lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 kiblnd_cksum (void *ptr, int nob)
  50 {
  51         char  *c  = ptr;
  52         __u32  sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return (sum == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 340         peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
 341         peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 342         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 343
 344         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 345         INIT_LIST_HEAD(&peer->ibp_conns);
 346         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer (kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(kiblnd_peer_idle(peer));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 387         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head        *tmp;
 389         kib_peer_t              *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394                 LASSERT(!kiblnd_peer_idle(peer));
 395
 396                 if (peer->ibp_nid != nid)
 397                         continue;
 398
 399                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 400                        peer, libcfs_nid2str(nid),
 401                        atomic_read(&peer->ibp_refcount),
 402                        peer->ibp_version);
 403                 return peer;
 404         }
 405         return NULL;
 406 }
 407
 408 void
 409 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 410 {
 411         LASSERT(list_empty(&peer->ibp_conns));
 412
 413         LASSERT (kiblnd_peer_active(peer));
 414         list_del_init(&peer->ibp_list);
 415         /* lose peerlist's ref */
 416         kiblnd_peer_decref(peer);
 417 }
 418
 419 static int
 420 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 421                      lnet_nid_t *nidp, int *count)
 422 {
 423         kib_peer_t              *peer;
 424         struct list_head        *ptmp;
 425         int                      i;
 426         unsigned long            flags;
 427
 428         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 429
 430         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 431
 432                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 433
 434                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 435                         LASSERT(!kiblnd_peer_idle(peer));
 436
 437                         if (peer->ibp_ni != ni)
 438                                 continue;
 439
 440                         if (index-- > 0)
 441                                 continue;
 442
 443                         *nidp = peer->ibp_nid;
 444                         *count = atomic_read(&peer->ibp_refcount);
 445
 446                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 447                                                flags);
 448                         return 0;
 449                 }
 450         }
 451
 452         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 453         return -ENOENT;
 454 }
 455
 456 static void
 457 kiblnd_del_peer_locked (kib_peer_t *peer)
 458 {
 459         struct list_head        *ctmp;
 460         struct list_head        *cnxt;
 461         kib_conn_t              *conn;
 462
 463         if (list_empty(&peer->ibp_conns)) {
 464                 kiblnd_unlink_peer_locked(peer);
 465         } else {
 466                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 467                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 468
 469                         kiblnd_close_conn_locked(conn, 0);
 470                 }
 471                 /* NB closing peer's last conn unlinked it. */
 472         }
 473         /* NB peer now unlinked; might even be freed if the peer table had the
 474          * last ref on it. */
 475 }
 476
 477 static int
 478 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 479 {
 480         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 481         struct list_head        *ptmp;
 482         struct list_head        *pnxt;
 483         kib_peer_t              *peer;
 484         int                     lo;
 485         int                     hi;
 486         int                     i;
 487         unsigned long           flags;
 488         int                     rc = -ENOENT;
 489
 490         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 491
 492         if (nid != LNET_NID_ANY) {
 493                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 494         } else {
 495                 lo = 0;
 496                 hi = kiblnd_data.kib_peer_hash_size - 1;
 497         }
 498
 499         for (i = lo; i <= hi; i++) {
 500                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 501                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 502                         LASSERT(!kiblnd_peer_idle(peer));
 503
 504                         if (peer->ibp_ni != ni)
 505                                 continue;
 506
 507                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 508                                 continue;
 509
 510                         if (!list_empty(&peer->ibp_tx_queue)) {
 511                                 LASSERT(list_empty(&peer->ibp_conns));
 512
 513                                 list_splice_init(&peer->ibp_tx_queue,
 514                                                  &zombies);
 515                         }
 516
 517                         kiblnd_del_peer_locked(peer);
 518                         rc = 0;         /* matched something */
 519                 }
 520         }
 521
 522         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 523
 524         kiblnd_txlist_done(ni, &zombies, -EIO);
 525
 526         return rc;
 527 }
 528
 529 static kib_conn_t *
 530 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 531 {
 532         kib_peer_t              *peer;
 533         struct list_head        *ptmp;
 534         kib_conn_t              *conn;
 535         struct list_head        *ctmp;
 536         int                     i;
 537         unsigned long           flags;
 538
 539         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 540
 541         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 542                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 543
 544                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 545                         LASSERT(!kiblnd_peer_idle(peer));
 546
 547                         if (peer->ibp_ni != ni)
 548                                 continue;
 549
 550                         list_for_each(ctmp, &peer->ibp_conns) {
 551                                 if (index-- > 0)
 552                                         continue;
 553
 554                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 555                                 kiblnd_conn_addref(conn);
 556                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 557                                                        flags);
 558                                 return conn;
 559                         }
 560                 }
 561         }
 562
 563         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 564         return NULL;
 565 }
 566
 567 static void
 568 kiblnd_debug_rx (kib_rx_t *rx)
 569 {
 570         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 571                rx, rx->rx_status, rx->rx_msg->ibm_type,
 572                rx->rx_msg->ibm_credits);
 573 }
 574
 575 static void
 576 kiblnd_debug_tx (kib_tx_t *tx)
 577 {
 578         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 579                "cookie "LPX64" msg %s%s type %x cred %d\n",
 580                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 581                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 582                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 583                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 584                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 585 }
 586
 587 void
 588 kiblnd_debug_conn (kib_conn_t *conn)
 589 {
 590         struct list_head        *tmp;
 591         int                     i;
 592
 593         spin_lock(&conn->ibc_lock);
 594
 595         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 596                atomic_read(&conn->ibc_refcount), conn,
 597                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 598         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 599                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 600                conn->ibc_nsends_posted, conn->ibc_credits,
 601                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 602         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 603
 604         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 605         list_for_each(tmp, &conn->ibc_early_rxs)
 606                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 607
 608         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 609         list_for_each(tmp, &conn->ibc_tx_noops)
 610                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 611
 612         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 613         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 614                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 615
 616         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 617         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 618                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 621         list_for_each(tmp, &conn->ibc_tx_queue)
 622                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 623
 624         CDEBUG(D_CONSOLE, "   active_txs:\n");
 625         list_for_each(tmp, &conn->ibc_active_txs)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   rxs:\n");
 629         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 630                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 631
 632         spin_unlock(&conn->ibc_lock);
 633 }
 634
 635 int
 636 kiblnd_translate_mtu(int value)
 637 {
 638         switch (value) {
 639         default:
 640                 return -1;
 641         case 0:
 642                 return 0;
 643         case 256:
 644                 return IB_MTU_256;
 645         case 512:
 646                 return IB_MTU_512;
 647         case 1024:
 648                 return IB_MTU_1024;
 649         case 2048:
 650                 return IB_MTU_2048;
 651         case 4096:
 652                 return IB_MTU_4096;
 653         }
 654 }
 655
 656 static void
 657 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 658 {
 659         int           mtu;
 660
 661         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 662         if (cmid->route.path_rec == NULL)
 663                 return;
 664
 665         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 666         LASSERT (mtu >= 0);
 667         if (mtu != 0)
 668                 cmid->route.path_rec->mtu = mtu;
 669 }
 670
 671 static int
 672 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 673 {
 674         cpumask_t       *mask;
 675         int             vectors;
 676         int             off;
 677         int             i;
 678
 679         vectors = conn->ibc_cmid->device->num_comp_vectors;
 680         if (vectors <= 1)
 681                 return 0;
 682
 683         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 684
 685         /* hash NID to CPU id in this partition... */
 686         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 687         for_each_cpu(i, mask) {
 688                 if (off-- == 0)
 689                         return i % vectors;
 690         }
 691
 692         LBUG();
 693         return 1;
 694 }
 695
 696 kib_conn_t *
 697 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 698                    int state, int version)
 699 {
 700         /* CAVEAT EMPTOR:
 701          * If the new conn is created successfully it takes over the caller's
 702          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 703          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 704          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 705          * to destroy 'cmid' here since I'm called from the CM which still has
 706          * its ref on 'cmid'). */
 707         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 708         kib_net_t              *net = peer->ibp_ni->ni_data;
 709         kib_dev_t              *dev;
 710         struct ib_qp_init_attr *init_qp_attr;
 711         struct kib_sched_info   *sched;
