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
 679         vectors = conn->ibc_cmid->device->num_comp_vectors;
 680         if (vectors <= 1)
 681                 return 0;
 682
 683         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 684
 685         /* hash NID to CPU id in this partition... */
 686         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 687         for_each_cpu(i, mask) {
 688                 if (off-- == 0)
 689                         return i % vectors;
 690         }
 691
 692         LBUG();
 693         return 1;
 694 }
 695
 696 kib_conn_t *
 697 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 698                    int state, int version)
 699 {
 700         /* CAVEAT EMPTOR:
 701          * If the new conn is created successfully it takes over the caller's
 702          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 703          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 704          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 705          * to destroy 'cmid' here since I'm called from the CM which still has
 706          * its ref on 'cmid'). */
 707         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 708         kib_net_t              *net = peer->ibp_ni->ni_data;
 709         kib_dev_t              *dev;
 710         struct ib_qp_init_attr *init_qp_attr;
 711         struct kib_sched_info   *sched;
 712 #ifdef HAVE_IB_CQ_INIT_ATTR
 713         struct ib_cq_init_attr  cq_attr = {};
 714 #endif
 715         kib_conn_t              *conn;
 716         struct ib_cq            *cq;
 717         unsigned long           flags;
 718         int                     cpt;
 719         int                     rc;
 720         int                     i;
 721
 722         LASSERT(net != NULL);
 723         LASSERT(!in_interrupt());
 724
 725         dev = net->ibn_dev;
 726
 727         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 728         sched = kiblnd_data.kib_scheds[cpt];
 729
 730         LASSERT(sched->ibs_nthreads > 0);
 731
 732         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 733                          sizeof(*init_qp_attr));
 734         if (init_qp_attr == NULL) {
 735                 CERROR("Can't allocate qp_attr for %s\n",
 736                        libcfs_nid2str(peer->ibp_nid));
 737                 goto failed_0;
 738         }
 739
 740         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 741         if (conn == NULL) {
 742                 CERROR("Can't allocate connection for %s\n",
 743                        libcfs_nid2str(peer->ibp_nid));
 744                 goto failed_1;
 745         }
 746
 747         conn->ibc_state = IBLND_CONN_INIT;
 748         conn->ibc_version = version;
 749         conn->ibc_peer = peer;                  /* I take the caller's ref */
 750         cmid->context = conn;                   /* for future CM callbacks */
 751         conn->ibc_cmid = cmid;
 752         conn->ibc_max_frags = peer->ibp_max_frags;
 753         conn->ibc_queue_depth = peer->ibp_queue_depth;
 754
 755         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 756         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 757         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 758         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 759         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 760         INIT_LIST_HEAD(&conn->ibc_active_txs);
 761         spin_lock_init(&conn->ibc_lock);
 762
 763         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 764                          sizeof(*conn->ibc_connvars));
 765         if (conn->ibc_connvars == NULL) {
 766                 CERROR("Can't allocate in-progress connection state\n");
 767                 goto failed_2;
 768         }
 769
 770         write_lock_irqsave(glock, flags);
 771         if (dev->ibd_failover) {
 772                 write_unlock_irqrestore(glock, flags);
 773                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 774                 goto failed_2;
 775         }
 776
 777         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 778                 /* wakeup failover thread and teardown connection */
 779                 if (kiblnd_dev_can_failover(dev)) {
 780                         list_add_tail(&dev->ibd_fail_list,
 781                                       &kiblnd_data.kib_failed_devs);
 782                         wake_up(&kiblnd_data.kib_failover_waitq);
 783                 }
 784
 785                 write_unlock_irqrestore(glock, flags);
 786                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 787                        cmid->device->name, dev->ibd_ifname);
 788                 goto failed_2;
 789         }
 790
 791         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 792         conn->ibc_hdev = dev->ibd_hdev;
 793
 794         kiblnd_setup_mtu_locked(cmid);
 795
 796         write_unlock_irqrestore(glock, flags);
 797
 798         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 799                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 800         if (conn->ibc_rxs == NULL) {
 801                 CERROR("Cannot allocate RX buffers\n");
 802                 goto failed_2;
 803         }
 804
 805         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 806                                 IBLND_RX_MSG_PAGES(conn));
 807         if (rc != 0)
 808                 goto failed_2;
 809
 810         kiblnd_map_rx_descs(conn);
 811
 812 #ifdef HAVE_IB_CQ_INIT_ATTR
 813         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 814         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 815         cq = ib_create_cq(cmid->device,
 816                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 817                           &cq_attr);
 818 #else
 819         cq = ib_create_cq(cmid->device,
 820                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 821                           IBLND_CQ_ENTRIES(conn),
 822                           kiblnd_get_completion_vector(conn, cpt));
 823 #endif
 824         if (IS_ERR(cq)) {
 825                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 826                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 827                 goto failed_2;
 828         }
 829
 830         conn->ibc_cq = cq;
 831
 832         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 833         if (rc != 0) {
 834                 CERROR("Can't request completion notification: %d\n", rc);
 835                 goto failed_2;
 836         }
 837
 838         init_qp_attr->event_handler = kiblnd_qp_event;
 839         init_qp_attr->qp_context = conn;
 840         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 841         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 842         init_qp_attr->cap.max_send_sge = 1;
 843         init_qp_attr->cap.max_recv_sge = 1;
 844         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 845         init_qp_attr->qp_type = IB_QPT_RC;
 846         init_qp_attr->send_cq = cq;
 847         init_qp_attr->recv_cq = cq;
 848
 849         conn->ibc_sched = sched;
 850
 851         do {
 852                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 853                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 854                         break;
 855
 856                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 857         } while (rc);
 858
 859         if (rc) {
 860                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 861                         rc, init_qp_attr->cap.max_send_wr,
 862                         init_qp_attr->cap.max_recv_wr);
 863                 goto failed_2;
 864         }
 865
 866         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 867                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 868                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 869
 870         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 871
 872         /* 1 ref for caller and each rxmsg */
 873         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 874         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 875
 876         /* post receives */
 877         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 878                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 879                 if (rc != 0) {
 880                         CERROR("Can't post rxmsg: %d\n", rc);
 881
 882                         /* Make posted receives complete */
 883                         kiblnd_abort_receives(conn);
 884
 885                         /* correct # of posted buffers
 886                          * NB locking needed now I'm racing with completion */
 887                         spin_lock_irqsave(&sched->ibs_lock, flags);
 888                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 889                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 890
 891                         /* cmid will be destroyed by CM(ofed) after cm_callback
 892                          * returned, so we can't refer it anymore
 893                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 894                         rdma_destroy_qp(conn->ibc_cmid);
 895                         conn->ibc_cmid = NULL;
 896
 897                         /* Drop my own and unused rxbuffer refcounts */
 898                         while (i++ <= IBLND_RX_MSGS(conn))
 899                                 kiblnd_conn_decref(conn);
 900
 901                         return NULL;
 902                 }
 903         }
 904
 905         /* Init successful! */
 906         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 907                  state == IBLND_CONN_PASSIVE_WAIT);
 908         conn->ibc_state = state;
 909
 910         /* 1 more conn */
 911         atomic_inc(&net->ibn_nconns);
 912         return conn;
 913
 914  failed_2:
 915         kiblnd_destroy_conn(conn, true);
 916  failed_1:
 917         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 918  failed_0:
 919         return NULL;
 920 }
 921
 922 void
 923 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 924 {
 925         struct rdma_cm_id *cmid = conn->ibc_cmid;
 926         kib_peer_t        *peer = conn->ibc_peer;
 927         int                rc;
 928
 929         LASSERT (!in_interrupt());
 930         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 931         LASSERT(list_empty(&conn->ibc_early_rxs));
 932         LASSERT(list_empty(&conn->ibc_tx_noops));
 933         LASSERT(list_empty(&conn->ibc_tx_queue));
 934         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 935         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 936         LASSERT(list_empty(&conn->ibc_active_txs));
 937         LASSERT (conn->ibc_noops_posted == 0);
 938         LASSERT (conn->ibc_nsends_posted == 0);
 939
 940         switch (conn->ibc_state) {
 941         default:
 942                 /* conn must be completely disengaged from the network */
 943                 LBUG();
 944
 945         case IBLND_CONN_DISCONNECTED:
 946                 /* connvars should have been freed already */
 947                 LASSERT (conn->ibc_connvars == NULL);
 948                 break;
 949
 950         case IBLND_CONN_INIT:
 951                 break;
 952         }
 953
 954         /* conn->ibc_cmid might be destroyed by CM already */
 955         if (cmid != NULL && cmid->qp != NULL)
 956                 rdma_destroy_qp(cmid);
 957
 958         if (conn->ibc_cq != NULL) {
 959                 rc = ib_destroy_cq(conn->ibc_cq);
 960                 if (rc != 0)
 961                         CWARN("Error destroying CQ: %d\n", rc);
 962         }
 963
 964         if (conn->ibc_rx_pages != NULL)
 965                 kiblnd_unmap_rx_descs(conn);
 966
 967         if (conn->ibc_rxs != NULL) {
 968                 LIBCFS_FREE(conn->ibc_rxs,
 969                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 970         }
 971
 972         if (conn->ibc_connvars != NULL)
 973                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 974
 975         if (conn->ibc_hdev != NULL)
 976                 kiblnd_hdev_decref(conn->ibc_hdev);
 977
 978         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 979         if (conn->ibc_state != IBLND_CONN_INIT) {
 980                 kib_net_t *net = peer->ibp_ni->ni_data;
 981
 982                 kiblnd_peer_decref(peer);
 983                 rdma_destroy_id(cmid);
 984                 atomic_dec(&net->ibn_nconns);
 985         }
 986
 987         if (free_conn)
 988                 LIBCFS_FREE(conn, sizeof(*conn));
 989 }
 990
 991 int
 992 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 993 {
 994         kib_conn_t              *conn;
 995         struct list_head        *ctmp;
 996         struct list_head        *cnxt;
 997         int                     count = 0;
 998
 999         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1000                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1001
1002                 CDEBUG(D_NET, "Closing conn -> %s, "
1003                               "version: %x, reason: %d\n",
1004                        libcfs_nid2str(peer->ibp_nid),
1005                        conn->ibc_version, why);
1006
1007                 kiblnd_close_conn_locked(conn, why);
1008                 count++;
1009         }
1010
1011         return count;
1012 }
1013
1014 int
1015 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1016                                 int version, __u64 incarnation)
