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 = kiblnd_cfg_rdma_frags(peer->ibp_ni);
 341         peer->ibp_queue_depth = ni->ni_peertxcredits;
 342         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 343
 344         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 345         INIT_LIST_HEAD(&peer->ibp_conns);
 346         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer (kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(kiblnd_peer_idle(peer));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 387         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head        *tmp;
 389         kib_peer_t              *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394                 LASSERT(!kiblnd_peer_idle(peer));
 395
 396                 if (peer->ibp_nid != nid)
 397                         continue;
 398
 399                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 400                        peer, libcfs_nid2str(nid),
 401                        atomic_read(&peer->ibp_refcount),
 402                        peer->ibp_version);
 403                 return peer;
 404         }
 405         return NULL;
 406 }
 407
 408 void
 409 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 410 {
 411         LASSERT(list_empty(&peer->ibp_conns));
 412
 413         LASSERT (kiblnd_peer_active(peer));
 414         list_del_init(&peer->ibp_list);
 415         /* lose peerlist's ref */
 416         kiblnd_peer_decref(peer);
 417 }
 418
 419 static int
 420 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 421                      lnet_nid_t *nidp, int *count)
 422 {
 423         kib_peer_t              *peer;
 424         struct list_head        *ptmp;
 425         int                      i;
 426         unsigned long            flags;
 427
 428         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 429
 430         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 431
 432                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 433
 434                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 435                         LASSERT(!kiblnd_peer_idle(peer));
 436
 437                         if (peer->ibp_ni != ni)
 438                                 continue;
 439
 440                         if (index-- > 0)
 441                                 continue;
 442
 443                         *nidp = peer->ibp_nid;
 444                         *count = atomic_read(&peer->ibp_refcount);
 445
 446                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 447                                                flags);
 448                         return 0;
 449                 }
 450         }
 451
 452         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 453         return -ENOENT;
 454 }
 455
 456 static void
 457 kiblnd_del_peer_locked (kib_peer_t *peer)
 458 {
 459         struct list_head        *ctmp;
 460         struct list_head        *cnxt;
 461         kib_conn_t              *conn;
 462
 463         if (list_empty(&peer->ibp_conns)) {
 464                 kiblnd_unlink_peer_locked(peer);
 465         } else {
 466                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 467                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 468
 469                         kiblnd_close_conn_locked(conn, 0);
 470                 }
 471                 /* NB closing peer's last conn unlinked it. */
 472         }
 473         /* NB peer now unlinked; might even be freed if the peer table had the
 474          * last ref on it. */
 475 }
 476
 477 static int
 478 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 479 {
 480         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 481         struct list_head        *ptmp;
 482         struct list_head        *pnxt;
 483         kib_peer_t              *peer;
 484         int                     lo;
 485         int                     hi;
 486         int                     i;
 487         unsigned long           flags;
 488         int                     rc = -ENOENT;
 489
 490         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 491
 492         if (nid != LNET_NID_ANY) {
 493                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 494         } else {
 495                 lo = 0;
 496                 hi = kiblnd_data.kib_peer_hash_size - 1;
 497         }
 498
 499         for (i = lo; i <= hi; i++) {
 500                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 501                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 502                         LASSERT(!kiblnd_peer_idle(peer));
 503
 504                         if (peer->ibp_ni != ni)
 505                                 continue;
 506
 507                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 508                                 continue;
 509
 510                         if (!list_empty(&peer->ibp_tx_queue)) {
 511                                 LASSERT(list_empty(&peer->ibp_conns));
 512
 513                                 list_splice_init(&peer->ibp_tx_queue,
 514                                                  &zombies);
 515                         }
 516
 517                         kiblnd_del_peer_locked(peer);
 518                         rc = 0;         /* matched something */
 519                 }
 520         }
 521
 522         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 523
 524         kiblnd_txlist_done(ni, &zombies, -EIO);
 525
 526         return rc;
 527 }
 528
 529 static kib_conn_t *
 530 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 531 {
 532         kib_peer_t              *peer;
 533         struct list_head        *ptmp;
 534         kib_conn_t              *conn;
 535         struct list_head        *ctmp;
 536         int                     i;
 537         unsigned long           flags;
 538
 539         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 540
 541         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 542                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 543
 544                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 545                         LASSERT(!kiblnd_peer_idle(peer));
 546
 547                         if (peer->ibp_ni != ni)
 548                                 continue;
 549
 550                         list_for_each(ctmp, &peer->ibp_conns) {
 551                                 if (index-- > 0)
 552                                         continue;
 553
 554                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 555                                 kiblnd_conn_addref(conn);
 556                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 557                                                        flags);
 558                                 return conn;
 559                         }
 560                 }
 561         }
 562
 563         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 564         return NULL;
 565 }
 566
 567 static void
 568 kiblnd_debug_rx (kib_rx_t *rx)
 569 {
 570         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 571                rx, rx->rx_status, rx->rx_msg->ibm_type,
 572                rx->rx_msg->ibm_credits);
 573 }
 574
 575 static void
 576 kiblnd_debug_tx (kib_tx_t *tx)
 577 {
 578         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 579                "cookie "LPX64" msg %s%s type %x cred %d\n",
 580                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 581                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 582                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 583                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 584                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 585 }
 586
 587 void
 588 kiblnd_debug_conn (kib_conn_t *conn)
 589 {
 590         struct list_head        *tmp;
 591         int                     i;
 592
 593         spin_lock(&conn->ibc_lock);
 594
 595         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 596                atomic_read(&conn->ibc_refcount), conn,
 597                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 598         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 599                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 600                conn->ibc_nsends_posted, conn->ibc_credits,
 601                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 602         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 603
 604         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 605         list_for_each(tmp, &conn->ibc_early_rxs)
 606                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 607
 608         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 609         list_for_each(tmp, &conn->ibc_tx_noops)
 610                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 611
 612         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 613         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 614                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 615
 616         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 617         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 618                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 621         list_for_each(tmp, &conn->ibc_tx_queue)
 622                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 623
 624         CDEBUG(D_CONSOLE, "   active_txs:\n");
 625         list_for_each(tmp, &conn->ibc_active_txs)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   rxs:\n");
 629         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 630                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 631
 632         spin_unlock(&conn->ibc_lock);
 633 }
 634
 635 int
 636 kiblnd_translate_mtu(int value)
 637 {
 638         switch (value) {
 639         default:
 640                 return -1;
 641         case 0:
 642                 return 0;
 643         case 256:
 644                 return IB_MTU_256;
 645         case 512:
 646                 return IB_MTU_512;
 647         case 1024:
 648                 return IB_MTU_1024;
 649         case 2048:
 650                 return IB_MTU_2048;
 651         case 4096:
 652                 return IB_MTU_4096;
 653         }
 654 }
 655
 656 static void
 657 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 658 {
 659         int           mtu;
 660
 661         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 662         if (cmid->route.path_rec == NULL)
 663                 return;
 664
 665         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 666         LASSERT (mtu >= 0);
 667         if (mtu != 0)
 668                 cmid->route.path_rec->mtu = mtu;
 669 }
 670
 671 static int
 672 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 673 {
 674         cpumask_t       *mask;
 675         int             vectors;
 676         int             off;
 677         int             i;
 678         lnet_nid_t      ibp_nid;
 679
 680         vectors = conn->ibc_cmid->device->num_comp_vectors;
 681         if (vectors <= 1)
 682                 return 0;
 683
 684         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 685
 686         /* hash NID to CPU id in this partition... */
 687         ibp_nid = conn->ibc_peer->ibp_nid;
 688         off = do_div(ibp_nid, cpumask_weight(mask));
 689         for_each_cpu(i, mask) {
 690                 if (off-- == 0)
 691                         return i % vectors;
 692         }
 693
 694         LBUG();
 695         return 1;
 696 }
 697
 698 kib_conn_t *
 699 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 700                    int state, int version)
 701 {
 702         /* CAVEAT EMPTOR:
 703          * If the new conn is created successfully it takes over the caller's
 704          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 705          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 706          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 707          * to destroy 'cmid' here since I'm called from the CM which still has
 708          * its ref on 'cmid'). */
 709         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 710         kib_net_t              *net = peer->ibp_ni->ni_data;
 711         kib_dev_t              *dev;
 712         struct ib_qp_init_attr *init_qp_attr;
 713         struct kib_sched_info   *sched;
 714 #ifdef HAVE_IB_CQ_INIT_ATTR
 715         struct ib_cq_init_attr  cq_attr = {};
 716 #endif
 717         kib_conn_t              *conn;
 718         struct ib_cq            *cq;
 719         unsigned long           flags;
 720         int                     cpt;
 721         int                     rc;
 722         int                     i;