 712 #ifdef HAVE_IB_CQ_INIT_ATTR
 713         struct ib_cq_init_attr  cq_attr = {};
 714 #endif
 715         kib_conn_t              *conn;
 716         struct ib_cq            *cq;
 717         unsigned long           flags;
 718         int                     cpt;
 719         int                     rc;
 720         int                     i;
 721
 722         LASSERT(net != NULL);
 723         LASSERT(!in_interrupt());
 724
 725         dev = net->ibn_dev;
 726
 727         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 728         sched = kiblnd_data.kib_scheds[cpt];
 729
 730         LASSERT(sched->ibs_nthreads > 0);
 731
 732         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 733                          sizeof(*init_qp_attr));
 734         if (init_qp_attr == NULL) {
 735                 CERROR("Can't allocate qp_attr for %s\n",
 736                        libcfs_nid2str(peer->ibp_nid));
 737                 goto failed_0;
 738         }
 739
 740         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 741         if (conn == NULL) {
 742                 CERROR("Can't allocate connection for %s\n",
 743                        libcfs_nid2str(peer->ibp_nid));
 744                 goto failed_1;
 745         }
 746
 747         conn->ibc_state = IBLND_CONN_INIT;
 748         conn->ibc_version = version;
 749         conn->ibc_peer = peer;                  /* I take the caller's ref */
 750         cmid->context = conn;                   /* for future CM callbacks */
 751         conn->ibc_cmid = cmid;
 752         conn->ibc_max_frags = peer->ibp_max_frags;
 753         conn->ibc_queue_depth = peer->ibp_queue_depth;
 754
 755         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 756         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 757         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 758         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 759         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 760         INIT_LIST_HEAD(&conn->ibc_active_txs);
 761         spin_lock_init(&conn->ibc_lock);
 762
 763         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 764                          sizeof(*conn->ibc_connvars));
 765         if (conn->ibc_connvars == NULL) {
 766                 CERROR("Can't allocate in-progress connection state\n");
 767                 goto failed_2;
 768         }
 769
 770         write_lock_irqsave(glock, flags);
 771         if (dev->ibd_failover) {
 772                 write_unlock_irqrestore(glock, flags);
 773                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 774                 goto failed_2;
 775         }
 776
 777         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 778                 /* wakeup failover thread and teardown connection */
 779                 if (kiblnd_dev_can_failover(dev)) {
 780                         list_add_tail(&dev->ibd_fail_list,
 781                                       &kiblnd_data.kib_failed_devs);
 782                         wake_up(&kiblnd_data.kib_failover_waitq);
 783                 }
 784
 785                 write_unlock_irqrestore(glock, flags);
 786                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 787                        cmid->device->name, dev->ibd_ifname);
 788                 goto failed_2;
 789         }
 790
 791         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 792         conn->ibc_hdev = dev->ibd_hdev;
 793
 794         kiblnd_setup_mtu_locked(cmid);
 795
 796         write_unlock_irqrestore(glock, flags);
 797
 798         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 799                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 800         if (conn->ibc_rxs == NULL) {
 801                 CERROR("Cannot allocate RX buffers\n");
 802                 goto failed_2;
 803         }
 804
 805         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 806                                 IBLND_RX_MSG_PAGES(conn));
 807         if (rc != 0)
 808                 goto failed_2;
 809
 810         kiblnd_map_rx_descs(conn);
 811
 812 #ifdef HAVE_IB_CQ_INIT_ATTR
 813         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 814         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 815         cq = ib_create_cq(cmid->device,
 816                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 817                           &cq_attr);
 818 #else
 819         cq = ib_create_cq(cmid->device,
 820                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 821                           IBLND_CQ_ENTRIES(conn),
 822                           kiblnd_get_completion_vector(conn, cpt));
 823 #endif
 824         if (IS_ERR(cq)) {
 825                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 826                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 827                 goto failed_2;
 828         }
 829
 830         conn->ibc_cq = cq;
 831
 832         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 833         if (rc != 0) {
 834                 CERROR("Can't request completion notification: %d\n", rc);
 835                 goto failed_2;
 836         }
 837
 838         init_qp_attr->event_handler = kiblnd_qp_event;
 839         init_qp_attr->qp_context = conn;
 840         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 841         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 842         init_qp_attr->cap.max_send_sge = 1;
 843         init_qp_attr->cap.max_recv_sge = 1;
 844         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 845         init_qp_attr->qp_type = IB_QPT_RC;
 846         init_qp_attr->send_cq = cq;
 847         init_qp_attr->recv_cq = cq;
 848
 849         conn->ibc_sched = sched;
 850
 851         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 852         if (rc != 0) {
 853                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 854                        rc, init_qp_attr->cap.max_send_wr,
 855                        init_qp_attr->cap.max_recv_wr);
 856                 goto failed_2;
 857         }
 858
 859         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 860
 861         /* 1 ref for caller and each rxmsg */
 862         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 863         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 864
 865         /* post receives */
 866         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 867                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 868                                     IBLND_POSTRX_NO_CREDIT);
 869                 if (rc != 0) {
 870                         CERROR("Can't post rxmsg: %d\n", rc);
 871
 872                         /* Make posted receives complete */
 873                         kiblnd_abort_receives(conn);
 874
 875                         /* correct # of posted buffers
 876                          * NB locking needed now I'm racing with completion */
 877                         spin_lock_irqsave(&sched->ibs_lock, flags);
 878                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 879                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 880
 881                         /* cmid will be destroyed by CM(ofed) after cm_callback
 882                          * returned, so we can't refer it anymore
 883                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 884                         rdma_destroy_qp(conn->ibc_cmid);
 885                         conn->ibc_cmid = NULL;
 886
 887                         /* Drop my own and unused rxbuffer refcounts */
 888                         while (i++ <= IBLND_RX_MSGS(conn))
 889                                 kiblnd_conn_decref(conn);
 890
 891                         return NULL;
 892                 }
 893         }
 894
 895         /* Init successful! */
 896         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 897                  state == IBLND_CONN_PASSIVE_WAIT);
 898         conn->ibc_state = state;
 899
 900         /* 1 more conn */
 901         atomic_inc(&net->ibn_nconns);
 902         return conn;
 903
 904  failed_2:
 905         kiblnd_destroy_conn(conn, true);
 906  failed_1:
 907         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 908  failed_0:
 909         return NULL;
 910 }
 911
 912 void
 913 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 914 {
 915         struct rdma_cm_id *cmid = conn->ibc_cmid;
 916         kib_peer_t        *peer = conn->ibc_peer;
 917         int                rc;
 918
 919         LASSERT (!in_interrupt());
 920         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 921         LASSERT(list_empty(&conn->ibc_early_rxs));
 922         LASSERT(list_empty(&conn->ibc_tx_noops));
 923         LASSERT(list_empty(&conn->ibc_tx_queue));
 924         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 925         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 926         LASSERT(list_empty(&conn->ibc_active_txs));
 927         LASSERT (conn->ibc_noops_posted == 0);
 928         LASSERT (conn->ibc_nsends_posted == 0);
 929
 930         switch (conn->ibc_state) {
 931         default:
 932                 /* conn must be completely disengaged from the network */
 933                 LBUG();
 934
 935         case IBLND_CONN_DISCONNECTED:
 936                 /* connvars should have been freed already */
 937                 LASSERT (conn->ibc_connvars == NULL);
 938                 break;
 939
 940         case IBLND_CONN_INIT:
 941                 break;
 942         }
 943
 944         /* conn->ibc_cmid might be destroyed by CM already */
 945         if (cmid != NULL && cmid->qp != NULL)
 946                 rdma_destroy_qp(cmid);
 947
 948         if (conn->ibc_cq != NULL) {
 949                 rc = ib_destroy_cq(conn->ibc_cq);
 950                 if (rc != 0)
 951                         CWARN("Error destroying CQ: %d\n", rc);
 952         }
 953
 954         if (conn->ibc_rx_pages != NULL)
 955                 kiblnd_unmap_rx_descs(conn);
 956
 957         if (conn->ibc_rxs != NULL) {
 958                 LIBCFS_FREE(conn->ibc_rxs,
 959                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 960         }
 961
 962         if (conn->ibc_connvars != NULL)
 963                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 964
 965         if (conn->ibc_hdev != NULL)
 966                 kiblnd_hdev_decref(conn->ibc_hdev);
 967
 968         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 969         if (conn->ibc_state != IBLND_CONN_INIT) {
 970                 kib_net_t *net = peer->ibp_ni->ni_data;
 971
 972                 kiblnd_peer_decref(peer);
 973                 rdma_destroy_id(cmid);
 974                 atomic_dec(&net->ibn_nconns);
 975         }
 976
 977         if (free_conn)
 978                 LIBCFS_FREE(conn, sizeof(*conn));
 979 }
 980
 981 int
 982 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 983 {
 984         kib_conn_t              *conn;
 985         struct list_head        *ctmp;
 986         struct list_head        *cnxt;
 987         int                     count = 0;
 988
 989         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 990                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 991
 992                 CDEBUG(D_NET, "Closing conn -> %s, "
 993                               "version: %x, reason: %d\n",
 994                        libcfs_nid2str(peer->ibp_nid),
 995                        conn->ibc_version, why);
 996
 997                 kiblnd_close_conn_locked(conn, why);
 998                 count++;
 999         }
1000
1001         return count;
1002 }
1003
1004 int
1005 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1006                                 int version, __u64 incarnation)
1007 {
1008         kib_conn_t              *conn;
1009         struct list_head        *ctmp;
1010         struct list_head        *cnxt;
1011         int                     count = 0;
1012
1013         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1014                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1015
1016                 if (conn->ibc_version     == version &&
1017                     conn->ibc_incarnation == incarnation)
1018                         continue;
1019
1020                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1021                               "incarnation:"LPX64"(%x, "LPX64")\n",
1022                        libcfs_nid2str(peer->ibp_nid),
1023                        conn->ibc_version, conn->ibc_incarnation,
1024                        version, incarnation);
1025
1026                 kiblnd_close_conn_locked(conn, -ESTALE);