1017 {
1018         kib_conn_t              *conn;
1019         struct list_head        *ctmp;
1020         struct list_head        *cnxt;
1021         int                     count = 0;
1022
1023         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1024                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1025
1026                 if (conn->ibc_version     == version &&
1027                     conn->ibc_incarnation == incarnation)
1028                         continue;
1029
1030                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1031                               "incarnation:"LPX64"(%x, "LPX64")\n",
1032                        libcfs_nid2str(peer->ibp_nid),
1033                        conn->ibc_version, conn->ibc_incarnation,
1034                        version, incarnation);
1035
1036                 kiblnd_close_conn_locked(conn, -ESTALE);
1037                 count++;
1038         }
1039
1040         return count;
1041 }
1042
1043 static int
1044 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1045 {
1046         kib_peer_t              *peer;
1047         struct list_head        *ptmp;
1048         struct list_head        *pnxt;
1049         int                     lo;
1050         int                     hi;
1051         int                     i;
1052         unsigned long           flags;
1053         int                     count = 0;
1054
1055         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1056
1057         if (nid != LNET_NID_ANY)
1058                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1059         else {
1060                 lo = 0;
1061                 hi = kiblnd_data.kib_peer_hash_size - 1;
1062         }
1063
1064         for (i = lo; i <= hi; i++) {
1065                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1066
1067                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1068                         LASSERT(!kiblnd_peer_idle(peer));
1069
1070                         if (peer->ibp_ni != ni)
1071                                 continue;
1072
1073                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1074                                 continue;
1075
1076                         count += kiblnd_close_peer_conns_locked(peer, 0);
1077                 }
1078         }
1079
1080         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1081
1082         /* wildcards always succeed */
1083         if (nid == LNET_NID_ANY)
1084                 return 0;
1085
1086         return (count == 0) ? -ENOENT : 0;
1087 }
1088
1089 static int
1090 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1091 {
1092         struct libcfs_ioctl_data *data = arg;
1093         int                       rc = -EINVAL;
1094
1095         switch(cmd) {
1096         case IOC_LIBCFS_GET_PEER: {
1097                 lnet_nid_t   nid = 0;
1098                 int          count = 0;
1099
1100                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1101                                           &nid, &count);
1102                 data->ioc_nid    = nid;
1103                 data->ioc_count  = count;
1104                 break;
1105         }
1106
1107         case IOC_LIBCFS_DEL_PEER: {
1108                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1109                 break;
1110         }
1111         case IOC_LIBCFS_GET_CONN: {
1112                 kib_conn_t *conn;
1113
1114                 rc = 0;
1115                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1116                 if (conn == NULL) {
1117                         rc = -ENOENT;
1118                         break;
1119                 }
1120
1121                 LASSERT (conn->ibc_cmid != NULL);
1122                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1123                 if (conn->ibc_cmid->route.path_rec == NULL)
1124                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1125                 else
1126                         data->ioc_u32[0] =
1127                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1128                 kiblnd_conn_decref(conn);
1129                 break;
1130         }
1131         case IOC_LIBCFS_CLOSE_CONNECTION: {
1132                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1133                 break;
1134         }
1135
1136         default:
1137                 break;
1138         }
1139
1140         return rc;
1141 }
1142
1143 static void
1144 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1145 {
1146         cfs_time_t      last_alive = 0;
1147         cfs_time_t      now = cfs_time_current();
1148         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1149         kib_peer_t      *peer;
1150         unsigned long   flags;
1151
1152         read_lock_irqsave(glock, flags);
1153
1154         peer = kiblnd_find_peer_locked(nid);
1155         if (peer != NULL)
1156                 last_alive = peer->ibp_last_alive;
1157
1158         read_unlock_irqrestore(glock, flags);
1159
1160         if (last_alive != 0)
1161                 *when = last_alive;
1162
1163         /* peer is not persistent in hash, trigger peer creation
1164          * and connection establishment with a NULL tx */
1165         if (peer == NULL)
1166                 kiblnd_launch_tx(ni, NULL, nid);
1167
1168         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1169                libcfs_nid2str(nid), peer,
1170                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1171         return;
1172 }
1173
1174 static void
1175 kiblnd_free_pages(kib_pages_t *p)
1176 {
1177         int     npages = p->ibp_npages;
1178         int     i;
1179
1180         for (i = 0; i < npages; i++) {
1181                 if (p->ibp_pages[i] != NULL)
1182                         __free_page(p->ibp_pages[i]);
1183         }
1184
1185         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1186 }
1187
1188 int
1189 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1190 {
1191         kib_pages_t     *p;
1192         int             i;
1193
1194         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1195                          offsetof(kib_pages_t, ibp_pages[npages]));
1196         if (p == NULL) {
1197                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1198                 return -ENOMEM;
1199         }
1200
1201         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1202         p->ibp_npages = npages;
1203
1204         for (i = 0; i < npages; i++) {
1205                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1206                                                      GFP_NOFS);
1207                 if (p->ibp_pages[i] == NULL) {
1208                         CERROR("Can't allocate page %d of %d\n", i, npages);
1209                         kiblnd_free_pages(p);
1210                         return -ENOMEM;
1211                 }
1212         }
1213
1214         *pp = p;
1215         return 0;
1216 }
1217
1218 void
1219 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1220 {
1221         kib_rx_t *rx;
1222         int       i;
1223
1224         LASSERT (conn->ibc_rxs != NULL);
1225         LASSERT (conn->ibc_hdev != NULL);
1226
1227         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1228                 rx = &conn->ibc_rxs[i];
1229
1230                 LASSERT(rx->rx_nob >= 0); /* not posted */
1231
1232                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1233                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1234                                                           rx->rx_msgaddr),
1235                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1236         }
1237
1238         kiblnd_free_pages(conn->ibc_rx_pages);
1239
1240         conn->ibc_rx_pages = NULL;
1241 }
1242
1243 void
1244 kiblnd_map_rx_descs(kib_conn_t *conn)
1245 {
1246         kib_rx_t       *rx;
1247         struct page    *pg;
1248         int             pg_off;
1249         int             ipg;
1250         int             i;
1251
1252         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1253                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1254                 rx = &conn->ibc_rxs[i];
1255
1256                 rx->rx_conn = conn;
1257                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1258
1259                 rx->rx_msgaddr =
1260                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1261                                               rx->rx_msg, IBLND_MSG_SIZE,
1262                                               DMA_FROM_DEVICE);
1263                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1264                                                   rx->rx_msgaddr));
1265                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1266
1267                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1268                        i, rx->rx_msg, rx->rx_msgaddr,
1269                        (__u64)(page_to_phys(pg) + pg_off));
1270
1271                 pg_off += IBLND_MSG_SIZE;
1272                 LASSERT(pg_off <= PAGE_SIZE);
1273
1274                 if (pg_off == PAGE_SIZE) {
1275                         pg_off = 0;
1276                         ipg++;
1277                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1278                 }
1279         }
1280 }
1281
1282 static void
1283 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1284 {
1285         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1286         kib_tx_t       *tx;
1287         int             i;
1288
1289         LASSERT (tpo->tpo_pool.po_allocated == 0);
1290
1291         if (hdev == NULL)
1292                 return;
1293
1294         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1295                 tx = &tpo->tpo_tx_descs[i];
1296                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1297                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1298                                                           tx->tx_msgaddr),
1299                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1300         }
1301
1302         kiblnd_hdev_decref(hdev);
1303         tpo->tpo_hdev = NULL;
1304 }
1305
1306 static kib_hca_dev_t *
1307 kiblnd_current_hdev(kib_dev_t *dev)
1308 {
1309         kib_hca_dev_t *hdev;
1310         unsigned long  flags;
1311         int            i = 0;
1312
1313         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1314         while (dev->ibd_failover) {
1315                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1316                 if (i++ % 50 == 0)
1317                         CDEBUG(D_NET, "%s: Wait for failover\n",
1318                                dev->ibd_ifname);
1319                 set_current_state(TASK_INTERRUPTIBLE);
1320                 schedule_timeout(cfs_time_seconds(1) / 100);
1321
1322                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1323         }
1324
1325         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1326         hdev = dev->ibd_hdev;
1327
1328         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1329
1330         return hdev;
1331 }
1332
1333 static void
1334 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1335 {
1336         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1337         kib_pool_t     *pool  = &tpo->tpo_pool;
1338         kib_net_t      *net   = pool->po_owner->ps_net;
1339         kib_dev_t      *dev;
1340         struct page    *page;
1341         kib_tx_t       *tx;
1342         int             page_offset;
1343         int             ipage;
1344         int             i;
1345
1346         LASSERT (net != NULL);
1347
1348         dev = net->ibn_dev;
1349
1350         /* pre-mapped messages are not bigger than 1 page */
1351         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1352
1353         /* No fancy arithmetic when we do the buffer calculations */
1354         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1355
1356         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1357
1358         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1359                 page = txpgs->ibp_pages[ipage];
1360                 tx = &tpo->tpo_tx_descs[i];
1361
1362                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1363                                            page_offset);
1364
1365                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1366                                                        tx->tx_msg,
1367                                                        IBLND_MSG_SIZE,
1368                                                        DMA_TO_DEVICE);
1369                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1370                                                   tx->tx_msgaddr));
1371                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1372
1373                 list_add(&tx->tx_list, &pool->po_free_list);
1374
1375                 page_offset += IBLND_MSG_SIZE;
1376                 LASSERT(page_offset <= PAGE_SIZE);
1377
1378                 if (page_offset == PAGE_SIZE) {
1379                         page_offset = 0;
1380                         ipage++;
1381                         LASSERT(ipage <= txpgs->ibp_npages);
1382                 }
1383         }
1384 }
1385
1386 struct ib_mr *
1387 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1388                       int negotiated_nfrags)
1389 {
1390         __u16   nfrags = (negotiated_nfrags != -1) ?