 723
 724         LASSERT(net != NULL);
 725         LASSERT(!in_interrupt());
 726
 727         dev = net->ibn_dev;
 728
 729         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 730         sched = kiblnd_data.kib_scheds[cpt];
 731
 732         LASSERT(sched->ibs_nthreads > 0);
 733
 734         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 735                          sizeof(*init_qp_attr));
 736         if (init_qp_attr == NULL) {
 737                 CERROR("Can't allocate qp_attr for %s\n",
 738                        libcfs_nid2str(peer->ibp_nid));
 739                 goto failed_0;
 740         }
 741
 742         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 743         if (conn == NULL) {
 744                 CERROR("Can't allocate connection for %s\n",
 745                        libcfs_nid2str(peer->ibp_nid));
 746                 goto failed_1;
 747         }
 748
 749         conn->ibc_state = IBLND_CONN_INIT;
 750         conn->ibc_version = version;
 751         conn->ibc_peer = peer;                  /* I take the caller's ref */
 752         cmid->context = conn;                   /* for future CM callbacks */
 753         conn->ibc_cmid = cmid;
 754         conn->ibc_max_frags = peer->ibp_max_frags;
 755         conn->ibc_queue_depth = peer->ibp_queue_depth;
 756
 757         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 758         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 759         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 760         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 761         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 762         INIT_LIST_HEAD(&conn->ibc_active_txs);
 763         spin_lock_init(&conn->ibc_lock);
 764
 765         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 766                          sizeof(*conn->ibc_connvars));
 767         if (conn->ibc_connvars == NULL) {
 768                 CERROR("Can't allocate in-progress connection state\n");
 769                 goto failed_2;
 770         }
 771
 772         write_lock_irqsave(glock, flags);
 773         if (dev->ibd_failover) {
 774                 write_unlock_irqrestore(glock, flags);
 775                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 776                 goto failed_2;
 777         }
 778
 779         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 780                 /* wakeup failover thread and teardown connection */
 781                 if (kiblnd_dev_can_failover(dev)) {
 782                         list_add_tail(&dev->ibd_fail_list,
 783                                       &kiblnd_data.kib_failed_devs);
 784                         wake_up(&kiblnd_data.kib_failover_waitq);
 785                 }
 786
 787                 write_unlock_irqrestore(glock, flags);
 788                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 789                        cmid->device->name, dev->ibd_ifname);
 790                 goto failed_2;
 791         }
 792
 793         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 794         conn->ibc_hdev = dev->ibd_hdev;
 795
 796         kiblnd_setup_mtu_locked(cmid);
 797
 798         write_unlock_irqrestore(glock, flags);
 799
 800         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 801                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 802         if (conn->ibc_rxs == NULL) {
 803                 CERROR("Cannot allocate RX buffers\n");
 804                 goto failed_2;
 805         }
 806
 807         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 808                                 IBLND_RX_MSG_PAGES(conn));
 809         if (rc != 0)
 810                 goto failed_2;
 811
 812         kiblnd_map_rx_descs(conn);
 813
 814 #ifdef HAVE_IB_CQ_INIT_ATTR
 815         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 816         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 817         cq = ib_create_cq(cmid->device,
 818                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 819                           &cq_attr);
 820 #else
 821         cq = ib_create_cq(cmid->device,
 822                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 823                           IBLND_CQ_ENTRIES(conn),
 824                           kiblnd_get_completion_vector(conn, cpt));
 825 #endif
 826         if (IS_ERR(cq)) {
 827                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 828                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 829                 goto failed_2;
 830         }
 831
 832         conn->ibc_cq = cq;
 833
 834         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 835         if (rc != 0) {
 836                 CERROR("Can't request completion notification: %d\n", rc);
 837                 goto failed_2;
 838         }
 839
 840         init_qp_attr->event_handler = kiblnd_qp_event;
 841         init_qp_attr->qp_context = conn;
 842         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 843         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 844         init_qp_attr->cap.max_send_sge = 1;
 845         init_qp_attr->cap.max_recv_sge = 1;
 846         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 847         init_qp_attr->qp_type = IB_QPT_RC;
 848         init_qp_attr->send_cq = cq;
 849         init_qp_attr->recv_cq = cq;
 850
 851         conn->ibc_sched = sched;
 852
 853         do {
 854                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 855                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 856                         break;
 857
 858                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 859         } while (rc);
 860
 861         if (rc) {
 862                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 863                         rc, init_qp_attr->cap.max_send_wr,
 864                         init_qp_attr->cap.max_recv_wr);
 865                 goto failed_2;
 866         }
 867
 868         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 869                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 870                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 871
 872         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 873
 874         /* 1 ref for caller and each rxmsg */
 875         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 876         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 877
 878         /* post receives */
 879         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 880                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 881                 if (rc != 0) {
 882                         CERROR("Can't post rxmsg: %d\n", rc);
 883
 884                         /* Make posted receives complete */
 885                         kiblnd_abort_receives(conn);
 886
 887                         /* correct # of posted buffers
 888                          * NB locking needed now I'm racing with completion */
 889                         spin_lock_irqsave(&sched->ibs_lock, flags);
 890                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 891                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 892
 893                         /* cmid will be destroyed by CM(ofed) after cm_callback
 894                          * returned, so we can't refer it anymore
 895                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 896                         rdma_destroy_qp(conn->ibc_cmid);
 897                         conn->ibc_cmid = NULL;
 898
 899                         /* Drop my own and unused rxbuffer refcounts */
 900                         while (i++ <= IBLND_RX_MSGS(conn))
 901                                 kiblnd_conn_decref(conn);
 902
 903                         return NULL;
 904                 }
 905         }
 906
 907         /* Init successful! */
 908         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 909                  state == IBLND_CONN_PASSIVE_WAIT);
 910         conn->ibc_state = state;
 911
 912         /* 1 more conn */
 913         atomic_inc(&net->ibn_nconns);
 914         return conn;
 915
 916  failed_2:
 917         kiblnd_destroy_conn(conn, true);
 918  failed_1:
 919         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 920  failed_0:
 921         return NULL;
 922 }
 923
 924 void
 925 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 926 {
 927         struct rdma_cm_id *cmid = conn->ibc_cmid;
 928         kib_peer_t        *peer = conn->ibc_peer;
 929         int                rc;
 930
 931         LASSERT (!in_interrupt());
 932         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 933         LASSERT(list_empty(&conn->ibc_early_rxs));
 934         LASSERT(list_empty(&conn->ibc_tx_noops));
 935         LASSERT(list_empty(&conn->ibc_tx_queue));
 936         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 937         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 938         LASSERT(list_empty(&conn->ibc_active_txs));
 939         LASSERT (conn->ibc_noops_posted == 0);
 940         LASSERT (conn->ibc_nsends_posted == 0);
 941
 942         switch (conn->ibc_state) {
 943         default:
 944                 /* conn must be completely disengaged from the network */
 945                 LBUG();
 946
 947         case IBLND_CONN_DISCONNECTED:
 948                 /* connvars should have been freed already */
 949                 LASSERT (conn->ibc_connvars == NULL);
 950                 break;
 951
 952         case IBLND_CONN_INIT:
 953                 break;
 954         }
 955
 956         /* conn->ibc_cmid might be destroyed by CM already */
 957         if (cmid != NULL && cmid->qp != NULL)
 958                 rdma_destroy_qp(cmid);
 959
 960         if (conn->ibc_cq != NULL) {
 961                 rc = ib_destroy_cq(conn->ibc_cq);
 962                 if (rc != 0)
 963                         CWARN("Error destroying CQ: %d\n", rc);
 964         }
 965
 966         if (conn->ibc_rx_pages != NULL)
 967                 kiblnd_unmap_rx_descs(conn);
 968
 969         if (conn->ibc_rxs != NULL) {
 970                 LIBCFS_FREE(conn->ibc_rxs,
 971                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 972         }
 973
 974         if (conn->ibc_connvars != NULL)
 975                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 976
 977         if (conn->ibc_hdev != NULL)
 978                 kiblnd_hdev_decref(conn->ibc_hdev);
 979
 980         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 981         if (conn->ibc_state != IBLND_CONN_INIT) {
 982                 kib_net_t *net = peer->ibp_ni->ni_data;
 983
 984                 kiblnd_peer_decref(peer);
 985                 rdma_destroy_id(cmid);
 986                 atomic_dec(&net->ibn_nconns);
 987         }
 988
 989         if (free_conn)
 990                 LIBCFS_FREE(conn, sizeof(*conn));
 991 }
 992
 993 int
 994 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 995 {
 996         kib_conn_t              *conn;
 997         struct list_head        *ctmp;
 998         struct list_head        *cnxt;
 999         int                     count = 0;
1000
1001         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1002                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1003
1004                 CDEBUG(D_NET, "Closing conn -> %s, "
1005                               "version: %x, reason: %d\n",
1006                        libcfs_nid2str(peer->ibp_nid),
1007                        conn->ibc_version, why);
1008
1009                 kiblnd_close_conn_locked(conn, why);
1010                 count++;
1011         }
1012
1013         return count;
1014 }
1015
1016 int
1017 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1018                                 int version, __u64 incarnation)
1019 {
1020         kib_conn_t              *conn;
1021         struct list_head        *ctmp;
1022         struct list_head        *cnxt;
1023         int                     count = 0;
1024
1025         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1026                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1027
1028                 if (conn->ibc_version     == version &&
1029                     conn->ibc_incarnation == incarnation)
1030                         continue;
1031
1032                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1033                               "incarnation:"LPX64"(%x, "LPX64")\n",
1034                        libcfs_nid2str(peer->ibp_nid),
1035                        conn->ibc_version, conn->ibc_incarnation,
1036                        version, incarnation);
1037
1038                 kiblnd_close_conn_locked(conn, -ESTALE);
1039                 count++;
1040         }
1041
