASSERT (dev->ibd_nnets == 0);
2643         LASSERT(list_empty(&dev->ibd_nets));
2644
2645         list_del(&dev->ibd_fail_list);
2646         list_del(&dev->ibd_list);
2647
2648         if (dev->ibd_hdev != NULL)
2649                 kiblnd_hdev_decref(dev->ibd_hdev);
2650
2651         LIBCFS_FREE(dev, sizeof(*dev));
2652 }
2653
2654 static kib_dev_t *
2655 kiblnd_create_dev(char *ifname)
2656 {
2657         struct net_device *netdev;
2658         kib_dev_t         *dev;
2659         __u32              netmask;
2660         __u32              ip;
2661         int                up;
2662         int                rc;
2663
2664         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2665         if (rc != 0) {
2666                 CERROR("Can't query IPoIB interface %s: %d\n",
2667                        ifname, rc);
2668                 return NULL;
2669         }
2670
2671         if (!up) {
2672                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2673                 return NULL;
2674         }
2675
2676         LIBCFS_ALLOC(dev, sizeof(*dev));
2677         if (dev == NULL)
2678                 return NULL;
2679
2680         netdev = dev_get_by_name(&init_net, ifname);
2681         if (netdev == NULL) {
2682                 dev->ibd_can_failover = 0;
2683         } else {
2684                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2685                 dev_put(netdev);
2686         }
2687
2688         INIT_LIST_HEAD(&dev->ibd_nets);
2689         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2690         INIT_LIST_HEAD(&dev->ibd_fail_list);
2691         dev->ibd_ifip = ip;
2692         strcpy(&dev->ibd_ifname[0], ifname);
2693
2694         /* initialize the device */
2695         rc = kiblnd_dev_failover(dev);
2696         if (rc != 0) {
2697                 CERROR("Can't initialize device: %d\n", rc);
2698                 LIBCFS_FREE(dev, sizeof(*dev));
2699                 return NULL;
2700         }
2701
2702         list_add_tail(&dev->ibd_list,
2703                           &kiblnd_data.kib_devs);
2704         return dev;
2705 }
2706
2707 static void
2708 kiblnd_base_shutdown(void)
2709 {
2710         struct kib_sched_info   *sched;
2711         int                     i;
2712
2713         LASSERT(list_empty(&kiblnd_data.kib_devs));
2714
2715         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2716                atomic_read(&libcfs_kmemory));
2717
2718         switch (kiblnd_data.kib_init) {
2719         default:
2720                 LBUG();
2721
2722         case IBLND_INIT_ALL:
2723         case IBLND_INIT_DATA:
2724                 LASSERT (kiblnd_data.kib_peers != NULL);
2725                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2726                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2727                 }
2728                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2729                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2730                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2731                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2732
2733                 /* flag threads to terminate; wake and wait for them to die */
2734                 kiblnd_data.kib_shutdown = 1;
2735
2736                 /* NB: we really want to stop scheduler threads net by net
2737                  * instead of the whole module, this should be improved
2738                  * with dynamic configuration LNet */
2739                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2740                         wake_up_all(&sched->ibs_waitq);
2741
2742                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2743                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2744
2745                 i = 2;
2746                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2747                         i++;
2748                         /* power of 2? */
2749                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2750                                "Waiting for %d threads to terminate\n",
2751                                atomic_read(&kiblnd_data.kib_nthreads));
2752                         set_current_state(TASK_UNINTERRUPTIBLE);
2753                         schedule_timeout(cfs_time_seconds(1));
2754                 }
2755
2756                 /* fall through */
2757
2758         case IBLND_INIT_NOTHING:
2759                 break;
2760         }
2761
2762         if (kiblnd_data.kib_peers != NULL) {
2763                 LIBCFS_FREE(kiblnd_data.kib_peers,
2764                             sizeof(struct list_head) *
2765                             kiblnd_data.kib_peer_hash_size);
2766         }
2767
2768         if (kiblnd_data.kib_scheds != NULL)
2769                 cfs_percpt_free(kiblnd_data.kib_scheds);
2770
2771         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2772                atomic_read(&libcfs_kmemory));
2773