1027                 count++;
1028         }
1029
1030         return count;
1031 }
1032
1033 static int
1034 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1035 {
1036         kib_peer_t              *peer;
1037         struct list_head        *ptmp;
1038         struct list_head        *pnxt;
1039         int                     lo;
1040         int                     hi;
1041         int                     i;
1042         unsigned long           flags;
1043         int                     count = 0;
1044
1045         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1046
1047         if (nid != LNET_NID_ANY)
1048                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1049         else {
1050                 lo = 0;
1051                 hi = kiblnd_data.kib_peer_hash_size - 1;
1052         }
1053
1054         for (i = lo; i <= hi; i++) {
1055                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1056
1057                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1058                         LASSERT(!kiblnd_peer_idle(peer));
1059
1060                         if (peer->ibp_ni != ni)
1061                                 continue;
1062
1063                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1064                                 continue;
1065
1066                         count += kiblnd_close_peer_conns_locked(peer, 0);
1067                 }
1068         }
1069
1070         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1071
1072         /* wildcards always succeed */
1073         if (nid == LNET_NID_ANY)
1074                 return 0;
1075
1076         return (count == 0) ? -ENOENT : 0;
1077 }
1078
1079 static int
1080 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1081 {
1082         struct libcfs_ioctl_data *data = arg;
1083         int                       rc = -EINVAL;
1084
1085         switch(cmd) {
1086         case IOC_LIBCFS_GET_PEER: {
1087                 lnet_nid_t   nid = 0;
1088                 int          count = 0;
1089
1090                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1091                                           &nid, &count);
1092                 data->ioc_nid    = nid;
1093                 data->ioc_count  = count;
1094                 break;
1095         }
1096
1097         case IOC_LIBCFS_DEL_PEER: {
1098                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1099                 break;
1100         }
1101         case IOC_LIBCFS_GET_CONN: {
1102                 kib_conn_t *conn;
1103
1104                 rc = 0;
1105                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1106                 if (conn == NULL) {
1107                         rc = -ENOENT;
1108                         break;
1109                 }
1110
1111                 LASSERT (conn->ibc_cmid != NULL);
1112                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1113                 if (conn->ibc_cmid->route.path_rec == NULL)
1114                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1115                 else
1116                         data->ioc_u32[0] =
1117                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1118                 kiblnd_conn_decref(conn);
1119                 break;
1120         }
1121         case IOC_LIBCFS_CLOSE_CONNECTION: {
1122                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1123                 break;
1124         }
1125
1126         default:
1127                 break;
1128         }
1129
1130         return rc;
1131 }
1132
1133 static void
1134 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1135 {
1136         cfs_time_t      last_alive = 0;
1137         cfs_time_t      now = cfs_time_current();
1138         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1139         kib_peer_t      *peer;
1140         unsigned long   flags;
1141
1142         read_lock_irqsave(glock, flags);
1143
1144         peer = kiblnd_find_peer_locked(nid);
1145         if (peer != NULL)
1146                 last_alive = peer->ibp_last_alive;
1147
1148         read_unlock_irqrestore(glock, flags);
1149
1150         if (last_alive != 0)
1151                 *when = last_alive;
1152
1153         /* peer is not persistent in hash, trigger peer creation
1154          * and connection establishment with a NULL tx */
1155         if (peer == NULL)
1156                 kiblnd_launch_tx(ni, NULL, nid);
1157
1158         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1159                libcfs_nid2str(nid), peer,
1160                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1161         return;
1162 }
1163
1164 static void
1165 kiblnd_free_pages(kib_pages_t *p)
1166 {
1167         int     npages = p->ibp_npages;
1168         int     i;
1169
1170         for (i = 0; i < npages; i++) {
1171                 if (p->ibp_pages[i] != NULL)
1172                         __free_page(p->ibp_pages[i]);
1173         }
1174
1175         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1176 }
1177
1178 int
1179 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1180 {
1181         kib_pages_t     *p;
1182         int             i;
1183
1184         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1185                          offsetof(kib_pages_t, ibp_pages[npages]));
1186         if (p == NULL) {
1187                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1188                 return -ENOMEM;
1189         }
1190
1191         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1192         p->ibp_npages = npages;
1193
1194         for (i = 0; i < npages; i++) {
1195                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1196                                                      GFP_NOFS);
1197                 if (p->ibp_pages[i] == NULL) {
1198                         CERROR("Can't allocate page %d of %d\n", i, npages);
1199                         kiblnd_free_pages(p);
1200                         return -ENOMEM;
1201                 }
1202         }
1203
1204         *pp = p;
1205         return 0;
1206 }
1207
1208 void
1209 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1210 {
1211         kib_rx_t *rx;
1212         int       i;
1213
1214         LASSERT (conn->ibc_rxs != NULL);
1215         LASSERT (conn->ibc_hdev != NULL);
1216
1217         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1218                 rx = &conn->ibc_rxs[i];
1219
1220                 LASSERT(rx->rx_nob >= 0); /* not posted */
1221
1222                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1223                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1224                                                           rx->rx_msgaddr),
1225                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1226         }
1227
1228         kiblnd_free_pages(conn->ibc_rx_pages);
1229
1230         conn->ibc_rx_pages = NULL;
1231 }
1232
1233 void
1234 kiblnd_map_rx_descs(kib_conn_t *conn)
1235 {
1236         kib_rx_t       *rx;
1237         struct page    *pg;
1238         int             pg_off;
1239         int             ipg;
1240         int             i;
1241
1242         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1243                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1244                 rx = &conn->ibc_rxs[i];
1245
1246                 rx->rx_conn = conn;
1247                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1248
1249                 rx->rx_msgaddr =
1250                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1251                                               rx->rx_msg, IBLND_MSG_SIZE,
1252                                               DMA_FROM_DEVICE);
1253                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1254                                                   rx->rx_msgaddr));
1255                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1256
1257                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1258                        i, rx->rx_msg, rx->rx_msgaddr,
1259                        (__u64)(page_to_phys(pg) + pg_off));
1260
1261                 pg_off += IBLND_MSG_SIZE;
1262                 LASSERT(pg_off <= PAGE_SIZE);
1263
1264                 if (pg_off == PAGE_SIZE) {
1265                         pg_off = 0;
1266                         ipg++;
1267                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1268                 }
1269         }
1270 }
1271
1272 static void
1273 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1274 {
1275         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1276         kib_tx_t       *tx;
1277         int             i;
1278
1279         LASSERT (tpo->tpo_pool.po_allocated == 0);
1280
1281         if (hdev == NULL)
1282                 return;
1283
1284         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1285                 tx = &tpo->tpo_tx_descs[i];
1286                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1287                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1288                                                           tx->tx_msgaddr),
1289                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1290         }
1291
1292         kiblnd_hdev_decref(hdev);
1293         tpo->tpo_hdev = NULL;
1294 }
1295
1296 static kib_hca_dev_t *
1297 kiblnd_current_hdev(kib_dev_t *dev)
1298 {
1299         kib_hca_dev_t *hdev;
1300         unsigned long  flags;
1301         int            i = 0;
1302
1303         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1304         while (dev->ibd_failover) {
1305                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1306                 if (i++ % 50 == 0)
1307                         CDEBUG(D_NET, "%s: Wait for failover\n",
1308                                dev->ibd_ifname);
1309                 set_current_state(TASK_INTERRUPTIBLE);
1310                 schedule_timeout(cfs_time_seconds(1) / 100);
1311
1312                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1313         }
1314
1315         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1316         hdev = dev->ibd_hdev;
1317
1318         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1319
1320         return hdev;
1321 }
1322
1323 static void
1324 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1325 {
1326         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1327         kib_pool_t     *pool  = &tpo->tpo_pool;
1328         kib_net_t      *net   = pool->po_owner->ps_net;
1329         kib_dev_t      *dev;
1330         struct page    *page;
1331         kib_tx_t       *tx;
1332         int             page_offset;
1333         int             ipage;
1334         int             i;
1335
1336         LASSERT (net != NULL);
1337
1338         dev = net->ibn_dev;
1339
1340         /* pre-mapped messages are not bigger than 1 page */
1341         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1342
1343         /* No fancy arithmetic when we do the buffer calculations */
1344         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1345
1346         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1347
1348         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1349                 page = txpgs->ibp_pages[ipage];
1350                 tx = &tpo->tpo_tx_descs[i];
1351
1352                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1353                                            page_offset);
1354
1355                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1356                                                        tx->tx_msg,
1357                                                        IBLND_MSG_SIZE,
1358                                                        DMA_TO_DEVICE);
1359                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1360                                                   tx->tx_msgaddr));
1361                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1362
1363                 list_add(&tx->tx_list, &pool->po_free_list);
1364
1365                 page_offset += IBLND_MSG_SIZE;
1366                 LASSERT(page_offset <= PAGE_SIZE);
1367
1368                 if (page_offset == PAGE_SIZE) {
1369                         page_offset = 0;
1370                         ipage++;
1371                         LASSERT(ipage <= txpgs->ibp_npages);
1372                 }
1373         }
1374 }
1375
1376 struct ib_mr *
1377 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1378                       int negotiated_nfrags)
1379 {
1380         __u16   nfrags = (negotiated_nfrags != -1) ?