1042         return count;
1043 }
1044
1045 static int
1046 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1047 {
1048         kib_peer_t              *peer;
1049         struct list_head        *ptmp;
1050         struct list_head        *pnxt;
1051         int                     lo;
1052         int                     hi;
1053         int                     i;
1054         unsigned long           flags;
1055         int                     count = 0;
1056
1057         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1058
1059         if (nid != LNET_NID_ANY)
1060                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1061         else {
1062                 lo = 0;
1063                 hi = kiblnd_data.kib_peer_hash_size - 1;
1064         }
1065
1066         for (i = lo; i <= hi; i++) {
1067                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1068
1069                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1070                         LASSERT(!kiblnd_peer_idle(peer));
1071
1072                         if (peer->ibp_ni != ni)
1073                                 continue;
1074
1075                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1076                                 continue;
1077
1078                         count += kiblnd_close_peer_conns_locked(peer, 0);
1079                 }
1080         }
1081
1082         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1083
1084         /* wildcards always succeed */
1085         if (nid == LNET_NID_ANY)
1086                 return 0;
1087
1088         return (count == 0) ? -ENOENT : 0;
1089 }
1090
1091 static int
1092 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1093 {
1094         struct libcfs_ioctl_data *data = arg;
1095         int                       rc = -EINVAL;
1096
1097         switch(cmd) {
1098         case IOC_LIBCFS_GET_PEER: {
1099                 lnet_nid_t   nid = 0;
1100                 int          count = 0;
1101
1102                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1103                                           &nid, &count);
1104                 data->ioc_nid    = nid;
1105                 data->ioc_count  = count;
1106                 break;
1107         }
1108
1109         case IOC_LIBCFS_DEL_PEER: {
1110                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1111                 break;
1112         }
1113         case IOC_LIBCFS_GET_CONN: {
1114                 kib_conn_t *conn;
1115
1116                 rc = 0;
1117                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1118                 if (conn == NULL) {
1119                         rc = -ENOENT;
1120                         break;
1121                 }
1122
1123                 LASSERT (conn->ibc_cmid != NULL);
1124                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1125                 if (conn->ibc_cmid->route.path_rec == NULL)
1126                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1127                 else
1128                         data->ioc_u32[0] =
1129                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1130                 kiblnd_conn_decref(conn);
1131                 break;
1132         }
1133         case IOC_LIBCFS_CLOSE_CONNECTION: {
1134                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1135                 break;
1136         }
1137
1138         default:
1139                 break;
1140         }
1141
1142         return rc;
1143 }
1144
1145 static void
1146 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1147 {
1148         cfs_time_t      last_alive = 0;
1149         cfs_time_t      now = cfs_time_current();
1150         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1151         kib_peer_t      *peer;
1152         unsigned long   flags;
1153
1154         read_lock_irqsave(glock, flags);
1155
1156         peer = kiblnd_find_peer_locked(nid);
1157         if (peer != NULL)
1158                 last_alive = peer->ibp_last_alive;
1159
1160         read_unlock_irqrestore(glock, flags);
1161
1162         if (last_alive != 0)
1163                 *when = last_alive;
1164
1165         /* peer is not persistent in hash, trigger peer creation
1166          * and connection establishment with a NULL tx */
1167         if (peer == NULL)
1168                 kiblnd_launch_tx(ni, NULL, nid);
1169
1170         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1171                libcfs_nid2str(nid), peer,
1172                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1173         return;
1174 }
1175
1176 static void
1177 kiblnd_free_pages(kib_pages_t *p)
1178 {
1179         int     npages = p->ibp_npages;
1180         int     i;
1181
1182         for (i = 0; i < npages; i++) {
1183                 if (p->ibp_pages[i] != NULL)
1184                         __free_page(p->ibp_pages[i]);
1185         }
1186
1187         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1188 }
1189
1190 int
1191 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1192 {
1193         kib_pages_t     *p;
1194         int             i;
1195
1196         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1197                          offsetof(kib_pages_t, ibp_pages[npages]));
1198         if (p == NULL) {
1199                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1200                 return -ENOMEM;
1201         }
1202
1203         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1204         p->ibp_npages = npages;
1205
1206         for (i = 0; i < npages; i++) {
1207                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1208                                                      GFP_NOFS);
1209                 if (p->ibp_pages[i] == NULL) {
1210                         CERROR("Can't allocate page %d of %d\n", i, npages);
1211                         kiblnd_free_pages(p);
1212                         return -ENOMEM;
1213                 }
1214         }
1215
1216         *pp = p;
1217         return 0;
1218 }
1219
1220 void
1221 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1222 {
1223         kib_rx_t *rx;
1224         int       i;
1225
1226         LASSERT (conn->ibc_rxs != NULL);
1227         LASSERT (conn->ibc_hdev != NULL);
1228
1229         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1230                 rx = &conn->ibc_rxs[i];
1231
1232                 LASSERT(rx->rx_nob >= 0); /* not posted */
1233
1234                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1235                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1236                                                           rx->rx_msgaddr),
1237                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1238         }
1239
1240         kiblnd_free_pages(conn->ibc_rx_pages);
1241
1242         conn->ibc_rx_pages = NULL;
1243 }
1244
1245 void
1246 kiblnd_map_rx_descs(kib_conn_t *conn)
1247 {
1248         kib_rx_t       *rx;
1249         struct page    *pg;
1250         int             pg_off;
1251         int             ipg;
1252         int             i;
1253
1254         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1255                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1256                 rx = &conn->ibc_rxs[i];
1257
1258                 rx->rx_conn = conn;
1259                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1260
1261                 rx->rx_msgaddr =
1262                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1263                                               rx->rx_msg, IBLND_MSG_SIZE,
1264                                               DMA_FROM_DEVICE);
1265                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1266                                                   rx->rx_msgaddr));
1267                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1268
1269                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1270                        i, rx->rx_msg, rx->rx_msgaddr,
1271                        (__u64)(page_to_phys(pg) + pg_off));
1272
1273                 pg_off += IBLND_MSG_SIZE;
1274                 LASSERT(pg_off <= PAGE_SIZE);
1275
1276                 if (pg_off == PAGE_SIZE) {
1277                         pg_off = 0;
1278                         ipg++;
1279                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1280                 }
1281         }
1282 }
1283
1284 static void
1285 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1286 {
1287         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1288         kib_tx_t       *tx;
1289         int             i;
1290
1291         LASSERT (tpo->tpo_pool.po_allocated == 0);
1292
1293         if (hdev == NULL)
1294                 return;
1295
1296         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1297                 tx = &tpo->tpo_tx_descs[i];
1298                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1299                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1300                                                           tx->tx_msgaddr),
1301                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1302         }
1303
1304         kiblnd_hdev_decref(hdev);
1305         tpo->tpo_hdev = NULL;
1306 }
1307
1308 static kib_hca_dev_t *
1309 kiblnd_current_hdev(kib_dev_t *dev)
1310 {
1311         kib_hca_dev_t *hdev;
1312         unsigned long  flags;
1313         int            i = 0;
1314
1315         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1316         while (dev->ibd_failover) {
1317                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1318                 if (i++ % 50 == 0)
1319                         CDEBUG(D_NET, "%s: Wait for failover\n",
1320                                dev->ibd_ifname);
1321                 set_current_state(TASK_INTERRUPTIBLE);
1322                 schedule_timeout(cfs_time_seconds(1) / 100);
1323
1324                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1325         }
1326
1327         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1328         hdev = dev->ibd_hdev;
1329
1330         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1331
1332         return hdev;
1333 }
1334
1335 static void
1336 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1337 {
1338         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1339         kib_pool_t     *pool  = &tpo->tpo_pool;
1340         kib_net_t      *net   = pool->po_owner->ps_net;
1341         kib_dev_t      *dev;
1342         struct page    *page;
1343         kib_tx_t       *tx;
1344         int             page_offset;
1345         int             ipage;
1346         int             i;
1347
1348         LASSERT (net != NULL);
1349
1350         dev = net->ibn_dev;
1351
1352         /* pre-mapped messages are not bigger than 1 page */
1353         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1354
1355         /* No fancy arithmetic when we do the buffer calculations */
1356         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1357
1358         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1359
1360         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1361                 page = txpgs->ibp_pages[ipage];
1362                 tx = &tpo->tpo_tx_descs[i];
1363
1364                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1365                                            page_offset);
1366
1367                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1368                                                        tx->tx_msg,
1369                                                        IBLND_MSG_SIZE,
1370                                                        DMA_TO_DEVICE);
1371                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1372                                                   tx->tx_msgaddr));
1373                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1374
1375                 list_add(&tx->tx_list, &pool->po_free_list);
1376
1377                 page_offset += IBLND_MSG_SIZE;
1378                 LASSERT(page_offset <= PAGE_SIZE);
1379
1380                 if (page_offset == PAGE_SIZE) {
1381                         page_offset = 0;
1382                         ipage++;
1383                         LASSERT(ipage <= txpgs->ibp_npages);
1384                 }
1385         }
1386 }
1387
1388 struct ib_mr *
1389 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1390                       int negotiated_nfrags)
1391 {
1392         kib_net_t     *net   = ni->ni_data;
1393         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
1394         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1395         int     mod;
1396         __u16   nfrags;
1397