1391           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1392
1393         LASSERT(hdev->ibh_mrs != NULL);
1394
1395         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1396             nfrags <= rd->rd_nfrags)
1397                 return NULL;
1398
1399         return hdev->ibh_mrs;
1400 }
1401
1402 static void
1403 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1404 {
1405         LASSERT(fpo->fpo_map_count == 0);
1406
1407         if (fpo->fpo_is_fmr) {
1408                 if (fpo->fmr.fpo_fmr_pool)
1409                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1410         } else {
1411                 struct kib_fast_reg_descriptor *frd, *tmp;
1412                 int i = 0;
1413
1414                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1415                                          frd_list) {
1416                         list_del(&frd->frd_list);
1417                         ib_free_fast_reg_page_list(frd->frd_frpl);
1418                         ib_dereg_mr(frd->frd_mr);
1419                         LIBCFS_FREE(frd, sizeof(*frd));
1420                         i++;
1421                 }
1422                 if (i < fpo->fast_reg.fpo_pool_size)
1423                         CERROR("FastReg pool still has %d regions registered\n",
1424                                 fpo->fast_reg.fpo_pool_size - i);
1425         }
1426
1427         if (fpo->fpo_hdev)
1428                 kiblnd_hdev_decref(fpo->fpo_hdev);
1429
1430         LIBCFS_FREE(fpo, sizeof(*fpo));
1431 }
1432
1433 static void
1434 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1435 {
1436         kib_fmr_pool_t *fpo, *tmp;
1437
1438         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1439                 list_del(&fpo->fpo_list);
1440                 kiblnd_destroy_fmr_pool(fpo);
1441         }
1442 }
1443
1444 static int kiblnd_fmr_pool_size(int ncpts)
1445 {
1446         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1447
1448         return max(IBLND_FMR_POOL, size);
1449 }
1450
1451 static int kiblnd_fmr_flush_trigger(int ncpts)
1452 {
1453         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1454
1455         return max(IBLND_FMR_POOL_FLUSH, size);
1456 }
1457
1458 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1459 {
1460         struct ib_fmr_pool_param param = {
1461                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1462                 .page_shift        = PAGE_SHIFT,
1463                 .access            = (IB_ACCESS_LOCAL_WRITE |
1464                                       IB_ACCESS_REMOTE_WRITE),
1465                 .pool_size         = fps->fps_pool_size,
1466                 .dirty_watermark   = fps->fps_flush_trigger,
1467                 .flush_function    = NULL,
1468                 .flush_arg         = NULL,
1469                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1470         int rc = 0;
1471
1472         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1473                                                    &param);
1474         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1475                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1476                 if (rc != -ENOSYS)
1477                         CERROR("Failed to create FMR pool: %d\n", rc);
1478                 else
1479                         CERROR("FMRs are not supported\n");
1480         }
1481
1482         return rc;
1483 }
1484
1485 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1486 {
1487         struct kib_fast_reg_descriptor *frd, *tmp;
1488         int i, rc;
1489
1490         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1491         fpo->fast_reg.fpo_pool_size = 0;
1492         for (i = 0; i < fps->fps_pool_size; i++) {
1493                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1494                                  sizeof(*frd));
1495                 if (!frd) {
1496                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1497                         rc = -ENOMEM;
1498                         goto out;
1499                 }
1500                 frd->frd_mr = NULL;
1501
1502                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1503                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1504                 if (IS_ERR(frd->frd_frpl)) {
1505                         rc = PTR_ERR(frd->frd_frpl);
1506                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1507                                 rc);
1508                         goto out_middle;
1509                 }
1510
1511                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1512                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1513                 if (IS_ERR(frd->frd_mr)) {
1514                         rc = PTR_ERR(frd->frd_mr);
1515                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1516                         goto out_middle;
1517                 }
1518
1519                 frd->frd_valid = true;
1520
1521                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1522                 fpo->fast_reg.fpo_pool_size++;
1523         }
1524
1525         return 0;
1526
1527 out_middle:
1528         if (frd->frd_mr)
1529                 ib_dereg_mr(frd->frd_mr);
1530         if (frd->frd_frpl)
1531                 ib_free_fast_reg_page_list(frd->frd_frpl);
1532         LIBCFS_FREE(frd, sizeof(*frd));
1533
1534 out:
1535         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1536                                  frd_list) {
1537                 list_del(&frd->frd_list);
1538                 ib_free_fast_reg_page_list(frd->frd_frpl);
1539                 ib_dereg_mr(frd->frd_mr);
1540                 LIBCFS_FREE(frd, sizeof(*frd));
1541         }
1542
1543         return rc;
1544 }
1545
1546 static int
1547 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1548 {
1549         struct ib_device_attr *dev_attr;
1550         kib_dev_t *dev = fps->fps_net->ibn_dev;
1551         kib_fmr_pool_t *fpo;
1552         int rc;
1553
1554         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1555         if (!dev_attr)
1556                 return -ENOMEM;
1557
1558         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1559         if (!fpo) {
1560                 rc = -ENOMEM;
1561                 goto out_dev_attr;
1562         }
1563
1564         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1565
1566         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1567         if (rc) {
1568                 CERROR("Query device failed for %s: %d\n",
1569                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1570                 goto out_dev_attr;
1571         }
1572
1573         /* Check for FMR or FastReg support */
1574         fpo->fpo_is_fmr = 0;
1575         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1576             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1577             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1578             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1579                 LCONSOLE_INFO("Using FMR for registration\n");
1580                 fpo->fpo_is_fmr = 1;
1581         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1582                 LCONSOLE_INFO("Using FastReg for registration\n");
1583         } else {
1584                 rc = -ENOSYS;
1585                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1586                 goto out_dev_attr;
1587         }
1588
1589         if (fpo->fpo_is_fmr)
1590                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1591         else
1592                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1593         if (rc)
1594                 goto out_fpo;
1595
1596         kfree(dev_attr);
1597         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1598         fpo->fpo_owner    = fps;
1599         *pp_fpo = fpo;
1600
1601         return 0;
1602
1603 out_fpo:
1604         kiblnd_hdev_decref(fpo->fpo_hdev);
1605         LIBCFS_FREE(fpo, sizeof(*fpo));
1606
1607 out_dev_attr:
1608         kfree(dev_attr);
1609
1610         return rc;
1611 }
1612
1613 static void
1614 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1615 {
1616         if (fps->fps_net == NULL) /* intialized? */
1617                 return;
1618
1619         spin_lock(&fps->fps_lock);
1620
1621         while (!list_empty(&fps->fps_pool_list)) {
1622                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1623                                                  kib_fmr_pool_t, fpo_list);
1624                 fpo->fpo_failed = 1;
1625                 list_del(&fpo->fpo_list);
1626                 if (fpo->fpo_map_count == 0)
1627                         list_add(&fpo->fpo_list, zombies);
1628                 else
1629                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1630         }
1631
1632         spin_unlock(&fps->fps_lock);
1633 }
1634
1635 static void
1636 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1637 {
1638         if (fps->fps_net != NULL) { /* initialized? */
1639                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1640                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1641         }
1642 }
1643
1644 static int
1645 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1646                         int pool_size, int flush_trigger)
1647 {
1648         kib_fmr_pool_t *fpo;
1649         int             rc;
1650
1651         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1652
1653         fps->fps_net = net;
1654         fps->fps_cpt = cpt;
1655         fps->fps_pool_size = pool_size;
1656         fps->fps_flush_trigger = flush_trigger;
1657         spin_lock_init(&fps->fps_lock);
1658         INIT_LIST_HEAD(&fps->fps_pool_list);
1659         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1660
1661         rc = kiblnd_create_fmr_pool(fps, &fpo);
1662         if (rc == 0)
1663                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1664
1665         return rc;
1666 }
1667
1668 static int
1669 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1670 {
1671         if (fpo->fpo_map_count != 0) /* still in use */
1672                 return 0;
1673         if (fpo->fpo_failed)
1674                 return 1;
1675         return cfs_time_aftereq(now, fpo->fpo_deadline);
1676 }
1677
1678 void
1679 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1680 {
1681         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1682         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1683         kib_fmr_poolset_t *fps;
1684         cfs_time_t         now = cfs_time_current();
1685         kib_fmr_pool_t    *tmp;
1686         int                rc;
1687
1688         if (!fpo)
