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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include <asm/page.h>
  38 #include "o2iblnd.h"
  39
  40 static lnd_t the_o2iblnd;
  41
  42 kib_data_t              kiblnd_data;
  43
  44 static __u32
  45 kiblnd_cksum (void *ptr, int nob)
  46 {
  47         char  *c  = ptr;
  48         __u32  sum = 0;
  49
  50         while (nob-- > 0)
  51                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  52
  53         /* ensure I don't return 0 (== no checksum) */
  54         return (sum == 0) ? 1 : sum;
  55 }
  56
  57 static char *
  58 kiblnd_msgtype2str(int type)
  59 {
  60         switch (type) {
  61         case IBLND_MSG_CONNREQ:
  62                 return "CONNREQ";
  63
  64         case IBLND_MSG_CONNACK:
  65                 return "CONNACK";
  66
  67         case IBLND_MSG_NOOP:
  68                 return "NOOP";
  69
  70         case IBLND_MSG_IMMEDIATE:
  71                 return "IMMEDIATE";
  72
  73         case IBLND_MSG_PUT_REQ:
  74                 return "PUT_REQ";
  75
  76         case IBLND_MSG_PUT_NAK:
  77                 return "PUT_NAK";
  78
  79         case IBLND_MSG_PUT_ACK:
  80                 return "PUT_ACK";
  81
  82         case IBLND_MSG_PUT_DONE:
  83                 return "PUT_DONE";
  84
  85         case IBLND_MSG_GET_REQ:
  86                 return "GET_REQ";
  87
  88         case IBLND_MSG_GET_DONE:
  89                 return "GET_DONE";
  90
  91         default:
  92                 return "???";
  93         }
  94 }
  95
  96 static int
  97 kiblnd_msgtype2size(int type)
  98 {
  99         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 100
 101         switch (type) {
 102         case IBLND_MSG_CONNREQ:
 103         case IBLND_MSG_CONNACK:
 104                 return hdr_size + sizeof(kib_connparams_t);
 105
 106         case IBLND_MSG_NOOP:
 107                 return hdr_size;
 108
 109         case IBLND_MSG_IMMEDIATE:
 110                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 111
 112         case IBLND_MSG_PUT_REQ:
 113                 return hdr_size + sizeof(kib_putreq_msg_t);
 114
 115         case IBLND_MSG_PUT_ACK:
 116                 return hdr_size + sizeof(kib_putack_msg_t);
 117
 118         case IBLND_MSG_GET_REQ:
 119                 return hdr_size + sizeof(kib_get_msg_t);
 120
 121         case IBLND_MSG_PUT_NAK:
 122         case IBLND_MSG_PUT_DONE:
 123         case IBLND_MSG_GET_DONE:
 124                 return hdr_size + sizeof(kib_completion_msg_t);
 125         default:
 126                 return -1;
 127         }
 128 }
 129
 130 static int
 131 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 132 {
 133         kib_rdma_desc_t   *rd;
 134         int msg_size;
 135         int                nob;
 136         int                n;
 137         int                i;
 138
 139         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 140                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 141
 142         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 143                               &msg->ibm_u.get.ibgm_rd :
 144                               &msg->ibm_u.putack.ibpam_rd;
 145
 146         if (flip) {
 147                 __swab32s(&rd->rd_key);
 148                 __swab32s(&rd->rd_nfrags);
 149         }
 150
 151         n = rd->rd_nfrags;
 152         nob = offsetof (kib_msg_t, ibm_u) +
 153               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 154
 155         if (msg->ibm_nob < nob) {
 156                 CERROR("Short %s: %d(%d)\n",
 157                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 158                 return 1;
 159         }
 160
 161         msg_size = kiblnd_rd_size(rd);
 162         if (msg_size <= 0 || msg_size > LNET_MAX_PAYLOAD) {
 163                 CERROR("Bad msg_size: %d, should be 0 < n <= %d\n",
 164                        msg_size, LNET_MAX_PAYLOAD);
 165                 return 1;
 166         }
 167
 168         if (!flip)
 169                 return 0;
 170
 171         for (i = 0; i < n; i++) {
 172                 __swab32s(&rd->rd_frags[i].rf_nob);
 173                 __swab64s(&rd->rd_frags[i].rf_addr);
 174         }
 175
 176         return 0;
 177 }
 178
 179 void
 180 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 181                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 182 {
 183         kib_net_t *net = ni->ni_data;
 184
 185         /* CAVEAT EMPTOR! all message fields not set here should have been
 186          * initialised previously. */
 187         msg->ibm_magic    = IBLND_MSG_MAGIC;
 188         msg->ibm_version  = version;
 189         /*   ibm_type */
 190         msg->ibm_credits  = credits;
 191         /*   ibm_nob */
 192         msg->ibm_cksum    = 0;
 193         msg->ibm_srcnid   = ni->ni_nid;
 194         msg->ibm_srcstamp = net->ibn_incarnation;
 195         msg->ibm_dstnid   = dstnid;
 196         msg->ibm_dststamp = dststamp;
 197
 198         if (*kiblnd_tunables.kib_cksum) {
 199                 /* NB ibm_cksum zero while computing cksum */
 200                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 201         }
 202 }
 203
 204 int