1398         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1399         mod = tunables->lnd_map_on_demand;
1400         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1401
1402         LASSERT(hdev->ibh_mrs != NULL);
1403
1404         if (mod > 0 && nfrags <= rd->rd_nfrags)
1405                 return NULL;
1406
1407         return hdev->ibh_mrs;
1408 }
1409
1410 static void
1411 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1412 {
1413         LASSERT(fpo->fpo_map_count == 0);
1414
1415         if (fpo->fpo_is_fmr) {
1416                 if (fpo->fmr.fpo_fmr_pool)
1417                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1418         } else {
1419                 struct kib_fast_reg_descriptor *frd, *tmp;
1420                 int i = 0;
1421
1422                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1423                                          frd_list) {
1424                         list_del(&frd->frd_list);
1425                         ib_free_fast_reg_page_list(frd->frd_frpl);
1426                         ib_dereg_mr(frd->frd_mr);
1427                         LIBCFS_FREE(frd, sizeof(*frd));
1428                         i++;
1429                 }
1430                 if (i < fpo->fast_reg.fpo_pool_size)
1431                         CERROR("FastReg pool still has %d regions registered\n",
1432                                 fpo->fast_reg.fpo_pool_size - i);
1433         }
1434
1435         if (fpo->fpo_hdev)
1436                 kiblnd_hdev_decref(fpo->fpo_hdev);
1437
1438         LIBCFS_FREE(fpo, sizeof(*fpo));
1439 }
1440
1441 static void
1442 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1443 {
1444         kib_fmr_pool_t *fpo, *tmp;
1445
1446         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1447                 list_del(&fpo->fpo_list);
1448                 kiblnd_destroy_fmr_pool(fpo);
1449         }
1450 }
1451
1452 static int
1453 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1454                      int ncpts)
1455 {
1456         int size = tunables->lnd_fmr_pool_size / ncpts;
1457
1458         return max(IBLND_FMR_POOL, size);
1459 }
1460
1461 static int
1462 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1463                          int ncpts)
1464 {
1465         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1466
1467         return max(IBLND_FMR_POOL_FLUSH, size);
1468 }
1469
1470 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1471 {
1472         struct ib_fmr_pool_param param = {
1473                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1474                 .page_shift        = PAGE_SHIFT,
1475                 .access            = (IB_ACCESS_LOCAL_WRITE |
1476                                       IB_ACCESS_REMOTE_WRITE),
1477                 .pool_size         = fps->fps_pool_size,
1478                 .dirty_watermark   = fps->fps_flush_trigger,
1479                 .flush_function    = NULL,
1480                 .flush_arg         = NULL,
1481                 .cache             = !!fps->fps_cache };
1482         int rc = 0;
1483
1484         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1485                                                    &param);
1486         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1487                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1488                 if (rc != -ENOSYS)
1489                         CERROR("Failed to create FMR pool: %d\n", rc);
1490                 else
1491                         CERROR("FMRs are not supported\n");
1492         }
1493
1494         return rc;
1495 }
1496
1497 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1498 {
1499         struct kib_fast_reg_descriptor *frd, *tmp;
1500         int i, rc;
1501
1502         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1503         fpo->fast_reg.fpo_pool_size = 0;
1504         for (i = 0; i < fps->fps_pool_size; i++) {
1505                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1506                                  sizeof(*frd));
1507                 if (!frd) {
1508                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1509                         rc = -ENOMEM;
1510                         goto out;
1511                 }
1512                 frd->frd_mr = NULL;
1513
1514                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1515                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1516                 if (IS_ERR(frd->frd_frpl)) {
1517                         rc = PTR_ERR(frd->frd_frpl);
1518                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1519                                 rc);
1520                         frd->frd_frpl = NULL;
1521                         goto out_middle;
1522                 }
1523
1524                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1525                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1526                 if (IS_ERR(frd->frd_mr)) {
1527                         rc = PTR_ERR(frd->frd_mr);
1528                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1529                         frd->frd_mr = NULL;
1530                         goto out_middle;
1531                 }
1532
1533                 frd->frd_valid = true;
1534
1535                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1536                 fpo->fast_reg.fpo_pool_size++;
1537         }
1538
1539         return 0;
1540
1541 out_middle:
1542         if (frd->frd_mr)
1543                 ib_dereg_mr(frd->frd_mr);
1544         if (frd->frd_frpl)
1545                 ib_free_fast_reg_page_list(frd->frd_frpl);
1546         LIBCFS_FREE(frd, sizeof(*frd));
1547
1548 out:
1549         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1550                                  frd_list) {
1551                 list_del(&frd->frd_list);
1552                 ib_free_fast_reg_page_list(frd->frd_frpl);
1553                 ib_dereg_mr(frd->frd_mr);
1554                 LIBCFS_FREE(frd, sizeof(*frd));
1555         }
1556
1557         return rc;
1558 }
1559
1560 static int
1561 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1562 {
1563         struct ib_device_attr *dev_attr;
1564         kib_dev_t *dev = fps->fps_net->ibn_dev;
1565         kib_fmr_pool_t *fpo;
1566         int rc;
1567
1568         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1569         if (!dev_attr)
1570                 return -ENOMEM;
1571
1572         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1573         if (!fpo) {
1574                 rc = -ENOMEM;
1575                 goto out_dev_attr;
1576         }
1577
1578         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1579
1580         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1581         if (rc) {
1582                 CERROR("Query device failed for %s: %d\n",
1583                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1584                 goto out_dev_attr;
1585         }
1586
1587         /* Check for FMR or FastReg support */
1588         fpo->fpo_is_fmr = 0;
1589         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1590             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1591             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1592             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1593                 LCONSOLE_INFO("Using FMR for registration\n");
1594                 fpo->fpo_is_fmr = 1;
1595         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1596                 LCONSOLE_INFO("Using FastReg for registration\n");
1597         } else {
1598                 rc = -ENOSYS;
1599                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1600                 goto out_dev_attr;
1601         }
1602
1603         if (fpo->fpo_is_fmr)
1604                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1605         else
1606                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1607         if (rc)
1608                 goto out_fpo;
1609
1610         kfree(dev_attr);
1611         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1612         fpo->fpo_owner    = fps;
1613         *pp_fpo = fpo;
1614
1615         return 0;
1616
1617 out_fpo:
1618         kiblnd_hdev_decref(fpo->fpo_hdev);
1619         LIBCFS_FREE(fpo, sizeof(*fpo));
1620
1621 out_dev_attr:
1622         kfree(dev_attr);
1623
1624         return rc;
1625 }
1626
1627 static void
1628 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1629 {
1630         if (fps->fps_net == NULL) /* intialized? */
1631                 return;
1632
1633         spin_lock(&fps->fps_lock);
1634
1635         while (!list_empty(&fps->fps_pool_list)) {
1636                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1637                                                  kib_fmr_pool_t, fpo_list);
1638                 fpo->fpo_failed = 1;
1639                 list_del(&fpo->fpo_list);
1640                 if (fpo->fpo_map_count == 0)
1641                         list_add(&fpo->fpo_list, zombies);
1642                 else
1643                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1644         }
1645
1646         spin_unlock(&fps->fps_lock);
1647 }
1648
1649 static void
1650 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1651 {
1652         if (fps->fps_net != NULL) { /* initialized? */
1653                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1654                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1655         }
1656 }
1657
1658 static int
1659 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1660                         kib_net_t *net,
1661                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1662 {
1663         kib_fmr_pool_t *fpo;
1664         int             rc;
1665
1666         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1667
1668         fps->fps_net = net;
1669         fps->fps_cpt = cpt;
1670
1671         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1672         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1673         fps->fps_cache = tunables->lnd_fmr_cache;
1674
1675         spin_lock_init(&fps->fps_lock);
1676         INIT_LIST_HEAD(&fps->fps_pool_list);
1677         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1678
1679         rc = kiblnd_create_fmr_pool(fps, &fpo);
1680         if (rc == 0)
1681                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1682
1683         return rc;
1684 }
1685
1686 static int
1687 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1688 {
1689         if (fpo->fpo_map_count != 0) /* still in use */
1690                 return 0;
1691         if (fpo->fpo_failed)
1692                 return 1;
1693         return cfs_time_aftereq(now, fpo->fpo_deadline);
1694 }
1695
1696 void
1697 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1698 {
1699         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1700         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1701         kib_fmr_poolset_t *fps;
1702         cfs_time_t         now = cfs_time_current();
1703         kib_fmr_pool_t    *tmp;
1704         int                rc;
1705
1706         if (!fpo)
1707                 return;
1708
1709         fps = fpo->fpo_owner;
1710         if (fpo->fpo_is_fmr) {
1711                 if (fmr->fmr_pfmr) {
1712                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1713                         LASSERT(!rc);
1714                         fmr->fmr_pfmr = NULL;
1715                 }
1716
1717                 if (status) {
1718                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1719                         LASSERT(!rc);
1720                 }
1721         } else {
1722                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1723
1724                 if (frd) {
1725                         frd->frd_valid = false;
1726                         spin_lock(&fps->fps_lock);
1727                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1728                         spin_unlock(&fps->fps_lock);
1729                         fmr->fmr_frd = NULL;
1730                 }
1731         }
1732         fmr->fmr_pool = NULL;
1733
1734         spin_lock(&fps->fps_lock);
1735         fpo->fpo_map_count--;   /* decref the pool */
1736
1737         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1738                 /* the first pool is persistent */
1739                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1740                         continue;
1741
1742                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1743                         list_move(&fpo->fpo_list, &zombies);
1744                         fps->fps_version++;
1745                 }
1746         }
1747         spin_unlock(&fps->fps_lock);
1748
1749         if (!list_empty(&zombies))