1381           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1382
1383         LASSERT(hdev->ibh_mrs != NULL);
1384
1385         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1386             nfrags <= rd->rd_nfrags)
1387                 return NULL;
1388
1389         return hdev->ibh_mrs;
1390 }
1391
1392 static void
1393 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1394 {
1395         LASSERT(fpo->fpo_map_count == 0);
1396
1397         if (fpo->fpo_is_fmr) {
1398                 if (fpo->fmr.fpo_fmr_pool)
1399                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1400         } else {
1401                 struct kib_fast_reg_descriptor *frd, *tmp;
1402                 int i = 0;
1403
1404                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1405                                          frd_list) {
1406                         list_del(&frd->frd_list);
1407                         ib_free_fast_reg_page_list(frd->frd_frpl);
1408                         ib_dereg_mr(frd->frd_mr);
1409                         LIBCFS_FREE(frd, sizeof(*frd));
1410                         i++;
1411                 }
1412                 if (i < fpo->fast_reg.fpo_pool_size)
1413                         CERROR("FastReg pool still has %d regions registered\n",
1414                                 fpo->fast_reg.fpo_pool_size - i);
1415         }
1416
1417         if (fpo->fpo_hdev)
1418                 kiblnd_hdev_decref(fpo->fpo_hdev);
1419
1420         LIBCFS_FREE(fpo, sizeof(*fpo));
1421 }
1422
1423 static void
1424 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1425 {
1426         kib_fmr_pool_t *fpo, *tmp;
1427
1428         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1429                 list_del(&fpo->fpo_list);
1430                 kiblnd_destroy_fmr_pool(fpo);
1431         }
1432 }
1433
1434 static int kiblnd_fmr_pool_size(int ncpts)
1435 {
1436         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1437
1438         return max(IBLND_FMR_POOL, size);
1439 }
1440
1441 static int kiblnd_fmr_flush_trigger(int ncpts)
1442 {
1443         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1444
1445         return max(IBLND_FMR_POOL_FLUSH, size);
1446 }
1447
1448 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1449 {
1450         struct ib_fmr_pool_param param = {
1451                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1452                 .page_shift        = PAGE_SHIFT,
1453                 .access            = (IB_ACCESS_LOCAL_WRITE |
1454                                       IB_ACCESS_REMOTE_WRITE),
1455                 .pool_size         = fps->fps_pool_size,
1456                 .dirty_watermark   = fps->fps_flush_trigger,
1457                 .flush_function    = NULL,
1458                 .flush_arg         = NULL,
1459                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1460         int rc = 0;
1461
1462         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1463                                                    &param);
1464         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1465                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1466                 if (rc != -ENOSYS)
1467                         CERROR("Failed to create FMR pool: %d\n", rc);
1468                 else
1469                         CERROR("FMRs are not supported\n");
1470         }
1471
1472         return rc;
1473 }
1474
1475 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1476 {
1477         struct kib_fast_reg_descriptor *frd, *tmp;
1478         int i, rc;
1479
1480         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1481         fpo->fast_reg.fpo_pool_size = 0;
1482         for (i = 0; i < fps->fps_pool_size; i++) {
1483                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1484                                  sizeof(*frd));
1485                 if (!frd) {
1486                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1487                         rc = -ENOMEM;
1488                         goto out;
1489                 }
1490                 frd->frd_mr = NULL;
1491
1492                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1493                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1494                 if (IS_ERR(frd->frd_frpl)) {
1495                         rc = PTR_ERR(frd->frd_frpl);
1496                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1497                                 rc);
1498                         goto out_middle;
1499                 }
1500
1501                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1502                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1503                 if (IS_ERR(frd->frd_mr)) {
1504                         rc = PTR_ERR(frd->frd_mr);
1505                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1506                         goto out_middle;
1507                 }
1508
1509                 frd->frd_valid = true;
1510
1511                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1512                 fpo->fast_reg.fpo_pool_size++;
1513         }
1514
1515         return 0;
1516
1517 out_middle:
1518         if (frd->frd_mr)
1519                 ib_dereg_mr(frd->frd_mr);
1520         if (frd->frd_frpl)
1521                 ib_free_fast_reg_page_list(frd->frd_frpl);
1522         LIBCFS_FREE(frd, sizeof(*frd));
1523
1524 out:
1525         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1526                                  frd_list) {
1527                 list_del(&frd->frd_list);
1528                 ib_free_fast_reg_page_list(frd->frd_frpl);
1529                 ib_dereg_mr(frd->frd_mr);
1530                 LIBCFS_FREE(frd, sizeof(*frd));
1531         }
1532
1533         return rc;
1534 }
1535
1536 static int
1537 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1538 {
1539         struct ib_device_attr *dev_attr;
1540         kib_dev_t *dev = fps->fps_net->ibn_dev;
1541         kib_fmr_pool_t *fpo;
1542         int rc;
1543
1544         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1545         if (!dev_attr)
1546                 return -ENOMEM;
1547
1548         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1549         if (!fpo) {
1550                 rc = -ENOMEM;
1551                 goto out_dev_attr;
1552         }
1553
1554         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1555
1556         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1557         if (rc) {
1558                 CERROR("Query device failed for %s: %d\n",
1559                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1560                 goto out_dev_attr;
1561         }
1562
1563         /* Check for FMR or FastReg support */
1564         fpo->fpo_is_fmr = 0;
1565         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1566             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1567             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1568             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1569                 LCONSOLE_INFO("Using FMR for registration\n");
1570                 fpo->fpo_is_fmr = 1;
1571         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1572                 LCONSOLE_INFO("Using FastReg for registration\n");
1573         } else {
1574                 rc = -ENOSYS;
1575                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1576                 goto out_dev_attr;
1577         }
1578
1579         if (fpo->fpo_is_fmr)
1580                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1581         else
1582                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1583         if (rc)
1584                 goto out_fpo;
1585
1586         kfree(dev_attr);
1587         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1588         fpo->fpo_owner    = fps;
1589         *pp_fpo = fpo;
1590
1591         return 0;
1592
1593 out_fpo:
1594         kiblnd_hdev_decref(fpo->fpo_hdev);
1595         LIBCFS_FREE(fpo, sizeof(*fpo));
1596
1597 out_dev_attr:
1598         kfree(dev_attr);
1599
1600         return rc;
1601 }
1602
1603 static void
1604 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1605 {
1606         if (fps->fps_net == NULL) /* intialized? */
1607                 return;
1608
1609         spin_lock(&fps->fps_lock);
1610
1611         while (!list_empty(&fps->fps_pool_list)) {
1612                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1613                                                  kib_fmr_pool_t, fpo_list);
1614                 fpo->fpo_failed = 1;
1615                 list_del(&fpo->fpo_list);
1616                 if (fpo->fpo_map_count == 0)
1617                         list_add(&fpo->fpo_list, zombies);
1618                 else
1619                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1620         }
1621
1622         spin_unlock(&fps->fps_lock);
1623 }
1624
1625 static void
1626 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1627 {
1628         if (fps->fps_net != NULL) { /* initialized? */
1629                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1630                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1631         }
1632 }
1633
1634 static int
1635 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1636                         int pool_size, int flush_trigger)
1637 {
1638         kib_fmr_pool_t *fpo;
1639         int             rc;
1640
1641         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1642
1643         fps->fps_net = net;
1644         fps->fps_cpt = cpt;
1645         fps->fps_pool_size = pool_size;
1646         fps->fps_flush_trigger = flush_trigger;
1647         spin_lock_init(&fps->fps_lock);
1648         INIT_LIST_HEAD(&fps->fps_pool_list);
1649         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1650
1651         rc = kiblnd_create_fmr_pool(fps, &fpo);
1652         if (rc == 0)
1653                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1654
1655         return rc;
1656 }
1657
1658 static int
1659 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1660 {
1661         if (fpo->fpo_map_count != 0) /* still in use */
1662                 return 0;
1663         if (fpo->fpo_failed)
1664                 return 1;
1665         return cfs_time_aftereq(now, fpo->fpo_deadline);
1666 }
1667
1668 void
1669 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1670 {
1671         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1672         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1673         kib_fmr_poolset_t *fps;
1674         cfs_time_t         now = cfs_time_current();
1675         kib_fmr_pool_t    *tmp;
1676         int                rc;
1677
1678         if (!fpo)
1679                 return;
1680
1681         fps = fpo->fpo_owner;
1682         if (fpo->fpo_is_fmr) {
1683                 if (fmr->fmr_pfmr) {
1684                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1685                         LASSERT(!rc);
1686                         fmr->fmr_pfmr = NULL;
1687                 }
1688
1689                 if (status) {
1690                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1691                         LASSERT(!rc);
1692                 }
1693         } else {
1694                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1695
1696                 if (frd) {
1697                         frd->frd_valid = false;
1698                         spin_lock(&fps->fps_lock);
1699                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1700                         spin_unlock(&fps->fps_lock);
1701                         fmr->fmr_frd = NULL;
1702                 }
1703         }
1704         fmr->fmr_pool = NULL;
1705
1706         spin_lock(&fps->fps_lock);
1707         fpo->fpo_map_count--;   /* decref the pool */
1708
1709         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1710                 /* the first pool is persistent */
1711                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1712                         continue;
1713
1714                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1715                         list_move(&fpo->fpo_list, &zombies);
1716                         fps->fps_version++;
1717                 }
1718         }
1719         spin_unlock(&fps->fps_lock);
1720
1721         if (!list_empty(&zombies))
1722                 kiblnd_destroy_fmr_pool_list(&zombies);
1723 }
1724
1725 int
1726 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1727                     __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr)
1728 {
1729         kib_fmr_pool_t *fpo;
1730         __u64 version;
1731         int rc;
1732
1733 again:
1734         spin_lock(&fps->fps_lock);
1735         version = fps->fps_version;
1736         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1737                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1738                 fpo->fpo_map_count++;
1739
1740                 if (fpo->fpo_is_fmr) {
1741                         struct ib_pool_fmr *pfmr;
1742
1743                         spin_unlock(&fps->fps_lock);
1744                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1745                                                     pages, npages, iov);
1746                         if (likely(!IS_ERR(pfmr))) {
1747                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1748                                                       : pfmr->fmr->lkey;