 205 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 206 {
 207         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 208         __u32     msg_cksum;
 209         __u16     version;
 210         int       msg_nob;
 211         int       flip;
 212
 213         /* 6 bytes are enough to have received magic + version */
 214         if (nob < 6) {
 215                 CERROR("Short message: %d\n", nob);
 216                 return -EPROTO;
 217         }
 218
 219         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 220                 flip = 0;
 221         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 222                 flip = 1;
 223         } else {
 224                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 225                 return -EPROTO;
 226         }
 227
 228         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 229         if (version != IBLND_MSG_VERSION &&
 230             version != IBLND_MSG_VERSION_1) {
 231                 CERROR("Bad version: %x\n", version);
 232                 return -EPROTO;
 233         }
 234
 235         if (nob < hdr_size) {
 236                 CERROR("Short message: %d\n", nob);
 237                 return -EPROTO;
 238         }
 239
 240         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 241         if (msg_nob > nob) {
 242                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 243                 return -EPROTO;
 244         }
 245
 246         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 247          * gets flipped */
 248         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 249         msg->ibm_cksum = 0;
 250         if (msg_cksum != 0 &&
 251             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 252                 CERROR("Bad checksum\n");
 253                 return -EPROTO;
 254         }
 255
 256         msg->ibm_cksum = msg_cksum;
 257
 258         if (flip) {
 259                 /* leave magic unflipped as a clue to peer endianness */
 260                 msg->ibm_version = version;
 261                 CLASSERT (sizeof(msg->ibm_type) == 1);
 262                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 263                 msg->ibm_nob     = msg_nob;
 264                 __swab64s(&msg->ibm_srcnid);
 265                 __swab64s(&msg->ibm_srcstamp);
 266                 __swab64s(&msg->ibm_dstnid);
 267                 __swab64s(&msg->ibm_dststamp);
 268         }
 269
 270         if (msg->ibm_srcnid == LNET_NID_ANY) {
 271                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 272                 return -EPROTO;
 273         }
 274
 275         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 276                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 277                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 278                 return -EPROTO;
 279         }
 280
 281         switch (msg->ibm_type) {
 282         default:
 283                 CERROR("Unknown message type %x\n", msg->ibm_type);
 284                 return -EPROTO;
 285
 286         case IBLND_MSG_NOOP:
 287         case IBLND_MSG_IMMEDIATE:
 288         case IBLND_MSG_PUT_REQ:
 289                 break;
 290
 291         case IBLND_MSG_PUT_ACK:
 292         case IBLND_MSG_GET_REQ:
 293                 if (kiblnd_unpack_rd(msg, flip))
 294                         return -EPROTO;
 295                 break;
 296
 297         case IBLND_MSG_PUT_NAK:
 298         case IBLND_MSG_PUT_DONE:
 299         case IBLND_MSG_GET_DONE:
 300                 if (flip)
 301                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 302                 break;
 303
 304         case IBLND_MSG_CONNREQ:
 305         case IBLND_MSG_CONNACK:
 306                 if (flip) {
 307                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 308                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 309                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 310                 }
 311                 break;
 312         }
 313         return 0;
 314 }
 315
 316 int
 317 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 318 {
 319         kib_peer_t      *peer;
 320         kib_net_t       *net = ni->ni_data;
 321         int             cpt = lnet_cpt_of_nid(nid);
 322         unsigned long   flags;
 323
 324         LASSERT(net != NULL);
 325         LASSERT(nid != LNET_NID_ANY);
 326
 327         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 328         if (peer == NULL) {
 329                 CERROR("Cannot allocate peer\n");
 330                 return -ENOMEM;
 331         }
 332
 333         peer->ibp_ni = ni;
 334         peer->ibp_nid = nid;
 335         peer->ibp_error = 0;
 336         peer->ibp_last_alive = 0;
 337         peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni);
 338         peer->ibp_queue_depth = ni->ni_peertxcredits;
 339         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 340
 341         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 342         INIT_LIST_HEAD(&peer->ibp_conns);
 343         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 344
 345         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 346