1750                 kiblnd_destroy_fmr_pool_list(&zombies);
1751 }
1752
1753 int
1754 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1755                     __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1756 {
1757         kib_fmr_pool_t *fpo;
1758         __u64 version;
1759         int rc;
1760
1761 again:
1762         spin_lock(&fps->fps_lock);
1763         version = fps->fps_version;
1764         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1765                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1766                 fpo->fpo_map_count++;
1767
1768                 if (fpo->fpo_is_fmr) {
1769                         struct ib_pool_fmr *pfmr;
1770
1771                         spin_unlock(&fps->fps_lock);
1772                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1773                                                     pages, npages, iov);
1774                         if (likely(!IS_ERR(pfmr))) {
1775                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1776                                                       : pfmr->fmr->lkey;
1777                                 fmr->fmr_frd  = NULL;
1778                                 fmr->fmr_pfmr = pfmr;
1779                                 fmr->fmr_pool = fpo;
1780                                 return 0;
1781                         }
1782                         rc = PTR_ERR(pfmr);
1783                 } else {
1784                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1785                                 struct ib_send_wr *wr;
1786                                 struct kib_fast_reg_descriptor *frd;
1787                                 struct ib_fast_reg_page_list *frpl;
1788                                 struct ib_mr *mr;
1789
1790                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1791                                                         struct kib_fast_reg_descriptor,
1792                                                         frd_list);
1793                                 list_del(&frd->frd_list);
1794                                 spin_unlock(&fps->fps_lock);
1795
1796                                 frpl = frd->frd_frpl;
1797                                 mr   = frd->frd_mr;
1798
1799                                 if (!frd->frd_valid) {
1800                                         struct ib_send_wr *inv_wr;
1801                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1802
1803                                         inv_wr = &frd->frd_inv_wr;
1804                                         memset(inv_wr, 0, sizeof(*inv_wr));
1805                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1806                                         inv_wr->wr_id = IBLND_WID_MR;
1807                                         inv_wr->ex.invalidate_rkey = key;
1808
1809                                         /* Bump the key */
1810                                         key = ib_inc_rkey(key);
1811                                         ib_update_fast_reg_key(mr, key);
1812                                 }
1813
1814                                 LASSERT(npages <= frpl->max_page_list_len);
1815                                 memcpy(frpl->page_list, pages,
1816                                         sizeof(*pages) * npages);
1817
1818                                 /* Prepare FastReg WR */
1819                                 wr = &frd->frd_fastreg_wr;
1820                                 memset(wr, 0, sizeof(*wr));
1821                                 wr->opcode = IB_WR_FAST_REG_MR;
1822                                 wr->wr_id = IBLND_WID_MR;
1823                                 wr->wr.fast_reg.iova_start = iov;
1824                                 wr->wr.fast_reg.page_list  = frpl;
1825                                 wr->wr.fast_reg.page_list_len = npages;
1826                                 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1827                                 wr->wr.fast_reg.length = nob;
1828                                 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1829                                                              : mr->lkey;
1830                                 wr->wr.fast_reg.access_flags =
1831                                                 (IB_ACCESS_LOCAL_WRITE |
1832                                                  IB_ACCESS_REMOTE_WRITE);
1833
1834                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1835                                 fmr->fmr_frd  = frd;
1836                                 fmr->fmr_pfmr = NULL;
1837                                 fmr->fmr_pool = fpo;
1838                                 return 0;
1839                         }
1840                         spin_unlock(&fps->fps_lock);
1841                         rc = -EBUSY;
1842                 }
1843
1844                 spin_lock(&fps->fps_lock);
1845                 fpo->fpo_map_count--;
1846                 if (rc != -EAGAIN) {
1847                         spin_unlock(&fps->fps_lock);
1848                         return rc;
1849                 }
1850
1851                 /* EAGAIN and ... */
1852                 if (version != fps->fps_version) {
1853                         spin_unlock(&fps->fps_lock);
1854                         goto again;
1855                 }
1856         }
1857
1858         if (fps->fps_increasing) {
1859                 spin_unlock(&fps->fps_lock);
1860                 CDEBUG(D_NET, "Another thread is allocating new "
1861                        "FMR pool, waiting for her to complete\n");
1862                 schedule();
1863                 goto again;
1864
1865         }
1866
1867         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1868                 /* someone failed recently */
1869                 spin_unlock(&fps->fps_lock);
1870                 return -EAGAIN;
1871         }
1872
1873         fps->fps_increasing = 1;
1874         spin_unlock(&fps->fps_lock);
1875
1876         CDEBUG(D_NET, "Allocate new FMR pool\n");
1877         rc = kiblnd_create_fmr_pool(fps, &fpo);
1878         spin_lock(&fps->fps_lock);
1879         fps->fps_increasing = 0;
1880         if (rc == 0) {
1881                 fps->fps_version++;
1882                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1883         } else {
1884                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1885         }
1886         spin_unlock(&fps->fps_lock);
1887
1888         goto again;
1889 }
1890
1891 static void
1892 kiblnd_fini_pool(kib_pool_t *pool)
1893 {
1894         LASSERT(list_empty(&pool->po_free_list));
1895         LASSERT(pool->po_allocated == 0);
1896
1897         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1898 }
1899
1900 static void
1901 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1902 {
1903         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1904
1905         memset(pool, 0, sizeof(kib_pool_t));
1906         INIT_LIST_HEAD(&pool->po_free_list);
1907         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1908         pool->po_owner    = ps;
1909         pool->po_size     = size;
1910 }
1911
1912 static void
1913 kiblnd_destroy_pool_list(struct list_head *head)
1914 {
1915         kib_pool_t *pool;
1916
1917         while (!list_empty(head)) {
1918                 pool = list_entry(head->next, kib_pool_t, po_list);
1919                 list_del(&pool->po_list);
1920
1921                 LASSERT(pool->po_owner != NULL);
1922                 pool->po_owner->ps_pool_destroy(pool);
1923         }
1924 }
1925
1926 static void
1927 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1928 {
1929         if (ps->ps_net == NULL) /* intialized? */
1930                 return;
1931
1932         spin_lock(&ps->ps_lock);
1933         while (!list_empty(&ps->ps_pool_list)) {
1934                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1935                                             kib_pool_t, po_list);
1936                 po->po_failed = 1;
1937                 list_del(&po->po_list);
1938                 if (po->po_allocated == 0)
1939                         list_add(&po->po_list, zombies);
1940                 else
1941                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1942         }
1943         spin_unlock(&ps->ps_lock);
1944 }
1945
1946 static void
1947 kiblnd_fini_poolset(kib_poolset_t *ps)
1948 {
1949         if (ps->ps_net != NULL) { /* initialized? */
1950                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1951                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1952         }
1953 }
1954
1955 static int
1956 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1957                     kib_net_t *net, char *name, int size,
1958                     kib_ps_pool_create_t po_create,
1959                     kib_ps_pool_destroy_t po_destroy,
1960                     kib_ps_node_init_t nd_init,
1961                     kib_ps_node_fini_t nd_fini)
1962 {
1963         kib_pool_t      *pool;
1964         int             rc;
1965
1966         memset(ps, 0, sizeof(kib_poolset_t));
1967
1968         ps->ps_cpt          = cpt;
1969         ps->ps_net          = net;
1970         ps->ps_pool_create  = po_create;
1971         ps->ps_pool_destroy = po_destroy;
1972         ps->ps_node_init    = nd_init;
1973         ps->ps_node_fini    = nd_fini;
1974         ps->ps_pool_size    = size;
1975         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1976             >= sizeof(ps->ps_name))
1977                 return -E2BIG;
1978         spin_lock_init(&ps->ps_lock);
1979         INIT_LIST_HEAD(&ps->ps_pool_list);
1980         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1981
1982         rc = ps->ps_pool_create(ps, size, &pool);
1983         if (rc == 0)
1984                 list_add(&pool->po_list, &ps->ps_pool_list);
1985         else
1986                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1987
1988         return rc;
1989 }
1990
1991 static int
1992 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1993 {
1994         if (pool->po_allocated != 0) /* still in use */
1995                 return 0;
1996         if (pool->po_failed)
1997                 return 1;
1998         return cfs_time_aftereq(now, pool->po_deadline);
1999 }
2000
2001 void
2002 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2003 {
2004         struct list_head zombies = LIST_HEAD_INIT(zombies);
2005         kib_poolset_t   *ps = pool->po_owner;
2006         kib_pool_t      *tmp;
2007         cfs_time_t       now = cfs_time_current();
2008
2009         spin_lock(&ps->ps_lock);
2010
2011         if (ps->ps_node_fini != NULL)
2012                 ps->ps_node_fini(pool, node);
2013
2014         LASSERT(pool->po_allocated > 0);
2015         list_add(node, &pool->po_free_list);
2016         pool->po_allocated--;
2017
2018         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2019                 /* the first pool is persistent */
2020                 if (ps->ps_pool_list.next == &pool->po_list)
2021                         continue;
2022
2023                 if (kiblnd_pool_is_idle(pool, now))
2024                         list_move(&pool->po_list, &zombies);
2025         }
2026         spin_unlock(&ps->ps_lock);
2027
2028         if (!list_empty(&zombies))
2029                 kiblnd_destroy_pool_list(&zombies);
2030 }
2031
2032 struct list_head *
2033 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2034 {
2035         struct list_head        *node;
2036         kib_pool_t              *pool;
2037         int                     rc;
2038         unsigned int            interval = 1;
2039         cfs_time_t              time_before;
2040         unsigned int            trips = 0;
2041
2042 again:
2043         spin_lock(&ps->ps_lock);
2044         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2045                 if (list_empty(&pool->po_free_list))
2046                         continue;
2047
2048                 pool->po_allocated++;
2049                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2050                 node = pool->po_free_list.next;
2051                 list_del(node);
2052
2053                 if (ps->ps_node_init != NULL) {
2054                         /* still hold the lock */
2055                         ps->ps_node_init(pool, node);
2056                 }
2057                 spin_unlock(&ps->ps_lock);
2058                 return node;
2059         }
2060
2061         /* no available tx pool and ... */
2062         if (ps->ps_increasing) {
2063                 /* another thread is allocating a new pool */
2064                 spin_unlock(&ps->ps_lock);
2065                 trips++;
2066                 CDEBUG(D_NET, "Another thread is allocating new "
2067                        "%s pool, waiting %d HZs for her to complete."