1749                                 fmr->fmr_frd  = NULL;
1750                                 fmr->fmr_pfmr = pfmr;
1751                                 fmr->fmr_pool = fpo;
1752                                 return 0;
1753                         }
1754                         rc = PTR_ERR(pfmr);
1755                 } else {
1756                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1757                                 struct ib_send_wr *wr;
1758                                 struct kib_fast_reg_descriptor *frd;
1759                                 struct ib_fast_reg_page_list *frpl;
1760                                 struct ib_mr *mr;
1761
1762                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1763                                                         struct kib_fast_reg_descriptor,
1764                                                         frd_list);
1765                                 list_del(&frd->frd_list);
1766                                 spin_unlock(&fps->fps_lock);
1767
1768                                 frpl = frd->frd_frpl;
1769                                 mr   = frd->frd_mr;
1770
1771                                 if (!frd->frd_valid) {
1772                                         struct ib_send_wr *inv_wr;
1773                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1774
1775                                         inv_wr = &frd->frd_inv_wr;
1776                                         memset(inv_wr, 0, sizeof(*inv_wr));
1777                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1778                                         inv_wr->wr_id = IBLND_WID_MR;
1779                                         inv_wr->ex.invalidate_rkey = key;
1780
1781                                         /* Bump the key */
1782                                         key = ib_inc_rkey(key);
1783                                         ib_update_fast_reg_key(mr, key);
1784                                 }
1785
1786                                 LASSERT(npages <= frpl->max_page_list_len);
1787                                 memcpy(frpl->page_list, pages,
1788                                         sizeof(*pages) * npages);
1789
1790                                 /* Prepare FastReg WR */
1791                                 wr = &frd->frd_fastreg_wr;
1792                                 memset(wr, 0, sizeof(*wr));
1793                                 wr->opcode = IB_WR_FAST_REG_MR;
1794                                 wr->wr_id = IBLND_WID_MR;
1795                                 wr->wr.fast_reg.iova_start = iov;
1796                                 wr->wr.fast_reg.page_list  = frpl;
1797                                 wr->wr.fast_reg.page_list_len = npages;
1798                                 wr->wr.fast_reg.page_shift = PAGE_SHIFT;
1799                                 wr->wr.fast_reg.length = nob;
1800                                 wr->wr.fast_reg.rkey = is_rx ? mr->rkey
1801                                                              : mr->lkey;
1802                                 wr->wr.fast_reg.access_flags =
1803                                                 (IB_ACCESS_LOCAL_WRITE |
1804                                                  IB_ACCESS_REMOTE_WRITE);
1805
1806                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1807                                 fmr->fmr_frd  = frd;
1808                                 fmr->fmr_pfmr = NULL;
1809                                 fmr->fmr_pool = fpo;
1810                                 return 0;
1811                         }
1812                         spin_unlock(&fps->fps_lock);
1813                         rc = -EBUSY;
1814                 }
1815
1816                 spin_lock(&fps->fps_lock);
1817                 fpo->fpo_map_count--;
1818                 if (rc != -EAGAIN) {
1819                         spin_unlock(&fps->fps_lock);
1820                         return rc;
1821                 }
1822
1823                 /* EAGAIN and ... */
1824                 if (version != fps->fps_version) {
1825                         spin_unlock(&fps->fps_lock);
1826                         goto again;
1827                 }
1828         }
1829
1830         if (fps->fps_increasing) {
1831                 spin_unlock(&fps->fps_lock);
1832                 CDEBUG(D_NET, "Another thread is allocating new "
1833                        "FMR pool, waiting for her to complete\n");
1834                 schedule();
1835                 goto again;
1836
1837         }
1838
1839         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1840                 /* someone failed recently */
1841                 spin_unlock(&fps->fps_lock);
1842                 return -EAGAIN;
1843         }
1844
1845         fps->fps_increasing = 1;
1846         spin_unlock(&fps->fps_lock);
1847
1848         CDEBUG(D_NET, "Allocate new FMR pool\n");
1849         rc = kiblnd_create_fmr_pool(fps, &fpo);
1850         spin_lock(&fps->fps_lock);
1851         fps->fps_increasing = 0;
1852         if (rc == 0) {
1853                 fps->fps_version++;
1854                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1855         } else {
1856                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1857         }
1858         spin_unlock(&fps->fps_lock);
1859
1860         goto again;
1861 }
1862
1863 static void
1864 kiblnd_fini_pool(kib_pool_t *pool)
1865 {
1866         LASSERT(list_empty(&pool->po_free_list));
1867         LASSERT(pool->po_allocated == 0);
1868
1869         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1870 }
1871
1872 static void
1873 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1874 {
1875         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1876
1877         memset(pool, 0, sizeof(kib_pool_t));
1878         INIT_LIST_HEAD(&pool->po_free_list);
1879         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1880         pool->po_owner    = ps;
1881         pool->po_size     = size;
1882 }
1883
1884 static void
1885 kiblnd_destroy_pool_list(struct list_head *head)
1886 {
1887         kib_pool_t *pool;
1888
1889         while (!list_empty(head)) {
1890                 pool = list_entry(head->next, kib_pool_t, po_list);
1891                 list_del(&pool->po_list);
1892
1893                 LASSERT(pool->po_owner != NULL);
1894                 pool->po_owner->ps_pool_destroy(pool);
1895         }
1896 }
1897
1898 static void
1899 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1900 {
1901         if (ps->ps_net == NULL) /* intialized? */
1902                 return;
1903
1904         spin_lock(&ps->ps_lock);
1905         while (!list_empty(&ps->ps_pool_list)) {
1906                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1907                                             kib_pool_t, po_list);
1908                 po->po_failed = 1;
1909                 list_del(&po->po_list);
1910                 if (po->po_allocated == 0)
1911                         list_add(&po->po_list, zombies);
1912                 else
1913                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1914         }
1915         spin_unlock(&ps->ps_lock);
1916 }
1917
1918 static void
1919 kiblnd_fini_poolset(kib_poolset_t *ps)
1920 {
1921         if (ps->ps_net != NULL) { /* initialized? */
1922                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1923                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1924         }
1925 }
1926
1927 static int
1928 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1929                     kib_net_t *net, char *name, int size,
1930                     kib_ps_pool_create_t po_create,
1931                     kib_ps_pool_destroy_t po_destroy,
1932                     kib_ps_node_init_t nd_init,
1933                     kib_ps_node_fini_t nd_fini)
1934 {
1935         kib_pool_t      *pool;
1936         int             rc;
1937
1938         memset(ps, 0, sizeof(kib_poolset_t));
1939
1940         ps->ps_cpt          = cpt;
1941         ps->ps_net          = net;
1942         ps->ps_pool_create  = po_create;
1943         ps->ps_pool_destroy = po_destroy;
1944         ps->ps_node_init    = nd_init;
1945         ps->ps_node_fini    = nd_fini;
1946         ps->ps_pool_size    = size;
1947         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1948             >= sizeof(ps->ps_name))
1949                 return -E2BIG;
1950         spin_lock_init(&ps->ps_lock);
1951         INIT_LIST_HEAD(&ps->ps_pool_list);
1952         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1953
1954         rc = ps->ps_pool_create(ps, size, &pool);
1955         if (rc == 0)
1956                 list_add(&pool->po_list, &ps->ps_pool_list);
1957         else
1958                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1959
1960         return rc;
1961 }
1962
1963 static int
1964 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1965 {
1966         if (pool->po_allocated != 0) /* still in use */
1967                 return 0;
1968         if (pool->po_failed)
1969                 return 1;
1970         return cfs_time_aftereq(now, pool->po_deadline);
1971 }
1972
1973 void
1974 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1975 {
1976         struct list_head zombies = LIST_HEAD_INIT(zombies);
1977         kib_poolset_t   *ps = pool->po_owner;
1978         kib_pool_t      *tmp;
1979         cfs_time_t       now = cfs_time_current();
1980
1981         spin_lock(&ps->ps_lock);
1982
1983         if (ps->ps_node_fini != NULL)
1984                 ps->ps_node_fini(pool, node);
1985
1986         LASSERT(pool->po_allocated > 0);
1987         list_add(node, &pool->po_free_list);
1988         pool->po_allocated--;
1989
1990         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1991                 /* the first pool is persistent */
1992                 if (ps->ps_pool_list.next == &pool->po_list)
1993                         continue;
1994
1995                 if (kiblnd_pool_is_idle(pool, now))
1996                         list_move(&pool->po_list, &zombies);
1997         }
1998         spin_unlock(&ps->ps_lock);
1999
2000         if (!list_empty(&zombies))
2001                 kiblnd_destroy_pool_list(&zombies);
2002 }
2003
2004 struct list_head *
2005 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2006 {
2007         struct list_head        *node;
2008         kib_pool_t              *pool;
2009         int                     rc;
2010         unsigned int            interval = 1;
2011         cfs_time_t              time_before;
2012         unsigned int            trips = 0;
2013
2014 again:
2015         spin_lock(&ps->ps_lock);
2016         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2017                 if (list_empty(&pool->po_free_list))
2018                         continue;
2019
2020                 pool->po_allocated++;
2021                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2022                 node = pool->po_free_list.next;
2023                 list_del(node);
2024
2025                 if (ps->ps_node_init != NULL) {
2026                         /* still hold the lock */
2027                         ps->ps_node_init(pool, node);
2028                 }
2029                 spin_unlock(&ps->ps_lock);
2030                 return node;
2031         }
2032
2033         /* no available tx pool and ... */
2034         if (ps->ps_increasing) {
2035                 /* another thread is allocating a new pool */
2036                 spin_unlock(&ps->ps_lock);
2037                 trips++;
2038                 CDEBUG(D_NET, "Another thread is allocating new "
2039                        "%s pool, waiting %d HZs for her to complete."
2040                        "trips = %d\n",
2041                        ps->ps_name, interval, trips);
2042
2043                 set_current_state(TASK_INTERRUPTIBLE);
2044                 schedule_timeout(interval);
2045                 if (interval < cfs_time_seconds(1))
2046                         interval *= 2;
2047
2048                 goto again;
2049         }
2050
2051         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2052                 /* someone failed recently */
2053                 spin_unlock(&ps->ps_lock);
2054                 return NULL;
2055         }
2056
2057         ps->ps_increasing = 1;
2058         spin_unlock(&ps->ps_lock);
2059
2060         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2061         time_before = cfs_time_current();
2062         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2063         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2064                cfs_time_current() - time_before);
2065
2066         spin_lock(&ps->ps_lock);
2067         ps->ps_increasing = 0;
2068         if (rc == 0) {
2069                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2070         } else {
2071                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2072                 CERROR("Can't allocate new %s pool because out of memory\n",
2073                        ps->ps_name);
2074         }
2075         spin_unlock(&ps->ps_lock);
2076
2077         goto again;
2078 }
2079
2080 static void
2081 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2082 {
2083         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2084         int             i;
2085
2086         LASSERT (pool->po_allocated == 0);
2087
2088         if (tpo->tpo_tx_pages != NULL) {
2089                 kiblnd_unmap_tx_pool(tpo);