1689                 return;
1690
1691         fps = fpo->fpo_owner;
1692         if (fpo->fpo_is_fmr) {
1693                 if (fmr->fmr_pfmr) {
1694                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1695                         LASSERT(!rc);
1696                         fmr->fmr_pfmr = NULL;
1697                 }
1698
1699                 if (status) {
1700                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1701                         LASSERT(!rc);
1702                 }
1703         } else {
1704                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1705
1706                 if (frd) {
1707                         frd->frd_valid = false;
1708                         spin_lock(&fps->fps_lock);
1709                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1710                         spin_unlock(&fps->fps_lock);
1711                         fmr->fmr_frd = NULL;
1712                 }
1713         }
1714         fmr->fmr_pool = NULL;
1715
1716         spin_lock(&fps->fps_lock);
1717         fpo->fpo_map_count--;   /* decref the pool */
1718
1719         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1720                 /* the first pool is persistent */
1721                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1722                         continue;
1723
1724                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1725                         list_move(&fpo->fpo_list, &zombies);
1726                         fps->fps_version++;
1727                 }
1728         }
1729         spin_unlock(&fps->fps_lock);
1730
1731         if (!list_empty(&zombies))
1732                 kiblnd_destroy_fmr_pool_list(&zombies);
1733 }
1734
1735 int
1736 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1737                     __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1738 {
1739         kib_fmr_pool_t *fpo;
1740         __u64 version;
1741         int rc;
1742
1743 again:
1744         spin_lock(&fps->fps_lock);
1745         version = fps->fps_version;
1746         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1747                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1748                 fpo->fpo_map_count++;
1749
1750                 if (fpo->fpo_is_fmr) {
1751                         struct ib_pool_fmr *pfmr;
1752
1753                         spin_unlock(&fps->fps_lock);
1754                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1755                                                     pages, npages, iov);
1756                         if (likely(!IS_ERR(pfmr))) {
1757                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1758                                                       : pfmr->fmr->lkey;
1759                                 fmr->fmr_frd  = NULL;
1760                                 fmr->fmr_pfmr = pfmr;
1761                                 fmr->fmr_pool = fpo;
1762                                 return 0;
1763                         }
1764                         rc = PTR_ERR(pfmr);
1765                 } else {
1766                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1767                                 struct ib_send_wr *wr;
1768                                 struct kib_fast_reg_descriptor *frd;
1769                                 struct ib_fast_reg_page_list *frpl;
1770                                 struct ib_mr *mr;
1771
1772                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1773                                                         struct kib_fast_reg_descriptor,
1774                                                         frd_list);
1775                                 list_del(&frd->frd_list);
1776                                 spin_unlock(&fps->fps_lock);
1777
1778                                 frpl = frd->frd_frpl;
1779                                 mr   = frd->frd_mr;
1780
1781                                 if (!frd->frd_valid) {
1782                                         struct ib_send_wr *inv_wr;
1783                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1784
1785                                         inv_wr = &frd->frd_inv_wr;
1786                                         memset(inv_wr, 0, sizeof(*inv_wr));
1787                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1788                                         inv_wr->wr_id = IBLND_WID_MR;
1789                                         inv_wr->ex.invalidate_rkey = key;
1790
1791                                         /* Bump the key */
1792                                         key = ib_inc_rkey(key);
1793                                         ib_update_fast_reg_key(mr, key);
1794                                 }
1795
1796                                 LASSERT(npages <= frpl->max_page_list_len);
1797                                 memcpy(frpl->page_list, pages,
1798                                         sizeof(*pages) * npages);
1799
1800                                 /* Prepare FastReg WR */
1801                                 wr = &frd->frd_fastreg_wr;
1802                                 memset(wr, 0, sizeof(*wr));
1803                                 wr->opcode = IB_WR_FAST_REG_MR;
1804                                 wr->wr_id = IBLND_WID_MR;
1805                                 wr->wr.fast_reg.iova_start = iov;
1806                                 wr->wr.fast_reg.page_list  = frpl;
1807                                 wr->wr.fast_reg.page_list_len = npages;
1808                                 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1809                                 wr->wr.fast_reg.length = nob;
1810                                 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1811                                                              : mr->lkey;
1812                                 wr->wr.fast_reg.access_flags =
1813                                                 (IB_ACCESS_LOCAL_WRITE |
1814                                                  IB_ACCESS_REMOTE_WRITE);
1815
1816                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1817                                 fmr->fmr_frd  = frd;
1818                                 fmr->fmr_pfmr = NULL;
1819                                 fmr->fmr_pool = fpo;
1820                                 return 0;
1821                         }
1822                         spin_unlock(&fps->fps_lock);
1823                         rc = -EBUSY;
1824                 }
1825
1826                 spin_lock(&fps->fps_lock);
1827                 fpo->fpo_map_count--;
1828                 if (rc != -EAGAIN) {
1829                         spin_unlock(&fps->fps_lock);
1830                         return rc;
1831                 }
1832
1833                 /* EAGAIN and ... */
1834                 if (version != fps->fps_version) {
1835                         spin_unlock(&fps->fps_lock);
1836                         goto again;
1837                 }
1838         }
1839
1840         if (fps->fps_increasing) {
1841                 spin_unlock(&fps->fps_lock);
1842                 CDEBUG(D_NET, "Another thread is allocating new "
1843                        "FMR pool, waiting for her to complete\n");
1844                 schedule();
1845                 goto again;
1846
1847         }
1848
1849         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1850                 /* someone failed recently */
1851                 spin_unlock(&fps->fps_lock);
1852                 return -EAGAIN;
1853         }
1854
1855         fps->fps_increasing = 1;
1856         spin_unlock(&fps->fps_lock);
1857
1858         CDEBUG(D_NET, "Allocate new FMR pool\n");
1859         rc = kiblnd_create_fmr_pool(fps, &fpo);
1860         spin_lock(&fps->fps_lock);
1861         fps->fps_increasing = 0;
1862         if (rc == 0) {
1863                 fps->fps_version++;
1864                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1865         } else {
1866                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1867         }
1868         spin_unlock(&fps->fps_lock);
1869
1870         goto again;
1871 }
1872
1873 static void
1874 kiblnd_fini_pool(kib_pool_t *pool)
1875 {
1876         LASSERT(list_empty(&pool->po_free_list));
1877         LASSERT(pool->po_allocated == 0);
1878
1879         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1880 }
1881
1882 static void
1883 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1884 {
1885         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1886
1887         memset(pool, 0, sizeof(kib_pool_t));
1888         INIT_LIST_HEAD(&pool->po_free_list);
1889         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1890         pool->po_owner    = ps;
1891         pool->po_size     = size;
1892 }
1893
1894 static void
1895 kiblnd_destroy_pool_list(struct list_head *head)
1896 {
1897         kib_pool_t *pool;
1898
1899         while (!list_empty(head)) {
1900                 pool = list_entry(head->next, kib_pool_t, po_list);
1901                 list_del(&pool->po_list);
1902
1903                 LASSERT(pool->po_owner != NULL);
1904                 pool->po_owner->ps_pool_destroy(pool);
1905         }
1906 }
1907
1908 static void
1909 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1910 {
1911         if (ps->ps_net == NULL) /* intialized? */
1912                 return;
1913
1914         spin_lock(&ps->ps_lock);
1915         while (!list_empty(&ps->ps_pool_list)) {
1916                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1917                                             kib_pool_t, po_list);
1918                 po->po_failed = 1;
1919                 list_del(&po->po_list);
1920                 if (po->po_allocated == 0)
1921                         list_add(&po->po_list, zombies);
1922                 else
1923                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1924         }
1925         spin_unlock(&ps->ps_lock);
1926 }
1927
1928 static void
1929 kiblnd_fini_poolset(kib_poolset_t *ps)
1930 {
1931         if (ps->ps_net != NULL) { /* initialized? */
1932                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1933                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1934         }
1935 }
1936
1937 static int
1938 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1939                     kib_net_t *net, char *name, int size,
1940                     kib_ps_pool_create_t po_create,
1941                     kib_ps_pool_destroy_t po_destroy,
1942                     kib_ps_node_init_t nd_init,
1943                     kib_ps_node_fini_t nd_fini)
1944 {
1945         kib_pool_t      *pool;
1946         int             rc;
1947
1948         memset(ps, 0, sizeof(kib_poolset_t));
1949