2068                        "trips = %d\n",
2069                        ps->ps_name, interval, trips);
2070
2071                 set_current_state(TASK_INTERRUPTIBLE);
2072                 schedule_timeout(interval);
2073                 if (interval < cfs_time_seconds(1))
2074                         interval *= 2;
2075
2076                 goto again;
2077         }
2078
2079         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2080                 /* someone failed recently */
2081                 spin_unlock(&ps->ps_lock);
2082                 return NULL;
2083         }
2084
2085         ps->ps_increasing = 1;
2086         spin_unlock(&ps->ps_lock);
2087
2088         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2089         time_before = cfs_time_current();
2090         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2091         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2092                cfs_time_current() - time_before);
2093
2094         spin_lock(&ps->ps_lock);
2095         ps->ps_increasing = 0;
2096         if (rc == 0) {
2097                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2098         } else {
2099                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2100                 CERROR("Can't allocate new %s pool because out of memory\n",
2101                        ps->ps_name);
2102         }
2103         spin_unlock(&ps->ps_lock);
2104
2105         goto again;
2106 }
2107
2108 static void
2109 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2110 {
2111         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2112         int             i;
2113
2114         LASSERT (pool->po_allocated == 0);
2115
2116         if (tpo->tpo_tx_pages != NULL) {
2117                 kiblnd_unmap_tx_pool(tpo);
2118                 kiblnd_free_pages(tpo->tpo_tx_pages);
2119         }
2120
2121         if (tpo->tpo_tx_descs == NULL)
2122                 goto out;
2123
2124         for (i = 0; i < pool->po_size; i++) {
2125                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2126
2127                 list_del(&tx->tx_list);
2128                 if (tx->tx_pages != NULL)
2129                         LIBCFS_FREE(tx->tx_pages,
2130                                     LNET_MAX_IOV *
2131                                     sizeof(*tx->tx_pages));
2132                 if (tx->tx_frags != NULL)
2133                         LIBCFS_FREE(tx->tx_frags,
2134                                     IBLND_MAX_RDMA_FRAGS *
2135                                             sizeof(*tx->tx_frags));
2136                 if (tx->tx_wrq != NULL)
2137                         LIBCFS_FREE(tx->tx_wrq,
2138                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2139                                     sizeof(*tx->tx_wrq));
2140                 if (tx->tx_sge != NULL)
2141                         LIBCFS_FREE(tx->tx_sge,
2142                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2143                                     sizeof(*tx->tx_sge));
2144                 if (tx->tx_rd != NULL)
2145                         LIBCFS_FREE(tx->tx_rd,
2146                                     offsetof(kib_rdma_desc_t,
2147                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2148         }
2149
2150         LIBCFS_FREE(tpo->tpo_tx_descs,
2151                     pool->po_size * sizeof(kib_tx_t));
2152 out:
2153         kiblnd_fini_pool(pool);
2154         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2155 }
2156
2157 static int kiblnd_tx_pool_size(int ncpts)
2158 {
2159         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2160
2161         return max(IBLND_TX_POOL, ntx);
2162 }
2163
2164 static int
2165 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2166 {
2167         int            i;
2168         int            npg;
2169         kib_pool_t    *pool;
2170         kib_tx_pool_t *tpo;
2171
2172         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2173         if (tpo == NULL) {
2174                 CERROR("Failed to allocate TX pool\n");
2175                 return -ENOMEM;
2176         }
2177
2178         pool = &tpo->tpo_pool;
2179         kiblnd_init_pool(ps, pool, size);
2180         tpo->tpo_tx_descs = NULL;
2181         tpo->tpo_tx_pages = NULL;
2182
2183         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2184         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2185                 CERROR("Can't allocate tx pages: %d\n", npg);
2186                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2187                 return -ENOMEM;
2188         }
2189
2190         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2191                          size * sizeof(kib_tx_t));
2192         if (tpo->tpo_tx_descs == NULL) {
2193                 CERROR("Can't allocate %d tx descriptors\n", size);
2194                 ps->ps_pool_destroy(pool);
2195                 return -ENOMEM;
2196         }
2197
2198         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2199
2200         for (i = 0; i < size; i++) {
2201                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2202
2203                 tx->tx_pool = tpo;
2204                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2205                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2206                                          lnet_cpt_table(), ps->ps_cpt,
2207                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2208                         if (tx->tx_pages == NULL)
2209                                 break;
2210                 }
2211
2212                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2213                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2214                 if (tx->tx_frags == NULL)
2215                         break;
2216
2217                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2218
2219                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2220                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2221                                  sizeof(*tx->tx_wrq));
2222                 if (tx->tx_wrq == NULL)
2223                         break;
2224
2225                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2226                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2227                                  sizeof(*tx->tx_sge));
2228                 if (tx->tx_sge == NULL)
2229                         break;
2230
2231                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2232                                  offsetof(kib_rdma_desc_t,
2233                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2234                 if (tx->tx_rd == NULL)
2235                         break;
2236         }
2237
2238         if (i == size) {
2239                 kiblnd_map_tx_pool(tpo);
2240                 *pp_po = pool;
2241                 return 0;
2242         }
2243
2244         ps->ps_pool_destroy(pool);
2245         return -ENOMEM;
2246 }
2247
2248 static void
2249 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2250 {
2251         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2252                                              tps_poolset);
2253         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2254
2255         tx->tx_cookie = tps->tps_next_tx_cookie++;
2256 }
2257
2258 static void
2259 kiblnd_net_fini_pools(kib_net_t *net)
2260 {
2261         int     i;
2262
2263         cfs_cpt_for_each(i, lnet_cpt_table()) {
2264                 kib_tx_poolset_t        *tps;
2265                 kib_fmr_poolset_t       *fps;
2266
2267                 if (net->ibn_tx_ps != NULL) {
2268                         tps = net->ibn_tx_ps[i];
2269                         kiblnd_fini_poolset(&tps->tps_poolset);
2270                 }
2271
2272                 if (net->ibn_fmr_ps != NULL) {
2273                         fps = net->ibn_fmr_ps[i];
2274                         kiblnd_fini_fmr_poolset(fps);
2275                 }
2276         }
2277
2278         if (net->ibn_tx_ps != NULL) {
2279                 cfs_percpt_free(net->ibn_tx_ps);
2280                 net->ibn_tx_ps = NULL;
2281         }
2282
2283         if (net->ibn_fmr_ps != NULL) {
2284                 cfs_percpt_free(net->ibn_fmr_ps);
2285                 net->ibn_fmr_ps = NULL;
2286         }
2287 }
2288
2289 static int
2290 kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts)
2291 {
2292         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2293         unsigned long   flags;
2294         int             cpt;
2295         int             rc;
2296         int             i;
2297
2298         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2299
2300         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2301         if (tunables->lnd_map_on_demand == 0) {
2302                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2303                                            flags);
2304                 goto create_tx_pool;
2305         }
2306
2307         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2308
2309         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2310                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2311                        tunables->lnd_fmr_pool_size,
2312                        *kiblnd_tunables.kib_ntx / 4);
2313                 rc = -EINVAL;
2314                 goto failed;
2315         }
2316
2317         /* TX pool must be created later than FMR, see LU-2268
2318          * for details */
2319         LASSERT(net->ibn_tx_ps == NULL);
2320
2321         /* premapping can fail if ibd_nmr > 1, so we always create
2322          * FMR pool and map-on-demand if premapping failed */
2323
2324         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2325                                            sizeof(kib_fmr_poolset_t));
2326         if (net->ibn_fmr_ps == NULL) {
2327                 CERROR("Failed to allocate FMR pool array\n");
2328                 rc = -ENOMEM;
2329                 goto failed;
2330         }
2331
2332         for (i = 0; i < ncpts; i++) {
2333                 cpt = (cpts == NULL) ? i : cpts[i];
2334                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2335                                              net, tunables);
2336                 if (rc != 0) {
2337                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2338                                cpt, rc);
2339                         goto failed;
2340                 }
2341         }
2342
2343         if (i > 0)
2344                 LASSERT(i == ncpts);
2345
2346  create_tx_pool:
2347         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2348                                           sizeof(kib_tx_poolset_t));
2349         if (net->ibn_tx_ps == NULL) {
2350                 CERROR("Failed to allocate tx pool array\n");
2351                 rc = -ENOMEM;
2352                 goto failed;
2353         }
2354
2355         for (i = 0; i < ncpts; i++) {
2356                 cpt = (cpts == NULL) ? i : cpts[i];
2357                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2358                                          cpt, net, "TX",
2359                                          kiblnd_tx_pool_size(ncpts),
2360                                          kiblnd_create_tx_pool,
2361                                          kiblnd_destroy_tx_pool,
2362                                          kiblnd_tx_init, NULL);
2363                 if (rc != 0) {
2364                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2365                                cpt, rc);
2366                         goto failed;
2367                 }
2368         }
2369
2370         return 0;
2371  failed:
2372         kiblnd_net_fini_pools(net);
2373         LASSERT(rc != 0);
2374         return rc;
2375 }
2376
2377 static int
2378 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2379 {
2380         struct ib_device_attr *attr;
2381         int                    rc;
2382
2383         /* It's safe to assume a HCA can handle a page size
2384          * matching that of the native system */
2385         hdev->ibh_page_shift = PAGE_SHIFT;
2386         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2387         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2388
2389         LIBCFS_ALLOC(attr, sizeof(*attr));
2390         if (attr == NULL) {
2391                 CERROR("Out of memory\n");
2392                 return -ENOMEM;
2393         }
2394
2395         rc = ib_query_device(hdev->ibh_ibdev, attr);
2396         if (rc == 0)
2397                 hdev->ibh_mr_size = attr->max_mr_size;
2398
2399         LIBCFS_FREE(attr, sizeof(*attr));
2400
2401         if (rc != 0) {
2402                 CERROR("Failed to query IB device: %d\n", rc);
2403                 return rc;
2404         }
2405
2406         if (hdev->ibh_mr_size == ~0ULL) {
2407                 hdev->ibh_mr_shift = 64;
2408                 return 0;
2409         }
2410
2411         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2412         return -EINVAL;
2413 }
2414
2415 static void
2416 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2417 {
2418         if (hdev->ibh_mrs == NULL)
2419                 return;
2420
2421         ib_dereg_mr(hdev->ibh_mrs);
2422
2423         hdev->ibh_mrs = NULL;
2424 }
2425
2426 void
2427 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2428 {
2429         kiblnd_hdev_cleanup_mrs(hdev);
2430
2431         if (hdev->ibh_pd != NULL)
2432                 ib_dealloc_pd(hdev->ibh_pd);
2433
2434         if (hdev->ibh_cmid != NULL)
2435                 rdma_destroy_id(hdev->ibh_cmid);
2436
2437         LIBCFS_FREE(hdev, sizeof(*hdev));
2438 }
2439
2440 static int
2441 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2442 {
2443         struct ib_mr *mr;
2444         int           rc;
2445         int           acflags = IB_ACCESS_LOCAL_WRITE |
2446                                 IB_ACCESS_REMOTE_WRITE;
2447
2448         rc = kiblnd_hdev_get_attr(hdev);
2449         if (rc != 0)
2450                 return rc;
2451
2452         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2453         if (IS_ERR(mr)) {
2454                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2455                 kiblnd_hdev_cleanup_mrs(hdev);
2456                 return PTR_ERR(mr);
2457         }
2458
2459         hdev->ibh_mrs = mr;
2460
2461         return 0;
2462 }
2463
2464 static int
2465 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2466 {       /* DUMMY */
2467         return 0;
2468 }
2469
2470 static int
2471 kiblnd_dev_need_failover(kib_dev_t *dev)
2472 {
2473         struct rdma_cm_id  *cmid;
2474         struct sockaddr_in  srcaddr;
2475         struct sockaddr_in  dstaddr;
2476         int                 rc;
2477
2478         if (dev->ibd_hdev == NULL || /* initializing */
2479             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2480             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2481                 return 1;
2482
2483         /* XXX: it's UGLY, but I don't have better way to find
2484          * ib-bonding HCA failover because:
2485          *
2486          * a. no reliable CM event for HCA failover...
2487          * b. no OFED API to get ib_device for current net_device...