2090                 kiblnd_free_pages(tpo->tpo_tx_pages);
2091         }
2092
2093         if (tpo->tpo_tx_descs == NULL)
2094                 goto out;
2095
2096         for (i = 0; i < pool->po_size; i++) {
2097                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2098
2099                 list_del(&tx->tx_list);
2100                 if (tx->tx_pages != NULL)
2101                         LIBCFS_FREE(tx->tx_pages,
2102                                     LNET_MAX_IOV *
2103                                     sizeof(*tx->tx_pages));
2104                 if (tx->tx_frags != NULL)
2105                         LIBCFS_FREE(tx->tx_frags,
2106                                     IBLND_MAX_RDMA_FRAGS *
2107                                             sizeof(*tx->tx_frags));
2108                 if (tx->tx_wrq != NULL)
2109                         LIBCFS_FREE(tx->tx_wrq,
2110                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2111                                     sizeof(*tx->tx_wrq));
2112                 if (tx->tx_sge != NULL)
2113                         LIBCFS_FREE(tx->tx_sge,
2114                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2115                                     sizeof(*tx->tx_sge));
2116                 if (tx->tx_rd != NULL)
2117                         LIBCFS_FREE(tx->tx_rd,
2118                                     offsetof(kib_rdma_desc_t,
2119                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2120         }
2121
2122         LIBCFS_FREE(tpo->tpo_tx_descs,
2123                     pool->po_size * sizeof(kib_tx_t));
2124 out:
2125         kiblnd_fini_pool(pool);
2126         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2127 }
2128
2129 static int kiblnd_tx_pool_size(int ncpts)
2130 {
2131         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2132
2133         return max(IBLND_TX_POOL, ntx);
2134 }
2135
2136 static int
2137 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2138 {
2139         int            i;
2140         int            npg;
2141         kib_pool_t    *pool;
2142         kib_tx_pool_t *tpo;
2143
2144         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2145         if (tpo == NULL) {
2146                 CERROR("Failed to allocate TX pool\n");
2147                 return -ENOMEM;
2148         }
2149
2150         pool = &tpo->tpo_pool;
2151         kiblnd_init_pool(ps, pool, size);
2152         tpo->tpo_tx_descs = NULL;
2153         tpo->tpo_tx_pages = NULL;
2154
2155         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2156         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2157                 CERROR("Can't allocate tx pages: %d\n", npg);
2158                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2159                 return -ENOMEM;
2160         }
2161
2162         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2163                          size * sizeof(kib_tx_t));
2164         if (tpo->tpo_tx_descs == NULL) {
2165                 CERROR("Can't allocate %d tx descriptors\n", size);
2166                 ps->ps_pool_destroy(pool);
2167                 return -ENOMEM;
2168         }
2169
2170         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2171
2172         for (i = 0; i < size; i++) {
2173                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2174
2175                 tx->tx_pool = tpo;
2176                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2177                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2178                                          lnet_cpt_table(), ps->ps_cpt,
2179                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2180                         if (tx->tx_pages == NULL)
2181                                 break;
2182                 }
2183
2184                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2185                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2186                 if (tx->tx_frags == NULL)
2187                         break;
2188
2189                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2190
2191                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2192                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2193                                  sizeof(*tx->tx_wrq));
2194                 if (tx->tx_wrq == NULL)
2195                         break;
2196
2197                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2198                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2199                                  sizeof(*tx->tx_sge));
2200                 if (tx->tx_sge == NULL)
2201                         break;
2202
2203                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2204                                  offsetof(kib_rdma_desc_t,
2205                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2206                 if (tx->tx_rd == NULL)
2207                         break;
2208         }
2209
2210         if (i == size) {
2211                 kiblnd_map_tx_pool(tpo);
2212                 *pp_po = pool;
2213                 return 0;
2214         }
2215
2216         ps->ps_pool_destroy(pool);
2217         return -ENOMEM;
2218 }
2219
2220 static void
2221 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2222 {
2223         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2224                                              tps_poolset);
2225         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2226
2227         tx->tx_cookie = tps->tps_next_tx_cookie++;
2228 }
2229
2230 static void
2231 kiblnd_net_fini_pools(kib_net_t *net)
2232 {
2233         int     i;
2234
2235         cfs_cpt_for_each(i, lnet_cpt_table()) {
2236                 kib_tx_poolset_t        *tps;
2237                 kib_fmr_poolset_t       *fps;
2238
2239                 if (net->ibn_tx_ps != NULL) {
2240                         tps = net->ibn_tx_ps[i];
2241                         kiblnd_fini_poolset(&tps->tps_poolset);
2242                 }
2243
2244                 if (net->ibn_fmr_ps != NULL) {
2245                         fps = net->ibn_fmr_ps[i];
2246                         kiblnd_fini_fmr_poolset(fps);
2247                 }
2248         }
2249
2250         if (net->ibn_tx_ps != NULL) {
2251                 cfs_percpt_free(net->ibn_tx_ps);
2252                 net->ibn_tx_ps = NULL;
2253         }
2254
2255         if (net->ibn_fmr_ps != NULL) {
2256                 cfs_percpt_free(net->ibn_fmr_ps);
2257                 net->ibn_fmr_ps = NULL;
2258         }
2259 }
2260
2261 static int
2262 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2263 {
2264         unsigned long   flags;
2265         int             cpt;
2266         int             rc;
2267         int             i;
2268
2269         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2270         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2271                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2272                                            flags);
2273                 goto create_tx_pool;
2274         }
2275
2276         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2277
2278         if (*kiblnd_tunables.kib_fmr_pool_size <
2279             *kiblnd_tunables.kib_ntx / 4) {
2280                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2281                        *kiblnd_tunables.kib_fmr_pool_size,
2282                        *kiblnd_tunables.kib_ntx / 4);
2283                 rc = -EINVAL;
2284                 goto failed;
2285         }
2286
2287         /* TX pool must be created later than FMR, see LU-2268
2288          * for details */
2289         LASSERT(net->ibn_tx_ps == NULL);
2290
2291         /* premapping can fail if ibd_nmr > 1, so we always create
2292          * FMR pool and map-on-demand if premapping failed */
2293
2294         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2295                                            sizeof(kib_fmr_poolset_t));
2296         if (net->ibn_fmr_ps == NULL) {
2297                 CERROR("Failed to allocate FMR pool array\n");
2298                 rc = -ENOMEM;
2299                 goto failed;
2300         }
2301
2302         for (i = 0; i < ncpts; i++) {
2303                 cpt = (cpts == NULL) ? i : cpts[i];
2304                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2305                                              kiblnd_fmr_pool_size(ncpts),
2306                                              kiblnd_fmr_flush_trigger(ncpts));
2307                 if (rc != 0) {
2308                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2309                                cpt, rc);
2310                         goto failed;
2311                 }
2312         }
2313
2314         if (i > 0)
2315                 LASSERT(i == ncpts);
2316
2317  create_tx_pool:
2318         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2319                                           sizeof(kib_tx_poolset_t));
2320         if (net->ibn_tx_ps == NULL) {
2321                 CERROR("Failed to allocate tx pool array\n");
2322                 rc = -ENOMEM;
2323                 goto failed;
2324         }
2325
2326         for (i = 0; i < ncpts; i++) {
2327                 cpt = (cpts == NULL) ? i : cpts[i];
2328                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2329                                          cpt, net, "TX",
2330                                          kiblnd_tx_pool_size(ncpts),
2331                                          kiblnd_create_tx_pool,
2332                                          kiblnd_destroy_tx_pool,
2333                                          kiblnd_tx_init, NULL);
2334                 if (rc != 0) {
2335                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2336                                cpt, rc);
2337                         goto failed;
2338                 }
2339         }
2340
2341         return 0;
2342  failed:
2343         kiblnd_net_fini_pools(net);
2344         LASSERT(rc != 0);
2345         return rc;
2346 }
2347
2348 static int
2349 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2350 {
2351         struct ib_device_attr *attr;
2352         int                    rc;
2353
2354         /* It's safe to assume a HCA can handle a page size
2355          * matching that of the native system */
2356         hdev->ibh_page_shift = PAGE_SHIFT;
2357         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2358         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2359
2360         LIBCFS_ALLOC(attr, sizeof(*attr));
2361         if (attr == NULL) {
2362                 CERROR("Out of memory\n");
2363                 return -ENOMEM;
2364         }
2365
2366         rc = ib_query_device(hdev->ibh_ibdev, attr);
2367         if (rc == 0)
2368                 hdev->ibh_mr_size = attr->max_mr_size;
2369
2370         LIBCFS_FREE(attr, sizeof(*attr));
2371
2372         if (rc != 0) {
2373                 CERROR("Failed to query IB device: %d\n", rc);
2374                 return rc;
2375         }
2376
2377         if (hdev->ibh_mr_size == ~0ULL) {
2378                 hdev->ibh_mr_shift = 64;
2379                 return 0;
2380         }
2381
2382         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2383         return -EINVAL;
2384 }
2385
2386 static void
2387 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2388 {
2389         if (hdev->ibh_mrs == NULL)
2390                 return;
2391
2392         ib_dereg_mr(hdev->ibh_mrs);
2393
2394         hdev->ibh_mrs = NULL;
2395 }
2396
2397 void
2398 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2399 {
2400         kiblnd_hdev_cleanup_mrs(hdev);
2401
2402         if (hdev->ibh_pd != NULL)
2403                 ib_dealloc_pd(hdev->ibh_pd);
2404
2405         if (hdev->ibh_cmid != NULL)
2406                 rdma_destroy_id(hdev->ibh_cmid);
2407
2408         LIBCFS_FREE(hdev, sizeof(*hdev));
2409 }
2410
2411 static int
2412 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2413 {
2414         struct ib_mr *mr;
2415         int           rc;
2416         int           acflags = IB_ACCESS_LOCAL_WRITE |
2417                                 IB_ACCESS_REMOTE_WRITE;
2418
2419         rc = kiblnd_hdev_get_attr(hdev);
2420         if (rc != 0)
2421                 return rc;
2422
2423         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2424         if (IS_ERR(mr)) {
2425                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2426                 kiblnd_hdev_cleanup_mrs(hdev);
2427                 return PTR_ERR(mr);
2428         }
2429
2430         hdev->ibh_mrs = mr;
2431
2432         return 0;
2433 }
2434
2435 static int
2436 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2437 {       /* DUMMY */
2438         return 0;
2439 }
2440
2441 static int
2442 kiblnd_dev_need_failover(kib_dev_t *dev)
2443 {
2444         struct rdma_cm_id  *cmid;
2445         struct sockaddr_in  srcaddr;
2446         struct sockaddr_in  dstaddr;
2447         int                 rc;
2448
2449         if (dev->ibd_hdev == NULL || /* initializing */
2450             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2451             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2452                 return 1;
2453
2454         /* XXX: it's UGLY, but I don't have better way to find
2455          * ib-bonding HCA failover because:
2456          *
2457          * a. no reliable CM event for HCA failover...
2458          * b. no OFED API to get ib_device for current net_device...