1950         ps->ps_cpt          = cpt;
1951         ps->ps_net          = net;
1952         ps->ps_pool_create  = po_create;
1953         ps->ps_pool_destroy = po_destroy;
1954         ps->ps_node_init    = nd_init;
1955         ps->ps_node_fini    = nd_fini;
1956         ps->ps_pool_size    = size;
1957         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1958             >= sizeof(ps->ps_name))
1959                 return -E2BIG;
1960         spin_lock_init(&ps->ps_lock);
1961         INIT_LIST_HEAD(&ps->ps_pool_list);
1962         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1963
1964         rc = ps->ps_pool_create(ps, size, &pool);
1965         if (rc == 0)
1966                 list_add(&pool->po_list, &ps->ps_pool_list);
1967         else
1968                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1969
1970         return rc;
1971 }
1972
1973 static int
1974 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1975 {
1976         if (pool->po_allocated != 0) /* still in use */
1977                 return 0;
1978         if (pool->po_failed)
1979                 return 1;
1980         return cfs_time_aftereq(now, pool->po_deadline);
1981 }
1982
1983 void
1984 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1985 {
1986         struct list_head zombies = LIST_HEAD_INIT(zombies);
1987         kib_poolset_t   *ps = pool->po_owner;
1988         kib_pool_t      *tmp;
1989         cfs_time_t       now = cfs_time_current();
1990
1991         spin_lock(&ps->ps_lock);
1992
1993         if (ps->ps_node_fini != NULL)
1994                 ps->ps_node_fini(pool, node);
1995
1996         LASSERT(pool->po_allocated > 0);
1997         list_add(node, &pool->po_free_list);
1998         pool->po_allocated--;
1999
2000         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2001                 /* the first pool is persistent */
2002                 if (ps->ps_pool_list.next == &pool->po_list)
2003                         continue;
2004
2005                 if (kiblnd_pool_is_idle(pool, now))
2006                         list_move(&pool->po_list, &zombies);
2007         }
2008         spin_unlock(&ps->ps_lock);
2009
2010         if (!list_empty(&zombies))
2011                 kiblnd_destroy_pool_list(&zombies);
2012 }
2013
2014 struct list_head *
2015 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2016 {
2017         struct list_head        *node;
2018         kib_pool_t              *pool;
2019         int                     rc;
2020         unsigned int            interval = 1;
2021         cfs_time_t              time_before;
2022         unsigned int            trips = 0;
2023
2024 again:
2025         spin_lock(&ps->ps_lock);
2026         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2027                 if (list_empty(&pool->po_free_list))
2028                         continue;
2029
2030                 pool->po_allocated++;
2031                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2032                 node = pool->po_free_list.next;
2033                 list_del(node);
2034
2035                 if (ps->ps_node_init != NULL) {
2036                         /* still hold the lock */
2037                         ps->ps_node_init(pool, node);
2038                 }
2039                 spin_unlock(&ps->ps_lock);
2040                 return node;
2041         }
2042
2043         /* no available tx pool and ... */
2044         if (ps->ps_increasing) {
2045                 /* another thread is allocating a new pool */
2046                 spin_unlock(&ps->ps_lock);
2047                 trips++;
2048                 CDEBUG(D_NET, "Another thread is allocating new "
2049                        "%s pool, waiting %d HZs for her to complete."
2050                        "trips = %d\n",
2051                        ps->ps_name, interval, trips);
2052
2053                 set_current_state(TASK_INTERRUPTIBLE);
2054                 schedule_timeout(interval);
2055                 if (interval < cfs_time_seconds(1))
2056                         interval *= 2;
2057
2058                 goto again;
2059         }
2060
2061         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2062                 /* someone failed recently */
2063                 spin_unlock(&ps->ps_lock);
2064                 return NULL;
2065         }
2066
2067         ps->ps_increasing = 1;
2068         spin_unlock(&ps->ps_lock);
2069
2070         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2071         time_before = cfs_time_current();
2072         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2073         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2074                cfs_time_current() - time_before);
2075
2076         spin_lock(&ps->ps_lock);
2077         ps->ps_increasing = 0;
2078         if (rc == 0) {
2079                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2080         } else {
2081                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2082                 CERROR("Can't allocate new %s pool because out of memory\n",
2083                        ps->ps_name);
2084         }
2085         spin_unlock(&ps->ps_lock);
2086
2087         goto again;
2088 }
2089
2090 static void
2091 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2092 {
2093         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2094         int             i;
2095
2096         LASSERT (pool->po_allocated == 0);
2097
2098         if (tpo->tpo_tx_pages != NULL) {
2099                 kiblnd_unmap_tx_pool(tpo);
2100                 kiblnd_free_pages(tpo->tpo_tx_pages);
2101         }
2102
2103         if (tpo->tpo_tx_descs == NULL)
2104                 goto out;
2105
2106         for (i = 0; i < pool->po_size; i++) {
2107                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2108
2109                 list_del(&tx->tx_list);
2110                 if (tx->tx_pages != NULL)
2111                         LIBCFS_FREE(tx->tx_pages,
2112                                     LNET_MAX_IOV *
2113                                     sizeof(*tx->tx_pages));
2114                 if (tx->tx_frags != NULL)
2115                         LIBCFS_FREE(tx->tx_frags,
2116                                     IBLND_MAX_RDMA_FRAGS *
2117                                             sizeof(*tx->tx_frags));
2118                 if (tx->tx_wrq != NULL)
2119                         LIBCFS_FREE(tx->tx_wrq,
2120                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2121                                     sizeof(*tx->tx_wrq));
2122                 if (tx->tx_sge != NULL)
2123                         LIBCFS_FREE(tx->tx_sge,
2124                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2125                                     sizeof(*tx->tx_sge));
2126                 if (tx->tx_rd != NULL)
2127                         LIBCFS_FREE(tx->tx_rd,
2128                                     offsetof(kib_rdma_desc_t,
2129                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2130         }
2131
2132         LIBCFS_FREE(tpo->tpo_tx_descs,
2133                     pool->po_size * sizeof(kib_tx_t));
2134 out:
2135         kiblnd_fini_pool(pool);
2136         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2137 }
2138
2139 static int kiblnd_tx_pool_size(int ncpts)
2140 {
2141         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2142
2143         return max(IBLND_TX_POOL, ntx);
2144 }
2145
2146 static int
2147 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2148 {
2149         int            i;
2150         int            npg;
2151         kib_pool_t    *pool;
2152         kib_tx_pool_t *tpo;
2153
2154         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2155         if (tpo == NULL) {
2156                 CERROR("Failed to allocate TX pool\n");
2157                 return -ENOMEM;
2158         }
2159
2160         pool = &tpo->tpo_pool;
2161         kiblnd_init_pool(ps, pool, size);
2162         tpo->tpo_tx_descs = NULL;
2163         tpo->tpo_tx_pages = NULL;
2164
2165         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2166         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2167                 CERROR("Can't allocate tx pages: %d\n", npg);
2168                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2169                 return -ENOMEM;
2170         }
2171
2172         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2173                          size * sizeof(kib_tx_t));
2174         if (tpo->tpo_tx_descs == NULL) {
2175                 CERROR("Can't allocate %d tx descriptors\n", size);
2176                 ps->ps_pool_destroy(pool);
2177                 return -ENOMEM;
2178         }
2179
2180         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2181
2182         for (i = 0; i < size; i++) {
2183                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2184
2185                 tx->tx_pool = tpo;
2186                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2187                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2188                                          lnet_cpt_table(), ps->ps_cpt,
2189                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2190                         if (tx->tx_pages == NULL)
2191                                 break;
2192                 }
2193
2194                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2195                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2196                 if (tx->tx_frags == NULL)
2197                         break;
2198
2199                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2200
2201                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2202                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2203                                  sizeof(*tx->tx_wrq));
2204                 if (tx->tx_wrq == NULL)
2205                         break;
2206
2207                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2208                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2209                                  sizeof(*tx->tx_sge));
2210                 if (tx->tx_sge == NULL)
2211                         break;
2212
2213                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2214                                  offsetof(kib_rdma_desc_t,
2215                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2216                 if (tx->tx_rd == NULL)
2217                         break;
2218         }
2219
2220         if (i == size) {
2221                 kiblnd_map_tx_pool(tpo);
2222                 *pp_po = pool;
2223                 return 0;
2224         }
2225
2226         ps->ps_pool_destroy(pool);
2227         return -ENOMEM;
2228 }
2229
2230 static void