2488          *
2489          * We have only two choices at this point:
2490          *
2491          * a. rdma_bind_addr(), it will conflict with listener cmid
2492          * b. rdma_resolve_addr() to zero addr */
2493         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2494                                      IB_QPT_RC);
2495         if (IS_ERR(cmid)) {
2496                 rc = PTR_ERR(cmid);
2497                 CERROR("Failed to create cmid for failover: %d\n", rc);
2498                 return rc;
2499         }
2500
2501         memset(&srcaddr, 0, sizeof(srcaddr));
2502         srcaddr.sin_family      = AF_INET;
2503         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2504
2505         memset(&dstaddr, 0, sizeof(dstaddr));
2506         dstaddr.sin_family = AF_INET;
2507         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2508                                (struct sockaddr *)&dstaddr, 1);
2509         if (rc != 0 || cmid->device == NULL) {
2510                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2511                        dev->ibd_ifname, &dev->ibd_ifip,
2512                        cmid->device, rc);
2513                 rdma_destroy_id(cmid);
2514                 return rc;
2515         }
2516
2517         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2518         rdma_destroy_id(cmid);
2519         return rc;
2520 }
2521
2522 int
2523 kiblnd_dev_failover(kib_dev_t *dev)
2524 {
2525         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2526         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2527         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2528         struct rdma_cm_id  *cmid  = NULL;
2529         kib_hca_dev_t      *hdev  = NULL;
2530         kib_hca_dev_t      *old;
2531         struct ib_pd       *pd;
2532         kib_net_t          *net;
2533         struct sockaddr_in  addr;
2534         unsigned long       flags;
2535         int                 rc = 0;
2536         int                 i;
2537
2538         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2539                  dev->ibd_can_failover ||
2540                  dev->ibd_hdev == NULL);
2541
2542         rc = kiblnd_dev_need_failover(dev);
2543         if (rc <= 0)
2544                 goto out;
2545
2546         if (dev->ibd_hdev != NULL &&
2547             dev->ibd_hdev->ibh_cmid != NULL) {
2548                 /* XXX it's not good to close old listener at here,
2549                  * because we can fail to create new listener.
2550                  * But we have to close it now, otherwise rdma_bind_addr
2551                  * will return EADDRINUSE... How crap! */
2552                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2553
2554                 cmid = dev->ibd_hdev->ibh_cmid;
2555                 /* make next schedule of kiblnd_dev_need_failover()
2556                  * return 1 for me */
2557                 dev->ibd_hdev->ibh_cmid  = NULL;
2558                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2559
2560                 rdma_destroy_id(cmid);
2561         }
2562
2563         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2564                                      IB_QPT_RC);
2565         if (IS_ERR(cmid)) {
2566                 rc = PTR_ERR(cmid);
2567                 CERROR("Failed to create cmid for failover: %d\n", rc);
2568                 goto out;
2569         }
2570
2571         memset(&addr, 0, sizeof(addr));
2572         addr.sin_family      = AF_INET;
2573         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2574         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2575
2576         /* Bind to failover device or port */
2577         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2578         if (rc != 0 || cmid->device == NULL) {
2579                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2580                        dev->ibd_ifname, &dev->ibd_ifip,
2581                        cmid->device, rc);
2582                 rdma_destroy_id(cmid);
2583                 goto out;
2584         }
2585
2586         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2587         if (hdev == NULL) {
2588                 CERROR("Failed to allocate kib_hca_dev\n");
2589                 rdma_destroy_id(cmid);
2590                 rc = -ENOMEM;
2591                 goto out;
2592         }
2593
2594         atomic_set(&hdev->ibh_ref, 1);
2595         hdev->ibh_dev   = dev;
2596         hdev->ibh_cmid  = cmid;
2597         hdev->ibh_ibdev = cmid->device;
2598
2599         pd = ib_alloc_pd(cmid->device);
2600         if (IS_ERR(pd)) {
2601                 rc = PTR_ERR(pd);
2602                 CERROR("Can't allocate PD: %d\n", rc);
2603                 goto out;
2604         }
2605
2606         hdev->ibh_pd = pd;
2607
2608         rc = rdma_listen(cmid, 0);
2609         if (rc != 0) {
2610                 CERROR("Can't start new listener: %d\n", rc);
2611                 goto out;
2612         }
2613
2614         rc = kiblnd_hdev_setup_mrs(hdev);
2615         if (rc != 0) {
2616                 CERROR("Can't setup device: %d\n", rc);
2617                 goto out;
2618         }
2619
2620         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2621
2622         old = dev->ibd_hdev;
2623         dev->ibd_hdev = hdev;   /* take over the refcount */
2624         hdev = old;
2625
2626         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2627                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2628                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2629                                             &zombie_tpo);
2630
2631                         if (net->ibn_fmr_ps != NULL)
2632                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2633                                                         &zombie_fpo);
2634                 }
2635         }
2636
2637         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2638  out:
2639         if (!list_empty(&zombie_tpo))
2640                 kiblnd_destroy_pool_list(&zombie_tpo);
2641         if (!list_empty(&zombie_ppo))
2642                 kiblnd_destroy_pool_list(&zombie_ppo);
2643         if (!list_empty(&zombie_fpo))
2644                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2645         if (hdev != NULL)
2646                 kiblnd_hdev_decref(hdev);
2647
2648         if (rc != 0)
2649                 dev->ibd_failed_failover++;
2650         else
2651                 dev->ibd_failed_failover = 0;
2652
2653         return rc;
2654 }
2655
2656 void
2657 kiblnd_destroy_dev (kib_dev_t *dev)
2658 {
2659         LASSERT (dev->ibd_nnets == 0);
2660         LASSERT(list_empty(&dev->ibd_nets));
2661
2662         list_del(&dev->ibd_fail_list);
2663         list_del(&dev->ibd_list);
2664
2665         if (dev->ibd_hdev != NULL)
2666                 kiblnd_hdev_decref(dev->ibd_hdev);
2667
2668         LIBCFS_FREE(dev, sizeof(*dev));
2669 }
2670
2671 static kib_dev_t *
2672 kiblnd_create_dev(char *ifname)
2673 {
2674         struct net_device *netdev;
2675         kib_dev_t         *dev;
2676         __u32              netmask;
2677         __u32              ip;
2678         int                up;
2679         int                rc;
2680
2681         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2682         if (rc != 0) {
2683                 CERROR("Can't query IPoIB interface %s: %d\n",
2684                        ifname, rc);
2685                 return NULL;
2686         }
2687
2688         if (!up) {
2689                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2690                 return NULL;
2691         }
2692
2693         LIBCFS_ALLOC(dev, sizeof(*dev));
2694         if (dev == NULL)
2695                 return NULL;
2696
2697         netdev = dev_get_by_name(&init_net, ifname);
2698         if (netdev == NULL) {
2699                 dev->ibd_can_failover = 0;
2700         } else {
2701                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2702                 dev_put(netdev);
2703         }
2704
2705         INIT_LIST_HEAD(&dev->ibd_nets);
2706         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2707         INIT_LIST_HEAD(&dev->ibd_fail_list);
2708         dev->ibd_ifip = ip;
2709         strcpy(&dev->ibd_ifname[0], ifname);
2710
2711         /* initialize the device */
2712         rc = kiblnd_dev_failover(dev);
2713         if (rc != 0) {
2714                 CERROR("Can't initialize device: %d\n", rc);
2715                 LIBCFS_FREE(dev, sizeof(*dev));
2716                 return NULL;
2717         }
2718
2719         list_add_tail(&dev->ibd_list,
2720                           &kiblnd_data.kib_devs);
2721         return dev;
2722 }
2723
2724 static void
2725 kiblnd_base_shutdown(void)
2726 {
2727         struct kib_sched_info   *sched;
2728         int                     i;
2729
2730         LASSERT(list_empty(&kiblnd_data.kib_devs));
2731
2732         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2733                atomic_read(&libcfs_kmemory));
2734
2735         switch (kiblnd_data.kib_init) {
2736         default:
2737                 LBUG();
2738
2739         case IBLND_INIT_ALL:
2740         case IBLND_INIT_DATA:
2741                 LASSERT (kiblnd_data.kib_peers != NULL);
2742                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2743                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2744                 }
2745                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2746                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2747                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2748                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2749
2750                 /* flag threads to terminate; wake and wait for them to die */
2751                 kiblnd_data.kib_shutdown = 1;
2752
2753                 /* NB: we really want to stop scheduler threads net by net
2754                  * instead of the whole module, this should be improved
2755                  * with dynamic configuration LNet */
2756                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2757                         wake_up_all(&sched->ibs_waitq);
2758
2759                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2760                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2761
2762                 i = 2;
2763                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2764                         i++;
2765                         /* power of 2? */
2766                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2767                                "Waiting for %d threads to terminate\n",
2768                                atomic_read(&kiblnd_data.kib_nthreads));
2769                         set_current_state(TASK_UNINTERRUPTIBLE);
2770                         schedule_timeout(cfs_time_seconds(1));
2771                 }
2772
2773                 /* fall through */
2774
2775         case IBLND_INIT_NOTHING:
2776                 break;
2777         }
2778
2779         if (kiblnd_data.kib_peers != NULL) {
2780                 LIBCFS_FREE(kiblnd_data.kib_peers,
2781                             sizeof(struct list_head) *
2782                             kiblnd_data.kib_peer_hash_size);
2783         }
2784
2785         if (kiblnd_data.kib_scheds != NULL)
2786                 cfs_percpt_free(kiblnd_data.kib_scheds);
2787
2788         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2789                atomic_read(&libcfs_kmemory));
2790
2791         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2792         module_put(THIS_MODULE);
2793 }
2794
2795 static void
2796 kiblnd_shutdown (lnet_ni_t *ni)
2797 {
2798         kib_net_t        *net = ni->ni_data;
2799         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2800         int               i;
2801         unsigned long     flags;
2802
2803         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2804
2805         if (net == NULL)
2806                 goto out;
2807
2808         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2809                atomic_read(&libcfs_kmemory));
2810
2811         write_lock_irqsave(g_lock, flags);
2812         net->ibn_shutdown = 1;
2813         write_unlock_irqrestore(g_lock, flags);
2814
2815         switch (net->ibn_init) {
2816         default:
2817                 LBUG();
2818
2819         case IBLND_INIT_ALL:
2820                 /* nuke all existing peers within this net */
2821                 kiblnd_del_peer(ni, LNET_NID_ANY);
2822
2823                 /* Wait for all peer state to clean up */
2824                 i = 2;
2825                 while (atomic_read(&net->ibn_npeers) != 0) {
2826                         i++;
2827                         /* power of 2? */
2828                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2829                                "%s: waiting for %d peers to disconnect\n",
2830                                libcfs_nid2str(ni->ni_nid),
2831                                atomic_read(&net->ibn_npeers));