2774         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2775         module_put(THIS_MODULE);
2776 }
2777
2778 static void
2779 kiblnd_shutdown (lnet_ni_t *ni)
2780 {
2781         kib_net_t        *net = ni->ni_data;
2782         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2783         int               i;
2784         unsigned long     flags;
2785
2786         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2787
2788         if (net == NULL)
2789                 goto out;
2790
2791         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2792                atomic_read(&libcfs_kmemory));
2793
2794         write_lock_irqsave(g_lock, flags);
2795         net->ibn_shutdown = 1;
2796         write_unlock_irqrestore(g_lock, flags);
2797
2798         switch (net->ibn_init) {
2799         default:
2800                 LBUG();
2801
2802         case IBLND_INIT_ALL:
2803                 /* nuke all existing peers within this net */
2804                 kiblnd_del_peer(ni, LNET_NID_ANY);
2805
2806                 /* Wait for all peer state to clean up */
2807                 i = 2;
2808                 while (atomic_read(&net->ibn_npeers) != 0) {
2809                         i++;
2810                         /* power of 2? */
2811                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2812                                "%s: waiting for %d peers to disconnect\n",
2813                                libcfs_nid2str(ni->ni_nid),
2814                                atomic_read(&net->ibn_npeers));
2815                         set_current_state(TASK_UNINTERRUPTIBLE);
2816                         schedule_timeout(cfs_time_seconds(1));
2817                 }
2818
2819                 kiblnd_net_fini_pools(net);
2820
2821                 write_lock_irqsave(g_lock, flags);
2822                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2823                 net->ibn_dev->ibd_nnets--;
2824                 list_del(&net->ibn_list);
2825                 write_unlock_irqrestore(g_lock, flags);
2826
2827                 /* fall through */
2828
2829         case IBLND_INIT_NOTHING:
2830                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2831
2832                 if (net->ibn_dev != NULL &&
2833                     net->ibn_dev->ibd_nnets == 0)
2834                         kiblnd_destroy_dev(net->ibn_dev);
2835
2836                 break;
2837         }
2838
2839         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2840                atomic_read(&libcfs_kmemory));
2841
2842         net->ibn_init = IBLND_INIT_NOTHING;
2843         ni->ni_data = NULL;
2844
2845         LIBCFS_FREE(net, sizeof(*net));
2846
2847 out:
2848         if (list_empty(&kiblnd_data.kib_devs))
2849                 kiblnd_base_shutdown();
2850         return;
2851 }
2852
2853 static int
2854 kiblnd_base_startup(void)
2855 {
2856         struct kib_sched_info   *sched;
2857         int                     rc;
2858         int                     i;
2859
2860         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2861
2862         try_module_get(THIS_MODULE);
2863         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2864
2865         rwlock_init(&kiblnd_data.kib_global_lock);
2866
2867         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2868         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2869
2870         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2871         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2872                      sizeof(struct list_head) *
2873                      kiblnd_data.kib_peer_hash_size);
2874         if (kiblnd_data.kib_peers == NULL)
2875                 goto failed;
2876
2877         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2878                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2879
2880         spin_lock_init(&kiblnd_data.kib_connd_lock);
2881         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2882         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2883         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2884         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2885
2886         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2887         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2888
2889         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2890                                                   sizeof(*sched));
2891         if (kiblnd_data.kib_scheds == NULL)
2892                 goto failed;
2893
2894         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2895                 int     nthrs;
2896
2897                 spin_lock_init(&sched->ibs_lock);
2898                 INIT_LIST_HEAD(&sched->ibs_conns);
2899                 init_waitqueue_head(&sched->ibs_waitq);
2900
2901                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2902                 if (*kiblnd_tunables.kib_nscheds > 0) {