2459          *
2460          * We have only two choices at this point:
2461          *
2462          * a. rdma_bind_addr(), it will conflict with listener cmid
2463          * b. rdma_resolve_addr() to zero addr */
2464         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2465                                      IB_QPT_RC);
2466         if (IS_ERR(cmid)) {
2467                 rc = PTR_ERR(cmid);
2468                 CERROR("Failed to create cmid for failover: %d\n", rc);
2469                 return rc;
2470         }
2471
2472         memset(&srcaddr, 0, sizeof(srcaddr));
2473         srcaddr.sin_family      = AF_INET;
2474         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2475
2476         memset(&dstaddr, 0, sizeof(dstaddr));
2477         dstaddr.sin_family = AF_INET;
2478         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2479                                (struct sockaddr *)&dstaddr, 1);
2480         if (rc != 0 || cmid->device == NULL) {
2481                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2482                        dev->ibd_ifname, &dev->ibd_ifip,
2483                        cmid->device, rc);
2484                 rdma_destroy_id(cmid);
2485                 return rc;
2486         }
2487
2488         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2489         rdma_destroy_id(cmid);
2490         return rc;
2491 }
2492
2493 int
2494 kiblnd_dev_failover(kib_dev_t *dev)
2495 {
2496         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2497         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2498         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2499         struct rdma_cm_id  *cmid  = NULL;
2500         kib_hca_dev_t      *hdev  = NULL;
2501         kib_hca_dev_t      *old;
2502         struct ib_pd       *pd;
2503         kib_net_t          *net;
2504         struct sockaddr_in  addr;
2505         unsigned long       flags;
2506         int                 rc = 0;
2507         int                 i;
2508
2509         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2510                  dev->ibd_can_failover ||
2511                  dev->ibd_hdev == NULL);
2512
2513         rc = kiblnd_dev_need_failover(dev);
2514         if (rc <= 0)
2515                 goto out;
2516
2517         if (dev->ibd_hdev != NULL &&
2518             dev->ibd_hdev->ibh_cmid != NULL) {
2519                 /* XXX it's not good to close old listener at here,
2520                  * because we can fail to create new listener.
2521                  * But we have to close it now, otherwise rdma_bind_addr
2522                  * will return EADDRINUSE... How crap! */
2523                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2524
2525                 cmid = dev->ibd_hdev->ibh_cmid;
2526                 /* make next schedule of kiblnd_dev_need_failover()
2527                  * return 1 for me */
2528                 dev->ibd_hdev->ibh_cmid  = NULL;
2529                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2530
2531                 rdma_destroy_id(cmid);
2532         }
2533
2534         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2535                                      IB_QPT_RC);
2536         if (IS_ERR(cmid)) {
2537                 rc = PTR_ERR(cmid);
2538                 CERROR("Failed to create cmid for failover: %d\n", rc);
2539                 goto out;
2540         }
2541
2542         memset(&addr, 0, sizeof(addr));
2543         addr.sin_family      = AF_INET;
2544         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2545         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2546
2547         /* Bind to failover device or port */
2548         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2549         if (rc != 0 || cmid->device == NULL) {
2550                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2551                        dev->ibd_ifname, &dev->ibd_ifip,
2552                        cmid->device, rc);
2553                 rdma_destroy_id(cmid);
2554                 goto out;
2555         }
2556
2557         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2558         if (hdev == NULL) {
2559                 CERROR("Failed to allocate kib_hca_dev\n");
2560                 rdma_destroy_id(cmid);
2561                 rc = -ENOMEM;
2562                 goto out;
2563         }
2564
2565         atomic_set(&hdev->ibh_ref, 1);
2566         hdev->ibh_dev   = dev;
2567         hdev->ibh_cmid  = cmid;
2568         hdev->ibh_ibdev = cmid->device;
2569
2570         pd = ib_alloc_pd(cmid->device);
2571         if (IS_ERR(pd)) {
2572                 rc = PTR_ERR(pd);
2573                 CERROR("Can't allocate PD: %d\n", rc);
2574                 goto out;
2575         }
2576
2577         hdev->ibh_pd = pd;
2578
2579         rc = rdma_listen(cmid, 0);
2580         if (rc != 0) {
2581                 CERROR("Can't start new listener: %d\n", rc);
2582                 goto out;
2583         }
2584
2585         rc = kiblnd_hdev_setup_mrs(hdev);
2586         if (rc != 0) {
2587                 CERROR("Can't setup device: %d\n", rc);
2588                 goto out;
2589         }
2590
2591         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2592
2593         old = dev->ibd_hdev;
2594         dev->ibd_hdev = hdev;   /* take over the refcount */
2595         hdev = old;
2596
2597         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2598                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2599                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2600                                             &zombie_tpo);
2601
2602                         if (net->ibn_fmr_ps != NULL)
2603                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2604                                                         &zombie_fpo);
2605                 }
2606         }
2607
2608         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2609  out:
2610         if (!list_empty(&zombie_tpo))
2611                 kiblnd_destroy_pool_list(&zombie_tpo);
2612         if (!list_empty(&zombie_ppo))
2613                 kiblnd_destroy_pool_list(&zombie_ppo);
2614         if (!list_empty(&zombie_fpo))
2615                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2616         if (hdev != NULL)
2617                 kiblnd_hdev_decref(hdev);
2618
2619         if (rc != 0)
2620                 dev->ibd_failed_failover++;
2621         else
2622                 dev->ibd_failed_failover = 0;
2623
2624         return rc;
2625 }
2626
2627 void
2628 kiblnd_destroy_dev (kib_dev_t *dev)
2629 {
2630         LASSERT (dev->ibd_nnets == 0);
2631         LASSERT(list_empty(&dev->ibd_nets));
2632
2633         list_del(&dev->ibd_fail_list);
2634         list_del(&dev->ibd_list);
2635
2636         if (dev->ibd_hdev != NULL)
2637                 kiblnd_hdev_decref(dev->ibd_hdev);
2638
2639         LIBCFS_FREE(dev, sizeof(*dev));
2640 }
2641
2642 static kib_dev_t *
2643 kiblnd_create_dev(char *ifname)
2644 {
2645         struct net_device *netdev;
2646         kib_dev_t         *dev;
2647         __u32              netmask;
2648         __u32              ip;
2649         int                up;
2650         int                rc;
2651
2652         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2653         if (rc != 0) {
2654                 CERROR("Can't query IPoIB interface %s: %d\n",
2655                        ifname, rc);
2656                 return NULL;
2657         }
2658
2659         if (!up) {
2660                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2661                 return NULL;
2662         }
2663
2664         LIBCFS_ALLOC(dev, sizeof(*dev));
2665         if (dev == NULL)
2666                 return NULL;
2667
2668         memset(dev, 0, sizeof(*dev));
2669         netdev = dev_get_by_name(&init_net, ifname);
2670         if (netdev == NULL) {
2671                 dev->ibd_can_failover = 0;
2672         } else {
2673                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2674                 dev_put(netdev);
2675         }
2676
2677         INIT_LIST_HEAD(&dev->ibd_nets);
2678         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2679         INIT_LIST_HEAD(&dev->ibd_fail_list);
2680         dev->ibd_ifip = ip;
2681         strcpy(&dev->ibd_ifname[0], ifname);
2682
2683         /* initialize the device */
2684         rc = kiblnd_dev_failover(dev);
2685         if (rc != 0) {
2686                 CERROR("Can't initialize device: %d\n", rc);
2687                 LIBCFS_FREE(dev, sizeof(*dev));
2688                 return NULL;
2689         }
2690
2691         list_add_tail(&dev->ibd_list,
2692                           &kiblnd_data.kib_devs);
2693         return dev;
2694 }
2695
2696 static void
2697 kiblnd_base_shutdown(void)
2698 {
2699         struct kib_sched_info   *sched;
2700         int                     i;
2701
2702         LASSERT(list_empty(&kiblnd_data.kib_devs));
2703
2704         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2705                atomic_read(&libcfs_kmemory));
2706
2707         switch (kiblnd_data.kib_init) {
2708         default:
2709                 LBUG();
2710
2711         case IBLND_INIT_ALL:
2712         case IBLND_INIT_DATA:
2713                 LASSERT (kiblnd_data.kib_peers != NULL);
2714                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2715                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2716                 }
2717                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2718                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2719                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2720                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2721
2722                 /* flag threads to terminate; wake and wait for them to die */
2723                 kiblnd_data.kib_shutdown = 1;
2724
2725                 /* NB: we really want to stop scheduler threads net by net
2726                  * instead of the whole module, this should be improved
2727                  * with dynamic configuration LNet */
2728                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2729                         wake_up_all(&sched->ibs_waitq);
2730
2731                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2732                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2733
2734                 i = 2;
2735                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2736                         i++;
2737                         /* power of 2? */
2738                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2739                                "Waiting for %d threads to terminate\n",
2740                                atomic_read(&kiblnd_data.kib_nthreads));
2741                         set_current_state(TASK_UNINTERRUPTIBLE);
2742                         schedule_timeout(cfs_time_seconds(1));
2743                 }
2744
2745                 /* fall through */
2746
2747         case IBLND_INIT_NOTHING:
2748                 break;
2749         }
2750
2751         if (kiblnd_data.kib_peers != NULL) {
2752                 LIBCFS_FREE(kiblnd_data.kib_peers,
2753                             sizeof(struct list_head) *
2754                             kiblnd_data.kib_peer_hash_size);
2755         }
2756
2757         if (kiblnd_data.kib_scheds != NULL)
2758                 cfs_percpt_free(kiblnd_data.kib_scheds);
2759
2760         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2761                atomic_read(&libcfs_kmemory));
2762
2763         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2764         module_put(THIS_MODULE);
2765 }
2766
2767 static void
2768 kiblnd_shutdown (lnet_ni_t *ni)
2769 {
2770         kib_net_t        *net = ni->ni_data;
2771         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2772         int               i;
2773         unsigned long     flags;
2774
2775         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2776
2777         if (net == NULL)
2778                 goto out;
2779
2780         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2781                atomic_read(&libcfs_kmemory));
2782
2783         write_lock_irqsave(g_lock, flags);
2784         net->ibn_shutdown = 1;
2785         write_unlock_irqrestore(g_lock, flags);
2786
2787         switch (net->ibn_init) {
2788         default:
2789                 LBUG();
2790
2791         case IBLND_INIT_ALL:
2792                 /* nuke all existing peers within this net */
2793                 kiblnd_del_peer(ni, LNET_NID_ANY);
2794
2795                 /* Wait for all peer state to clean up */
2796                 i = 2;
2797                 while (atomic_read(&net->ibn_npeers) != 0) {
2798                         i++;
2799                         /* power of 2? */
2800                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2801                                "%s: waiting for %d peers to disconnect\n",
2802                                libcfs_nid2str(ni->ni_nid),
2803                                atomic_read(&net->ibn_npeers));
2804                         set_current_state(TASK_UNINTERRUPTIBLE);