2832                         set_current_state(TASK_UNINTERRUPTIBLE);
2833                         schedule_timeout(cfs_time_seconds(1));
2834                 }
2835
2836                 kiblnd_net_fini_pools(net);
2837
2838                 write_lock_irqsave(g_lock, flags);
2839                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2840                 net->ibn_dev->ibd_nnets--;
2841                 list_del(&net->ibn_list);
2842                 write_unlock_irqrestore(g_lock, flags);
2843
2844                 /* fall through */
2845
2846         case IBLND_INIT_NOTHING:
2847                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2848
2849                 if (net->ibn_dev != NULL &&
2850                     net->ibn_dev->ibd_nnets == 0)
2851                         kiblnd_destroy_dev(net->ibn_dev);
2852
2853                 break;
2854         }
2855
2856         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2857                atomic_read(&libcfs_kmemory));
2858
2859         net->ibn_init = IBLND_INIT_NOTHING;
2860         ni->ni_data = NULL;
2861
2862         LIBCFS_FREE(net, sizeof(*net));
2863
2864 out:
2865         if (list_empty(&kiblnd_data.kib_devs))
2866                 kiblnd_base_shutdown();
2867         return;
2868 }
2869
2870 static int
2871 kiblnd_base_startup(void)
2872 {
2873         struct kib_sched_info   *sched;
2874         int                     rc;
2875         int                     i;
2876
2877         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2878
2879         try_module_get(THIS_MODULE);
2880         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2881
2882         rwlock_init(&kiblnd_data.kib_global_lock);
2883
2884         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2885         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2886
2887         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2888         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2889                      sizeof(struct list_head) *
2890                      kiblnd_data.kib_peer_hash_size);
2891         if (kiblnd_data.kib_peers == NULL)
2892                 goto failed;
2893
2894         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2895                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2896
2897         spin_lock_init(&kiblnd_data.kib_connd_lock);
2898         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2899         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2900         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2901         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2902
2903         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2904         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2905
2906         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2907                                                   sizeof(*sched));
2908         if (kiblnd_data.kib_scheds == NULL)
2909                 goto failed;
2910
2911         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2912                 int     nthrs;
2913
2914                 spin_lock_init(&sched->ibs_lock);
2915                 INIT_LIST_HEAD(&sched->ibs_conns);
2916                 init_waitqueue_head(&sched->ibs_waitq);
2917
2918                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2919                 if (*kiblnd_tunables.kib_nscheds > 0) {
2920                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2921                 } else {
2922                         /* max to half of CPUs, another half is reserved for
2923                          * upper layer modules */
2924                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2925                 }
2926
2927                 sched->ibs_nthreads_max = nthrs;
2928                 sched->ibs_cpt = i;
2929         }
2930
2931         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2932
2933         /* lists/ptrs/locks initialised */
2934         kiblnd_data.kib_init = IBLND_INIT_DATA;
2935         /*****************************************************/
2936
2937         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2938         if (rc != 0) {
2939                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2940                 goto failed;
2941         }
2942
2943         if (*kiblnd_tunables.kib_dev_failover != 0)
2944                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2945                                          "kiblnd_failover");
2946
2947         if (rc != 0) {
2948                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2949                 goto failed;
2950         }
2951
2952         /* flag everything initialised */
2953         kiblnd_data.kib_init = IBLND_INIT_ALL;
2954         /*****************************************************/
2955
2956         return 0;
2957
2958  failed:
2959         kiblnd_base_shutdown();
2960         return -ENETDOWN;
2961 }
2962
2963 static int
2964 kiblnd_start_schedulers(struct kib_sched_info *sched)
2965 {
2966         int     rc = 0;
2967         int     nthrs;
2968         int     i;
2969
2970         if (sched->ibs_nthreads == 0) {
2971                 if (*kiblnd_tunables.kib_nscheds > 0) {
2972                         nthrs = sched->ibs_nthreads_max;
2973                 } else {
2974                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2975                                                sched->ibs_cpt);
2976                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2977                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2978                 }
2979         } else {
2980                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2981                 /* increase one thread if there is new interface */
2982                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2983         }
2984
2985         for (i = 0; i < nthrs; i++) {
2986                 long    id;
2987                 char    name[20];
2988                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2989                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2990                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2991                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2992                 if (rc == 0)
2993                         continue;
2994
2995                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2996                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2997                 break;
2998         }
2999
3000         sched->ibs_nthreads += i;
3001         return rc;
3002 }
3003
3004 static int
3005 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3006 {
3007         int     cpt;
3008         int     rc;
3009         int     i;
3010
3011         for (i = 0; i < ncpts; i++) {
3012                 struct kib_sched_info *sched;
3013
3014                 cpt = (cpts == NULL) ? i : cpts[i];
3015                 sched = kiblnd_data.kib_scheds[cpt];
3016
3017                 if (!newdev && sched->ibs_nthreads > 0)
3018                         continue;
3019
3020                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3021                 if (rc != 0) {
3022                         CERROR("Failed to start scheduler threads for %s\n",
3023                                dev->ibd_ifname);
3024                         return rc;
3025                 }
3026         }
3027         return 0;
3028 }
3029
3030 static kib_dev_t *
3031 kiblnd_dev_search(char *ifname)
3032 {
3033         kib_dev_t       *alias = NULL;
3034         kib_dev_t       *dev;
3035         char            *colon;
3036         char            *colon2;
3037
3038         colon = strchr(ifname, ':');
3039         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3040                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3041                         return dev;
3042
3043                 if (alias != NULL)
3044                         continue;
3045
3046                 colon2 = strchr(dev->ibd_ifname, ':');
3047                 if (colon != NULL)
3048                         *colon = 0;
3049                 if (colon2 != NULL)
3050                         *colon2 = 0;
3051
3052                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3053                         alias = dev;
3054
3055                 if (colon != NULL)
3056                         *colon = ':';
3057                 if (colon2 != NULL)
3058                         *colon2 = ':';
3059         }
3060         return alias;
3061 }
3062
3063 static int
3064 kiblnd_startup (lnet_ni_t *ni)
3065 {
3066         char                     *ifname;
3067         kib_dev_t                *ibdev = NULL;
3068         kib_net_t                *net;
3069         struct timeval            tv;
3070         unsigned long             flags;
3071         int                       rc;
3072         int                       newdev;
3073
3074         LASSERT (ni->ni_lnd == &the_o2iblnd);
3075
3076         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3077                 rc = kiblnd_base_startup();
3078                 if (rc != 0)
3079                         return rc;
3080         }
3081
3082         LIBCFS_ALLOC(net, sizeof(*net));
3083         ni->ni_data = net;
3084         if (net == NULL)
3085                 goto failed;
3086
3087         do_gettimeofday(&tv);
3088         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3089
3090         kiblnd_tunables_setup(ni);
3091
3092         if (ni->ni_interfaces[0] != NULL) {
3093                 /* Use the IPoIB interface specified in 'networks=' */
3094
3095                 CLASSERT (LNET_MAX_INTERFACES > 1);
3096                 if (ni->ni_interfaces[1] != NULL) {
3097                         CERROR("Multiple interfaces not supported\n");
3098                         goto failed;
3099                 }
3100
3101                 ifname = ni->ni_interfaces[0];
3102         } else {
3103                 ifname = *kiblnd_tunables.kib_default_ipif;
3104         }
3105
3106         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3107                 CERROR("IPoIB interface name too long: %s\n", ifname);
3108                 goto failed;
3109         }
3110
3111         ibdev = kiblnd_dev_search(ifname);
3112
3113         newdev = ibdev == NULL;
3114         /* hmm...create kib_dev even for alias */
3115         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3116                 ibdev = kiblnd_create_dev(ifname);
3117
3118         if (ibdev == NULL)
3119                 goto failed;
3120
3121         net->ibn_dev = ibdev;
3122         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3123
3124         rc = kiblnd_dev_start_threads(ibdev, newdev,
3125                                       ni->ni_cpts, ni->ni_ncpts);
3126         if (rc != 0)
3127                 goto failed;
3128
3129         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3130         if (rc != 0) {
3131                 CERROR("Failed to initialize NI pools: %d\n", rc);
3132                 goto failed;
3133         }
3134
3135         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3136         ibdev->ibd_nnets++;
3137         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3138         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3139
3140         net->ibn_init = IBLND_INIT_ALL;
3141
3142         return 0;
3143
3144 failed:
3145         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3146                 kiblnd_destroy_dev(ibdev);
3147
3148         kiblnd_shutdown(ni);
3149
3150         CDEBUG(D_NET, "kiblnd_startup failed\n");
3151         return -ENETDOWN;
3152 }
3153
3154 static lnd_t the_o2iblnd = {
3155         .lnd_type       = O2IBLND,
3156         .lnd_startup    = kiblnd_startup,
3157         .lnd_shutdown   = kiblnd_shutdown,
3158         .lnd_ctl        = kiblnd_ctl,
3159         .lnd_query      = kiblnd_query,
3160         .lnd_send       = kiblnd_send,
3161         .lnd_recv       = kiblnd_recv,
3162 };
3163
3164 static void __exit ko2iblnd_exit(void)
3165 {
3166         lnet_unregister_lnd(&the_o2iblnd);
3167         kiblnd_tunables_fini();
3168 }
3169
3170 static int __init ko2iblnd_init(void)
3171 {
3172         int rc;
3173
3174         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3175         CLASSERT(offsetof(kib_msg_t,
3176                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3177                  IBLND_MSG_SIZE);
3178         CLASSERT(offsetof(kib_msg_t,
3179                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3180                  <= IBLND_MSG_SIZE);
3181
3182         rc = kiblnd_tunables_init();
3183         if (rc != 0)
3184                 return rc;
3185
3186         lnet_register_lnd(&the_o2iblnd);
3187
3188         return 0;
3189 }
3190
3191 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3192 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3193 MODULE_VERSION("2.8.0");
3194 MODULE_LICENSE("GPL");
3195
3196 module_init(ko2iblnd_init);
3197 module_exit(ko2iblnd_exit);