2903                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2904                 } else {
2905                         /* max to half of CPUs, another half is reserved for
2906                          * upper layer modules */
2907                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2908                 }
2909
2910                 sched->ibs_nthreads_max = nthrs;
2911                 sched->ibs_cpt = i;
2912         }
2913
2914         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2915
2916         /* lists/ptrs/locks initialised */
2917         kiblnd_data.kib_init = IBLND_INIT_DATA;
2918         /*****************************************************/
2919
2920         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2921         if (rc != 0) {
2922                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2923                 goto failed;
2924         }
2925
2926         if (*kiblnd_tunables.kib_dev_failover != 0)
2927                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2928                                          "kiblnd_failover");
2929
2930         if (rc != 0) {
2931                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2932                 goto failed;
2933         }
2934
2935         /* flag everything initialised */
2936         kiblnd_data.kib_init = IBLND_INIT_ALL;
2937         /*****************************************************/
2938
2939         return 0;
2940
2941  failed:
2942         kiblnd_base_shutdown();
2943         return -ENETDOWN;
2944 }
2945
2946 static int
2947 kiblnd_start_schedulers(struct kib_sched_info *sched)
2948 {
2949         int     rc = 0;
2950         int     nthrs;
2951         int     i;
2952
2953         if (sched->ibs_nthreads == 0) {
2954                 if (*kiblnd_tunables.kib_nscheds > 0) {
2955                         nthrs = sched->ibs_nthreads_max;
2956                 } else {
2957                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2958                                                sched->ibs_cpt);
2959                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2960                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2961                 }
2962         } else {
2963                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2964                 /* increase one thread if there is new interface */
2965                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2966         }
2967
2968         for (i = 0; i < nthrs; i++) {
2969                 long    id;
2970                 char    name[20];
2971                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2972                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2973                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2974                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2975                 if (rc == 0)
2976                         continue;
2977
2978                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2979                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2980                 break;
2981         }
2982
2983         sched->ibs_nthreads += i;
2984         return rc;
2985 }
2986
2987 static int
2988 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2989 {
2990         int     cpt;
2991         int     rc;
2992         int     i;
2993
2994         for (i = 0; i < ncpts; i++) {
2995                 struct kib_sched_info *sched;
2996
2997                 cpt = (cpts == NULL) ? i : cpts[i];
2998                 sched = kiblnd_data.kib_scheds[cpt];
2999
3000                 if (!newdev && sched->ibs_nthreads > 0)
3001                         continue;
3002
3003                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3004                 if (rc != 0) {
3005                         CERROR("Failed to start scheduler threads for %s\n",
3006                                dev->ibd_ifname);
3007                         return rc;
3008                 }
3009         }
3010         return 0;
3011 }
3012
3013 static kib_dev_t *
3014 kiblnd_dev_search(char *ifname)
3015 {
3016         kib_dev_t       *alias = NULL;
3017         kib_dev_t       *dev;
3018         char            *colon;
3019         char            *colon2;
3020
3021         colon = strchr(ifname, ':');
3022         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3023                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3024                         return dev;
3025
3026                 if (alias != NULL)
3027                         continue;
3028
3029                 colon2 = strchr(dev->ibd_ifname, ':');
3030                 if (colon != NULL)
3031                         *colon = 0;
3032                 if (colon2 != NULL)
3033                         *colon2 = 0;
3034
3035                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3036                         alias = dev;