2231 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2232 {
2233         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2234                                              tps_poolset);
2235         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2236
2237         tx->tx_cookie = tps->tps_next_tx_cookie++;
2238 }
2239
2240 static void
2241 kiblnd_net_fini_pools(kib_net_t *net)
2242 {
2243         int     i;
2244
2245         cfs_cpt_for_each(i, lnet_cpt_table()) {
2246                 kib_tx_poolset_t        *tps;
2247                 kib_fmr_poolset_t       *fps;
2248
2249                 if (net->ibn_tx_ps != NULL) {
2250                         tps = net->ibn_tx_ps[i];
2251                         kiblnd_fini_poolset(&tps->tps_poolset);
2252                 }
2253
2254                 if (net->ibn_fmr_ps != NULL) {
2255                         fps = net->ibn_fmr_ps[i];
2256                         kiblnd_fini_fmr_poolset(fps);
2257                 }
2258         }
2259
2260         if (net->ibn_tx_ps != NULL) {
2261                 cfs_percpt_free(net->ibn_tx_ps);
2262                 net->ibn_tx_ps = NULL;
2263         }
2264
2265         if (net->ibn_fmr_ps != NULL) {
2266                 cfs_percpt_free(net->ibn_fmr_ps);
2267                 net->ibn_fmr_ps = NULL;
2268         }
2269 }
2270
2271 static int
2272 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2273 {
2274         unsigned long   flags;
2275         int             cpt;
2276         int             rc;
2277         int             i;
2278
2279         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2280         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2281                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2282                                            flags);
2283                 goto create_tx_pool;
2284         }
2285
2286         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2287
2288         if (*kiblnd_tunables.kib_fmr_pool_size <
2289             *kiblnd_tunables.kib_ntx / 4) {
2290                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2291                        *kiblnd_tunables.kib_fmr_pool_size,
2292                        *kiblnd_tunables.kib_ntx / 4);
2293                 rc = -EINVAL;
2294                 goto failed;
2295         }
2296
2297         /* TX pool must be created later than FMR, see LU-2268
2298          * for details */
2299         LASSERT(net->ibn_tx_ps == NULL);
2300
2301         /* premapping can fail if ibd_nmr > 1, so we always create
2302          * FMR pool and map-on-demand if premapping failed */
2303
2304         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2305                                            sizeof(kib_fmr_poolset_t));
2306         if (net->ibn_fmr_ps == NULL) {
2307                 CERROR("Failed to allocate FMR pool array\n");
2308                 rc = -ENOMEM;
2309                 goto failed;
2310         }
2311
2312         for (i = 0; i < ncpts; i++) {
2313                 cpt = (cpts == NULL) ? i : cpts[i];
2314                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2315                                              kiblnd_fmr_pool_size(ncpts),
2316                                              kiblnd_fmr_flush_trigger(ncpts));
2317                 if (rc != 0) {
2318                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2319                                cpt, rc);
2320                         goto failed;
2321                 }
2322         }
2323
2324         if (i > 0)
2325                 LASSERT(i == ncpts);
2326
2327  create_tx_pool:
2328         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2329                                           sizeof(kib_tx_poolset_t));
2330         if (net->ibn_tx_ps == NULL) {
2331                 CERROR("Failed to allocate tx pool array\n");
2332                 rc = -ENOMEM;
2333                 goto failed;
2334         }
2335
2336         for (i = 0; i < ncpts; i++) {
2337                 cpt = (cpts == NULL) ? i : cpts[i];
2338                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2339                                          cpt, net, "TX",
2340                                          kiblnd_tx_pool_size(ncpts),
2341                                          kiblnd_create_tx_pool,
2342                                          kiblnd_destroy_tx_pool,
2343                                          kiblnd_tx_init, NULL);
2344                 if (rc != 0) {
2345                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2346                                cpt, rc);
2347                         goto failed;
2348                 }
2349         }
2350
2351         return 0;
2352  failed:
2353         kiblnd_net_fini_pools(net);
2354         LASSERT(rc != 0);
2355         return rc;
2356 }
2357
2358 static int
2359 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2360 {
2361         struct ib_device_attr *attr;
2362         int                    rc;
2363
2364         /* It's safe to assume a HCA can handle a page size
2365          * matching that of the native system */
2366         hdev->ibh_page_shift = PAGE_SHIFT;
2367         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2368         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2369
2370         LIBCFS_ALLOC(attr, sizeof(*attr));
2371         if (attr == NULL) {
2372                 CERROR("Out of memory\n");
2373                 return -ENOMEM;
2374         }
2375
2376         rc = ib_query_device(hdev->ibh_ibdev, attr);
2377         if (rc == 0)
2378                 hdev->ibh_mr_size = attr->max_mr_size;
2379
2380         LIBCFS_FREE(attr, sizeof(*attr));
2381
2382         if (rc != 0) {
2383                 CERROR("Failed to query IB device: %d\n", rc);
2384                 return rc;
2385         }
2386
2387         if (hdev->ibh_mr_size == ~0ULL) {
2388                 hdev->ibh_mr_shift = 64;
2389                 return 0;
2390         }
2391
2392         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2393         return -EINVAL;
2394 }
2395
2396 static void
2397 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2398 {
2399         if (hdev->ibh_mrs == NULL)
2400                 return;
2401
2402         ib_dereg_mr(hdev->ibh_mrs);
2403
2404         hdev->ibh_mrs = NULL;
2405 }
2406
2407 void
2408 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2409 {
2410         kiblnd_hdev_cleanup_mrs(hdev);
2411
2412         if (hdev->ibh_pd != NULL)
2413                 ib_dealloc_pd(hdev->ibh_pd);
2414
2415         if (hdev->ibh_cmid != NULL)
2416                 rdma_destroy_id(hdev->ibh_cmid);
2417
2418         LIBCFS_FREE(hdev, sizeof(*hdev));
2419 }
2420
2421 static int
2422 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2423 {
2424         struct ib_mr *mr;
2425         int           rc;
2426         int           acflags = IB_ACCESS_LOCAL_WRITE |
2427                                 IB_ACCESS_REMOTE_WRITE;
2428
2429         rc = kiblnd_hdev_get_attr(hdev);
2430         if (rc != 0)
2431                 return rc;
2432
2433         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2434         if (IS_ERR(mr)) {
2435                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2436                 kiblnd_hdev_cleanup_mrs(hdev);
2437                 return PTR_ERR(mr);
2438         }
2439
2440         hdev->ibh_mrs = mr;
2441
2442         return 0;
2443 }
2444
2445 static int
2446 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2447 {       /* DUMMY */
2448         return 0;
2449 }
2450
2451 static int
2452 kiblnd_dev_need_failover(kib_dev_t *dev)
2453 {
2454         struct rdma_cm_id  *cmid;
2455         struct sockaddr_in  srcaddr;
2456         struct sockaddr_in  dstaddr;
2457         int                 rc;
2458
2459         if (dev->ibd_hdev == NULL || /* initializing */
2460             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2461             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2462                 return 1;
2463
2464         /* XXX: it's UGLY, but I don't have better way to find
2465          * ib-bonding HCA failover because:
2466          *
2467          * a. no reliable CM event for HCA failover...
2468          * b. no OFED API to get ib_device for current net_device...
2469          *
2470          * We have only two choices at this point:
2471          *
2472          * a. rdma_bind_addr(), it will conflict with listener cmid
2473          * b. rdma_resolve_addr() to zero addr */
2474         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2475                                      IB_QPT_RC);
2476         if (IS_ERR(cmid)) {
2477                 rc = PTR_ERR(cmid);
2478                 CERROR("Failed to create cmid for failover: %d\n", rc);
2479                 return rc;
2480         }
2481
2482         memset(&srcaddr, 0, sizeof(srcaddr));
2483         srcaddr.sin_family      = AF_INET;
2484         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2485
2486         memset(&dstaddr, 0, sizeof(dstaddr));
2487         dstaddr.sin_family = AF_INET;
2488         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2489                                (struct sockaddr *)&dstaddr, 1);
2490         if (rc != 0 || cmid->device == NULL) {
2491                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2492                        dev->ibd_ifname, &dev->ibd_ifip,
2493                        cmid->device, rc);
2494                 rdma_destroy_id(cmid);
2495                 return rc;
2496         }
2497
2498         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2499         rdma_destroy_id(cmid);
2500         return rc;
2501 }
2502
2503 int
2504 kiblnd_dev_failover(kib_dev_t *dev)
2505 {
2506         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2507         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2508         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2509         struct rdma_cm_id  *cmid  = NULL;
2510         kib_hca_dev_t      *hdev  = NULL;
2511         kib_hca_dev_t      *old;
2512         struct ib_pd       *pd;
2513         kib_net_t          *net;
2514         struct sockaddr_in  addr;
2515         unsigned long       flags;
2516         int                 rc = 0;
2517         int                 i;
2518
2519         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2520                  dev->ibd_can_failover ||
2521                  dev->ibd_hdev == NULL);
2522
2523         rc = kiblnd_dev_need_failover(dev);
2524         if (rc <= 0)
2525                 goto out;
2526
2527         if (dev->ibd_hdev != NULL &&
2528             dev->ibd_hdev->ibh_cmid != NULL) {
2529                 /* XXX it's not good to close old listener at here,
2530                  * because we can fail to create new listener.