2805                         schedule_timeout(cfs_time_seconds(1));
2806                 }
2807
2808                 kiblnd_net_fini_pools(net);
2809
2810                 write_lock_irqsave(g_lock, flags);
2811                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2812                 net->ibn_dev->ibd_nnets--;
2813                 list_del(&net->ibn_list);
2814                 write_unlock_irqrestore(g_lock, flags);
2815
2816                 /* fall through */
2817
2818         case IBLND_INIT_NOTHING:
2819                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2820
2821                 if (net->ibn_dev != NULL &&
2822                     net->ibn_dev->ibd_nnets == 0)
2823                         kiblnd_destroy_dev(net->ibn_dev);
2824
2825                 break;
2826         }
2827
2828         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2829                atomic_read(&libcfs_kmemory));
2830
2831         net->ibn_init = IBLND_INIT_NOTHING;
2832         ni->ni_data = NULL;
2833
2834         LIBCFS_FREE(net, sizeof(*net));
2835
2836 out:
2837         if (list_empty(&kiblnd_data.kib_devs))
2838                 kiblnd_base_shutdown();
2839         return;
2840 }
2841
2842 static int
2843 kiblnd_base_startup(void)
2844 {
2845         struct kib_sched_info   *sched;
2846         int                     rc;
2847         int                     i;
2848
2849         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2850
2851         try_module_get(THIS_MODULE);
2852         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2853
2854         rwlock_init(&kiblnd_data.kib_global_lock);
2855
2856         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2857         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2858
2859         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2860         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2861                      sizeof(struct list_head) *
2862                      kiblnd_data.kib_peer_hash_size);
2863         if (kiblnd_data.kib_peers == NULL)
2864                 goto failed;
2865
2866         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2867                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2868
2869         spin_lock_init(&kiblnd_data.kib_connd_lock);
2870         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2871         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2872         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2873         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2874
2875         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2876         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2877
2878         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2879                                                   sizeof(*sched));
2880         if (kiblnd_data.kib_scheds == NULL)
2881                 goto failed;
2882
2883         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2884                 int     nthrs;
2885
2886                 spin_lock_init(&sched->ibs_lock);
2887                 INIT_LIST_HEAD(&sched->ibs_conns);
2888                 init_waitqueue_head(&sched->ibs_waitq);
2889
2890                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2891                 if (*kiblnd_tunables.kib_nscheds > 0) {
2892                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2893                 } else {
2894                         /* max to half of CPUs, another half is reserved for
2895                          * upper layer modules */
2896                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2897                 }
2898
2899                 sched->ibs_nthreads_max = nthrs;
2900                 sched->ibs_cpt = i;
2901         }
2902
2903         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2904
2905         /* lists/ptrs/locks initialised */
2906         kiblnd_data.kib_init = IBLND_INIT_DATA;
2907         /*****************************************************/
2908
2909         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2910         if (rc != 0) {
2911                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2912                 goto failed;
2913         }
2914
2915         if (*kiblnd_tunables.kib_dev_failover != 0)
2916                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2917                                          "kiblnd_failover");
2918
2919         if (rc != 0) {
2920                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2921                 goto failed;
2922         }
2923
2924         /* flag everything initialised */
2925         kiblnd_data.kib_init = IBLND_INIT_ALL;
2926         /*****************************************************/
2927
2928         return 0;
2929
2930  failed:
2931         kiblnd_base_shutdown();
2932         return -ENETDOWN;
2933 }
2934
2935 static int
2936 kiblnd_start_schedulers(struct kib_sched_info *sched)
2937 {
2938         int     rc = 0;
2939         int     nthrs;
2940         int     i;
2941
2942         if (sched->ibs_nthreads == 0) {
2943                 if (*kiblnd_tunables.kib_nscheds > 0) {
2944                         nthrs = sched->ibs_nthreads_max;
2945                 } else {
2946                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2947                                                sched->ibs_cpt);
2948                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2949                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2950                 }
2951         } else {
2952                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2953                 /* increase one thread if there is new interface */
2954                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2955         }
2956
2957         for (i = 0; i < nthrs; i++) {
2958                 long    id;
2959                 char    name[20];
2960                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2961                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2962                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2963                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2964                 if (rc == 0)
2965                         continue;
2966
2967                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2968                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2969                 break;
2970         }
2971
2972         sched->ibs_nthreads += i;
2973         return rc;
2974 }
2975
2976 static int
2977 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2978 {
2979         int     cpt;
2980         int     rc;
2981         int     i;
2982
2983         for (i = 0; i < ncpts; i++) {
2984                 struct kib_sched_info *sched;
2985
2986                 cpt = (cpts == NULL) ? i : cpts[i];
2987                 sched = kiblnd_data.kib_scheds[cpt];
2988
2989                 if (!newdev && sched->ibs_nthreads > 0)
2990                         continue;
2991
2992                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2993                 if (rc != 0) {
2994                         CERROR("Failed to start scheduler threads for %s\n",
2995                                dev->ibd_ifname);
2996                         return rc;
2997                 }
2998         }
2999         return 0;
3000 }
3001
3002 static kib_dev_t *
3003 kiblnd_dev_search(char *ifname)
3004 {
3005         kib_dev_t       *alias = NULL;
3006         kib_dev_t       *dev;
3007         char            *colon;
3008         char            *colon2;
3009
3010         colon = strchr(ifname, ':');
3011         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3012                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3013                         return dev;
3014
3015                 if (alias != NULL)
3016                         continue;
3017
3018                 colon2 = strchr(dev->ibd_ifname, ':');
3019                 if (colon != NULL)
3020                         *colon = 0;
3021                 if (colon2 != NULL)
3022                         *colon2 = 0;
3023
3024                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3025                         alias = dev;
3026
3027                 if (colon != NULL)
3028                         *colon = ':';
3029                 if (colon2 != NULL)
3030                         *colon2 = ':';
3031         }
3032         return alias;
3033 }
3034
3035 static int
3036 kiblnd_startup (lnet_ni_t *ni)
3037 {
3038         char                     *ifname;
3039         kib_dev_t                *ibdev = NULL;
3040         kib_net_t                *net;
3041         struct timeval            tv;
3042         unsigned long             flags;
3043         int                       rc;
3044         int                       newdev;
3045
3046         LASSERT (ni->ni_lnd == &the_o2iblnd);
3047
3048         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3049                 rc = kiblnd_base_startup();
3050                 if (rc != 0)
3051                         return rc;
3052         }
3053
3054         LIBCFS_ALLOC(net, sizeof(*net));
3055         ni->ni_data = net;
3056         if (net == NULL)
3057                 goto failed;
3058
3059         memset(net, 0, sizeof(*net));
3060
3061         do_gettimeofday(&tv);
3062         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3063
3064         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3065         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3066         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3067         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3068
3069         if (ni->ni_interfaces[0] != NULL) {
3070                 /* Use the IPoIB interface specified in 'networks=' */
3071
3072                 CLASSERT (LNET_MAX_INTERFACES > 1);
3073                 if (ni->ni_interfaces[1] != NULL) {
3074                         CERROR("Multiple interfaces not supported\n");
3075                         goto failed;
3076                 }
3077
3078                 ifname = ni->ni_interfaces[0];
3079         } else {
3080                 ifname = *kiblnd_tunables.kib_default_ipif;
3081         }
3082
3083         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3084                 CERROR("IPoIB interface name too long: %s\n", ifname);
3085                 goto failed;
3086         }
3087
3088         ibdev = kiblnd_dev_search(ifname);
3089
3090         newdev = ibdev == NULL;
3091         /* hmm...create kib_dev even for alias */
3092         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3093                 ibdev = kiblnd_create_dev(ifname);
3094
3095         if (ibdev == NULL)
3096                 goto failed;
3097
3098         net->ibn_dev = ibdev;
3099         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3100
3101         rc = kiblnd_dev_start_threads(ibdev, newdev,
3102                                       ni->ni_cpts, ni->ni_ncpts);
3103         if (rc != 0)
3104                 goto failed;
3105
3106         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3107         if (rc != 0) {
3108                 CERROR("Failed to initialize NI pools: %d\n", rc);
3109                 goto failed;
3110         }
3111
3112         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3113         ibdev->ibd_nnets++;
3114         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3115         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3116
3117         net->ibn_init = IBLND_INIT_ALL;
3118
3119         return 0;
3120
3121 failed:
3122         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3123                 kiblnd_destroy_dev(ibdev);
3124
3125         kiblnd_shutdown(ni);
3126
3127         CDEBUG(D_NET, "kiblnd_startup failed\n");
3128         return -ENETDOWN;
3129 }
3130
3131 static lnd_t the_o2iblnd = {
3132         .lnd_type       = O2IBLND,
3133         .lnd_startup    = kiblnd_startup,
3134         .lnd_shutdown   = kiblnd_shutdown,
3135         .lnd_ctl        = kiblnd_ctl,
3136         .lnd_query      = kiblnd_query,
3137         .lnd_send       = kiblnd_send,
3138         .lnd_recv       = kiblnd_recv,
3139 };
3140
3141 static void __exit ko2iblnd_exit(void)
3142 {
3143         lnet_unregister_lnd(&the_o2iblnd);
3144         kiblnd_tunables_fini();
3145 }
3146
3147 static int __init ko2iblnd_init(void)
3148 {
3149         int rc;
3150
3151         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3152         CLASSERT(offsetof(kib_msg_t,
3153                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3154                  IBLND_MSG_SIZE);
3155         CLASSERT(offsetof(kib_msg_t,
3156                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3157                  <= IBLND_MSG_SIZE);
3158
3159         rc = kiblnd_tunables_init();
3160         if (rc != 0)
3161                 return rc;
3162
3163         lnet_register_lnd(&the_o2iblnd);
3164
3165         return 0;
3166 }
3167
3168 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3169 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3170 MODULE_VERSION("2.8.0");
3171 MODULE_LICENSE("GPL");
3172
3173 module_init(ko2iblnd_init);
3174 module_exit(ko2iblnd_exit);