3037
3038                 if (colon != NULL)
3039                         *colon = ':';
3040                 if (colon2 != NULL)
3041                         *colon2 = ':';
3042         }
3043         return alias;
3044 }
3045
3046 static int
3047 kiblnd_startup (lnet_ni_t *ni)
3048 {
3049         char                     *ifname;
3050         kib_dev_t                *ibdev = NULL;
3051         kib_net_t                *net;
3052         struct timeval            tv;
3053         unsigned long             flags;
3054         int                       rc;
3055         int                       newdev;
3056
3057         LASSERT (ni->ni_lnd == &the_o2iblnd);
3058
3059         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3060                 rc = kiblnd_base_startup();
3061                 if (rc != 0)
3062                         return rc;
3063         }
3064
3065         LIBCFS_ALLOC(net, sizeof(*net));
3066         ni->ni_data = net;
3067         if (net == NULL)
3068                 goto failed;
3069
3070         do_gettimeofday(&tv);
3071         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3072
3073         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3074         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3075         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3076         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3077
3078         if (ni->ni_interfaces[0] != NULL) {
3079                 /* Use the IPoIB interface specified in 'networks=' */
3080
3081                 CLASSERT (LNET_MAX_INTERFACES > 1);
3082                 if (ni->ni_interfaces[1] != NULL) {
3083                         CERROR("Multiple interfaces not supported\n");
3084                         goto failed;
3085                 }
3086
3087                 ifname = ni->ni_interfaces[0];
3088         } else {
3089                 ifname = *kiblnd_tunables.kib_default_ipif;
3090         }
3091
3092         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3093                 CERROR("IPoIB interface name too long: %s\n", ifname);
3094                 goto failed;
3095         }
3096
3097         ibdev = kiblnd_dev_search(ifname);
3098
3099         newdev = ibdev == NULL;
3100         /* hmm...create kib_dev even for alias */
3101         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3102                 ibdev = kiblnd_create_dev(ifname);
3103
3104         if (ibdev == NULL)
3105                 goto failed;
3106
3107         net->ibn_dev = ibdev;
3108         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3109
3110         rc = kiblnd_dev_start_threads(ibdev, newdev,
3111                                       ni->ni_cpts, ni->ni_ncpts);
3112         if (rc != 0)
3113                 goto failed;
3114
3115         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3116         if (rc != 0) {
3117                 CERROR("Failed to initialize NI pools: %d\n", rc);
3118                 goto failed;
3119         }
3120
3121         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3122         ibdev->ibd_nnets++;
3123         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3124         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3125
3126         net->ibn_init = IBLND_INIT_ALL;
3127
3128         return 0;
3129
3130 failed:
3131         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3132                 kiblnd_destroy_dev(ibdev);
3133
3134         kiblnd_shutdown(ni);
3135
3136         CDEBUG(D_NET, "kiblnd_startup failed\n");
3137         return -ENETDOWN;
3138 }
3139
3140 static lnd_t the_o2iblnd = {
3141         .lnd_type       = O2IBLND,
3142         .lnd_startup    = kiblnd_startup,
3143         .lnd_shutdown   = kiblnd_shutdown,
3144         .lnd_ctl        = kiblnd_ctl,
3145         .lnd_query      = kiblnd_query,
3146         .lnd_send       = kiblnd_send,
3147         .lnd_recv       = kiblnd_recv,
3148 };
3149
3150 static void __exit ko2iblnd_exit(void)
3151 {
3152         lnet_unregister_lnd(&the_o2iblnd);
3153         kiblnd_tunables_fini();
3154 }
3155
3156 static int __init ko2iblnd_init(void)
3157 {
3158         int rc;
3159
3160         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3161         CLASSERT(offsetof(kib_msg_t,
3162                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3163                  IBLND_MSG_SIZE);
3164         CLASSERT(offsetof(kib_msg_t,
3165                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3166                  <= IBLND_MSG_SIZE);
3167
3168         rc = kiblnd_tunables_init();
3169         if (rc != 0)
3170                 return rc;
3171
3172         lnet_register_lnd(&the_o2iblnd);
3173
3174         return 0;
3175 }
3176
3177 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3178 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3179 MODULE_VERSION("2.8.0");
3180 MODULE_LICENSE("GPL");
3181
3182 module_init(ko2iblnd_init);
3183 module_exit(ko2iblnd_exit);