2531                  * But we have to close it now, otherwise rdma_bind_addr
2532                  * will return EADDRINUSE... How crap! */
2533                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2534
2535                 cmid = dev->ibd_hdev->ibh_cmid;
2536                 /* make next schedule of kiblnd_dev_need_failover()
2537                  * return 1 for me */
2538                 dev->ibd_hdev->ibh_cmid  = NULL;
2539                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2540
2541                 rdma_destroy_id(cmid);
2542         }
2543
2544         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2545                                      IB_QPT_RC);
2546         if (IS_ERR(cmid)) {
2547                 rc = PTR_ERR(cmid);
2548                 CERROR("Failed to create cmid for failover: %d\n", rc);
2549                 goto out;
2550         }
2551
2552         memset(&addr, 0, sizeof(addr));
2553         addr.sin_family      = AF_INET;
2554         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2555         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2556
2557         /* Bind to failover device or port */
2558         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2559         if (rc != 0 || cmid->device == NULL) {
2560                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2561                        dev->ibd_ifname, &dev->ibd_ifip,
2562                        cmid->device, rc);
2563                 rdma_destroy_id(cmid);
2564                 goto out;
2565         }
2566
2567         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2568         if (hdev == NULL) {
2569                 CERROR("Failed to allocate kib_hca_dev\n");
2570                 rdma_destroy_id(cmid);
2571                 rc = -ENOMEM;
2572                 goto out;
2573         }
2574
2575         atomic_set(&hdev->ibh_ref, 1);
2576         hdev->ibh_dev   = dev;
2577         hdev->ibh_cmid  = cmid;
2578         hdev->ibh_ibdev = cmid->device;
2579
2580         pd = ib_alloc_pd(cmid->device);
2581         if (IS_ERR(pd)) {
2582                 rc = PTR_ERR(pd);
2583                 CERROR("Can't allocate PD: %d\n", rc);
2584                 goto out;
2585         }
2586
2587         hdev->ibh_pd = pd;
2588
2589         rc = rdma_listen(cmid, 0);
2590         if (rc != 0) {
2591                 CERROR("Can't start new listener: %d\n", rc);
2592                 goto out;
2593         }
2594
2595         rc = kiblnd_hdev_setup_mrs(hdev);
2596         if (rc != 0) {
2597                 CERROR("Can't setup device: %d\n", rc);
2598                 goto out;
2599         }
2600
2601         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2602
2603         old = dev->ibd_hdev;
2604         dev->ibd_hdev = hdev;   /* take over the refcount */
2605         hdev = old;
2606
2607         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2608                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2609                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2610                                             &zombie_tpo);
2611
2612                         if (net->ibn_fmr_ps != NULL)
2613                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2614                                                         &zombie_fpo);
2615                 }
2616         }
2617
2618         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2619  out:
2620         if (!list_empty(&zombie_tpo))
2621                 kiblnd_destroy_pool_list(&zombie_tpo);
2622         if (!list_empty(&zombie_ppo))
2623                 kiblnd_destroy_pool_list(&zombie_ppo);
2624         if (!list_empty(&zombie_fpo))
2625                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2626         if (hdev != NULL)
2627                 kiblnd_hdev_decref(hdev);
2628
2629         if (rc != 0)
2630                 dev->ibd_failed_failover++;
2631         else
2632                 dev->ibd_failed_failover = 0;
2633
2634         return rc;
2635 }
2636
2637 void
2638 kiblnd_destroy_dev (kib_dev_t *dev)
2639 {
2640         LASSERT (dev->ibd_nnets == 0);
2641         LASSERT(list_empty(&dev->ibd_nets));
2642
2643         list_del(&dev->ibd_fail_list);
2644         list_del(&dev->ibd_list);
2645
2646         if (dev->ibd_hdev != NULL)
2647                 kiblnd_hdev_decref(dev->ibd_hdev);
2648
2649         LIBCFS_FREE(dev, sizeof(*dev));
2650 }
2651
2652 static kib_dev_t *
2653 kiblnd_create_dev(char *ifname)
2654 {
2655         struct net_device *netdev;
2656         kib_dev_t         *dev;
2657         __u32              netmask;
2658         __u32              ip;
2659         int                up;
2660         int                rc;
2661
2662         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2663         if (rc != 0) {
2664                 CERROR("Can't query IPoIB interface %s: %d\n",
2665                        ifname, rc);
2666                 return NULL;
2667         }
2668
2669         if (!up) {
2670                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2671                 return NULL;
2672         }
2673
2674         LIBCFS_ALLOC(dev, sizeof(*dev));
2675         if (dev == NULL)
2676                 return NULL;
2677
2678         memset(dev, 0, sizeof(*dev));
2679         netdev = dev_get_by_name(&init_net, ifname);
2680         if (netdev == NULL) {
2681                 dev->ibd_can_failover = 0;
2682         } else {
2683                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2684                 dev_put(netdev);
2685         }
2686
2687         INIT_LIST_HEAD(&dev->ibd_nets);
2688         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2689         INIT_LIST_HEAD(&dev->ibd_fail_list);
2690         dev->ibd_ifip = ip;
2691         strcpy(&dev->ibd_ifname[0], ifname);
2692
2693         /* initialize the device */
2694         rc = kiblnd_dev_failover(dev);
2695         if (rc != 0) {
2696                 CERROR("Can't initialize device: %d\n", rc);
2697                 LIBCFS_FREE(dev, sizeof(*dev));
2698                 return NULL;
2699         }
2700
2701         list_add_tail(&dev->ibd_list,
2702                           &kiblnd_data.kib_devs);
2703         return dev;
2704 }
2705
2706 static void
2707 kiblnd_base_shutdown(void)
2708 {
2709         struct kib_sched_info   *sched;
2710         int                     i;
2711
2712         LASSERT(list_empty(&kiblnd_data.kib_devs));
2713
2714         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2715                atomic_read(&libcfs_kmemory));
2716
2717         switch (kiblnd_data.kib_init) {
2718         default:
2719                 LBUG();
2720
2721         case IBLND_INIT_ALL:
2722         case IBLND_INIT_DATA:
2723                 LASSERT (kiblnd_data.kib_peers != NULL);
2724                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2725                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2726                 }
2727                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2728                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2729                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2730                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2731
2732                 /* flag threads to terminate; wake and wait for them to die */
2733                 kiblnd_data.kib_shutdown = 1;
2734
2735                 /* NB: we really want to stop scheduler threads net by net
2736                  * instead of the whole module, this should be improved
2737                  * with dynamic configuration LNet */
2738                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2739                         wake_up_all(&sched->ibs_waitq);
2740
2741                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2742                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2743
2744                 i = 2;
2745                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2746                         i++;
2747                         /* power of 2? */
2748                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2749                                "Waiting for %d threads to terminate\n",
2750                                atomic_read(&kiblnd_data.kib_nthreads));
2751                         set_current_state(TASK_UNINTERRUPTIBLE);
2752                         schedule_timeout(cfs_time_seconds(1));
2753                 }
2754
2755                 /* fall through */
2756
2757         case IBLND_INIT_NOTHING:
2758                 break;
2759         }
2760
2761         if (kiblnd_data.kib_peers != NULL) {
2762                 LIBCFS_FREE(kiblnd_data.kib_peers,
2763                             sizeof(struct list_head) *
2764                             kiblnd_data.kib_peer_hash_size);
2765         }
2766
2767         if (kiblnd_data.kib_scheds != NULL)
2768                 cfs_percpt_free(kiblnd_data.kib_scheds);
2769
2770         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2771                atomic_read(&libcfs_kmemory));
2772
2773         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2774         module_put(THIS_MODULE);
2775 }
2776
2777 static void
2778 kiblnd_shutdown (lnet_ni_t *ni)
2779 {
2780         kib_net_t        *net = ni->ni_data;
2781         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2782         int               i;
2783         unsigned long     flags;
2784
2785         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2786
2787         if (net == NULL)
2788                 goto out;
2789
2790         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2791                atomic_read(&libcfs_kmemory));
2792
2793         write_lock_irqsave(g_lock, flags);
2794         net->ibn_shutdown = 1;
2795         write_unlock_irqrestore(g_lock, flags);
2796
2797         switch (net->ibn_init) {
2798         default:
2799                 LBUG();
2800
2801         case IBLND_INIT_ALL:
2802                 /* nuke all existing peers within this net */
2803                 kiblnd_del_peer(ni, LNET_NID_ANY);
2804
2805                 /* Wait for all peer state to clean up */
2806                 i = 2;
2807                 while (atomic_read(&net->ibn_npeers) != 0) {
2808                         i++;
2809                         /* power of 2? */
2810                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2811                                "%s: waiting for %d peers to disconnect\n",
2812                                libcfs_nid2str(ni->ni_nid),
2813                                atomic_read(&net->ibn_npeers));
2814                         set_current_state(TASK_UNINTERRUPTIBLE);
2815                         schedule_timeout(cfs_time_seconds(1));
2816                 }
2817
2818                 kiblnd_net_fini_pools(net);
2819
2820                 write_lock_irqsave(g_lock, flags);
2821                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2822                 net->ibn_dev->ibd_nnets--;
2823                 list_del(&net->ibn_list);
2824                 write_unlock_irqrestore(g_lock, flags);
2825
2826                 /* fall through */
2827
2828         case IBLND_INIT_NOTHING:
2829                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2830
2831                 if (net->ibn_dev != NULL &&
2832                     net->ibn_dev->ibd_nnets == 0)
2833                         kiblnd_destroy_dev(net->ibn_dev);
2834
2835                 break;
2836         }
2837
2838         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2839                atomic_read(&libcfs_kmemory));
2840
2841         net->ibn_init = IBLND_INIT_NOTHING;
2842         ni->ni_data = NULL;
2843
2844         LIBCFS_FREE(net, sizeof(*net));
2845
2846 out:
2847         if (list_empty(&kiblnd_data.kib_devs))
2848                 kiblnd_base_shutdown();
2849         return;
2850 }
2851
2852 static int
2853 kiblnd_base_startup(void)
2854 {
2855         struct kib_sched_info   *sched;
2856         int                     rc;
2857         int                     i;
2858
2859         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2860
2861         try_module_get(THIS_MODULE);
2862         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2863
2864         rwlock_init(&kiblnd_data.kib_global_lock);
2865
2866         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2867         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2868
2869         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2870         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2871                      sizeof(struct list_head) *
2872                      kiblnd_data.kib_peer_hash_size);
2873         if (kiblnd_data.kib_peers == NULL)
2874                 goto failed;
2875
2876         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2877                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2878
2879         spin_lock_init(&kiblnd_data.kib_connd_lock);
2880         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2881         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2882         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2883         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2884
2885         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2886         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2887
2888         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2889                                                   sizeof(*sched));
2890         if (kiblnd_data.kib_scheds == NULL)
2891                 goto failed;
2892
2893         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2894                 int     nthrs;
2895
2896                 spin_lock_init(&sched->ibs_lock);
2897                 INIT_LIST_HEAD(&sched->ibs_conns);
2898                 init_waitqueue_head(&sched->ibs_waitq);
2899
2900                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2901                 if (*kiblnd_tunables.kib_nscheds > 0) {
2902                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2903                 } else {
2904                         /* max to half of CPUs, another half is reserved for
2905                          * upper layer modules */
2906                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2907                 }
2908
2909                 sched->ibs_nthreads_max = nthrs;
2910                 sched->ibs_cpt = i;
2911         }
2912
2913         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2914
2915         /* lists/ptrs/locks initialised */
2916         kiblnd_data.kib_init = IBLND_INIT_DATA;
2917         /*****************************************************/
2918
2919         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2920         if (rc != 0) {
2921                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2922                 goto failed;
2923         }
2924
2925         if (*kiblnd_tunables.kib_dev_failover != 0)
2926                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2927                                          "kiblnd_failover");
2928
2929         if (rc != 0) {
2930                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2931                 goto failed;
2932         }
2933
2934         /* flag everything initialised */
2935         kiblnd_data.kib_init = IBLND_INIT_ALL;
2936         /*****************************************************/
2937
2938         return 0;
2939
2940  failed:
2941         kiblnd_base_shutdown();
2942         return -ENETDOWN;
2943 }
2944
2945 static int
2946 kiblnd_start_schedulers(struct kib_sched_info *sched)
2947 {
2948         int     rc = 0;
2949         int     nthrs;
2950         int     i;
2951
2952         if (sched->ibs_nthreads == 0) {
2953                 if (*kiblnd_tunables.kib_nscheds > 0) {
2954                         nthrs = sched->ibs_nthreads_max;
2955                 } else {
2956                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2957                                                sched->ibs_cpt);
2958                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2959                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2960                 }
2961         } else {
2962                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2963                 /* increase one thread if there is new interface */
2964                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2965         }
2966
2967         for (i = 0; i < nthrs; i++) {
2968                 long    id;
2969                 char    name[20];
2970                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2971                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2972                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2973                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2974                 if (rc == 0)
2975                         continue;
2976
2977                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2978                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2979                 break;
2980         }
2981
2982         sched->ibs_nthreads += i;
2983         return rc;
2984 }
2985
2986 static int
2987 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2988 {
2989         int     cpt;
2990         int     rc;
2991         int     i;
2992
2993         for (i = 0; i < ncpts; i++) {
2994                 struct kib_sched_info *sched;
2995
2996                 cpt = (cpts == NULL) ? i : cpts[i];
2997                 sched = kiblnd_data.kib_scheds[cpt];
2998
2999                 if (!newdev && sched->ibs_nthreads > 0)
3000                         continue;
3001
3002                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3003                 if (rc != 0) {
3004                         CERROR("Failed to start scheduler threads for %s\n",
3005                                dev->ibd_ifname);
3006                         return rc;
3007                 }
3008         }
3009         return 0;
3010 }
3011
3012 static kib_dev_t *
3013 kiblnd_dev_search(char *ifname)
3014 {
3015         kib_dev_t       *alias = NULL;
3016         kib_dev_t       *dev;
3017         char            *colon;
3018         char            *colon2;
3019
3020         colon = strchr(ifname, ':');
3021         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3022                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3023                         return dev;
3024
3025                 if (alias != NULL)
3026                         continue;
3027
3028                 colon2 = strchr(dev->ibd_ifname, ':');
3029                 if (colon != NULL)
3030                         *colon = 0;
3031                 if (colon2 != NULL)
3032                         *colon2 = 0;
3033
3034                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3035                         alias = dev;
3036
3037                 if (colon != NULL)
3038                         *colon = ':';
3039                 if (colon2 != NULL)
3040                         *colon2 = ':';
3041         }
3042         return alias;
3043 }
3044
3045 static int
3046 kiblnd_startup (lnet_ni_t *ni)
3047 {
3048         char                     *ifname;
3049         kib_dev_t                *ibdev = NULL;
3050         kib_net_t                *net;
3051         struct timeval            tv;
3052         unsigned long             flags;
3053         int                       rc;
3054         int                       newdev;
3055
3056         LASSERT (ni->ni_lnd == &the_o2iblnd);
3057
3058         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3059                 rc = kiblnd_base_startup();
3060                 if (rc != 0)
3061                         return rc;
3062         }
3063
3064         LIBCFS_ALLOC(net, sizeof(*net));
3065         ni->ni_data = net;
3066         if (net == NULL)
3067                 goto failed;
3068
3069         memset(net, 0, sizeof(*net));
3070
3071         do_gettimeofday(&tv);
3072         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3073
3074         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3075         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3076         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3077         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3078
3079         if (ni->ni_interfaces[0] != NULL) {
3080                 /* Use the IPoIB interface specified in 'networks=' */
3081
3082                 CLASSERT (LNET_MAX_INTERFACES > 1);
3083                 if (ni->ni_interfaces[1] != NULL) {
3084                         CERROR("Multiple interfaces not supported\n");
3085                         goto failed;
3086                 }
3087
3088                 ifname = ni->ni_interfaces[0];
3089         } else {
3090                 ifname = *kiblnd_tunables.kib_default_ipif;
3091         }
3092
3093         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3094                 CERROR("IPoIB interface name too long: %s\n", ifname);
3095                 goto failed;
3096         }
3097
3098         ibdev = kiblnd_dev_search(ifname);
3099
3100         newdev = ibdev == NULL;
3101         /* hmm...create kib_dev even for alias */
3102         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3103                 ibdev = kiblnd_create_dev(ifname);
3104
3105         if (ibdev == NULL)
3106                 goto failed;
3107
3108         net->ibn_dev = ibdev;
3109         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3110
3111         rc = kiblnd_dev_start_threads(ibdev, newdev,
3112                                       ni->ni_cpts, ni->ni_ncpts);
3113         if (rc != 0)
3114                 goto failed;
3115
3116         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3117         if (rc != 0) {
3118                 CERROR("Failed to initialize NI pools: %d\n", rc);
3119                 goto failed;
3120         }
3121
3122         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3123         ibdev->ibd_nnets++;
3124         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3125         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3126
3127         net->ibn_init = IBLND_INIT_ALL;
3128
3129         return 0;
3130
3131 failed:
3132         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3133                 kiblnd_destroy_dev(ibdev);
3134
3135         kiblnd_shutdown(ni);
3136
3137         CDEBUG(D_NET, "kiblnd_startup failed\n");
3138         return -ENETDOWN;
3139 }
3140
3141 static lnd_t the_o2iblnd = {
3142         .lnd_type       = O2IBLND,
3143         .lnd_startup    = kiblnd_startup,
3144         .lnd_shutdown   = kiblnd_shutdown,
3145         .lnd_ctl        = kiblnd_ctl,
3146         .lnd_query      = kiblnd_query,
3147         .lnd_send       = kiblnd_send,
3148         .lnd_recv       = kiblnd_recv,
3149 };
3150
3151 static void __exit ko2iblnd_exit(void)
3152 {
3153         lnet_unregister_lnd(&the_o2iblnd);
3154         kiblnd_tunables_fini();
3155 }
3156
3157 static int __init ko2iblnd_init(void)
3158 {
3159         int rc;
3160
3161         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3162         CLASSERT(offsetof(kib_msg_t,
3163                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3164                  IBLND_MSG_SIZE);
3165         CLASSERT(offsetof(kib_msg_t,
3166                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3167                  <= IBLND_MSG_SIZE);
3168
3169         rc = kiblnd_tunables_init();
3170         if (rc != 0)
3171                 return rc;
3172
3173         lnet_register_lnd(&the_o2iblnd);
3174
3175         return 0;
3176 }
3177
3178 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3179 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3180 MODULE_VERSION("2.8.0");
3181 MODULE_LICENSE("GPL");
3182
3183 module_init(ko2iblnd_init);
3184 module_exit(ko2iblnd_exit);