 347         /* always called with a ref on ni, which prevents ni being shutdown */
 348         LASSERT(net->ibn_shutdown == 0);
 349
 350         /* npeers only grows with the global lock held */
 351         atomic_inc(&net->ibn_npeers);
 352
 353         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 354
 355         *peerp = peer;
 356         return 0;
 357 }
 358
 359 void
 360 kiblnd_destroy_peer (kib_peer_t *peer)
 361 {
 362         kib_net_t *net = peer->ibp_ni->ni_data;
 363
 364         LASSERT(net != NULL);
 365         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 366         LASSERT(!kiblnd_peer_active(peer));
 367         LASSERT(kiblnd_peer_idle(peer));
 368         LASSERT(list_empty(&peer->ibp_tx_queue));
 369
 370         LIBCFS_FREE(peer, sizeof(*peer));
 371
 372         /* NB a peer's connections keep a reference on their peer until
 373          * they are destroyed, so we can be assured that _all_ state to do
 374          * with this peer has been cleaned up when its refcount drops to
 375          * zero. */
 376         atomic_dec(&net->ibn_npeers);
 377 }
 378
 379 kib_peer_t *
 380 kiblnd_find_peer_locked (lnet_nid_t nid)
 381 {
 382         /* the caller is responsible for accounting the additional reference
 383          * that this creates */
 384         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 385         struct list_head        *tmp;
 386         kib_peer_t              *peer;
 387
 388         list_for_each(tmp, peer_list) {
 389
 390                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 391                 LASSERT(!kiblnd_peer_idle(peer));
 392
 393                 if (peer->ibp_nid != nid)
 394                         continue;
 395
 396                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 397                        peer, libcfs_nid2str(nid),
 398                        atomic_read(&peer->ibp_refcount),
 399                        peer->ibp_version);
 400                 return peer;
 401         }
 402         return NULL;
 403 }
 404
 405 void
 406 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 407 {
 408         LASSERT(list_empty(&peer->ibp_conns));
 409
 410         LASSERT (kiblnd_peer_active(peer));
 411         list_del_init(&peer->ibp_list);
 412         /* lose peerlist's ref */
 413         kiblnd_peer_decref(peer);
 414 }
 415
 416 static int
 417 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 418                      lnet_nid_t *nidp, int *count)
 419 {
 420         kib_peer_t              *peer;
 421         struct list_head        *ptmp;
 422         int                      i;
 423         unsigned long            flags;
 424
 425         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 426
 427         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 428
 429                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 430
 431                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 432                         LASSERT(!kiblnd_peer_idle(peer));
 433
 434                         if (peer->ibp_ni != ni)
 435                                 continue;
 436
 437                         if (index-- > 0)
 438                                 continue;
 439
 440                         *nidp = peer->ibp_nid;
 441                         *count = atomic_read(&peer->ibp_refcount);
 442
 443                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 444                                                flags);
 445                         return 0;
 446                 }
 447         }
 448
 449         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 450         return -ENOENT;
 451 }
 452
 453 static void
 454 kiblnd_del_peer_locked (kib_peer_t *peer)
 455 {
 456         struct list_head        *ctmp;
 457         struct list_head        *cnxt;
 458         kib_conn_t              *conn;
 459
 460         if (list_empty(&peer->ibp_conns)) {
 461                 kiblnd_unlink_peer_locked(peer);
 462         } else {
 463                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 464                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 465
 466                         kiblnd_close_conn_locked(conn, 0);
 467                 }
 468                 /* NB closing peer's last conn unlinked it. */
 469         }
 470         /* NB peer now unlinked; might even be freed if the peer table had the
 471          * last ref on it. */
 472 }
 473
 474 static int
 475 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 476 {
 477         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 478         struct list_head        *ptmp;
 479         struct list_head        *pnxt;
 480         kib_peer_t              *peer;
 481         int                     lo;
 482         int                     hi;
 483         int                     i;
 484         unsigned long           flags;
 485         int                     rc = -ENOENT;
 486
 487         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 488
 489         if (nid != LNET_NID_ANY) {
 490                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 491         } else {
 492                 lo = 0;
 493                 hi = kiblnd_data.kib_peer_hash_size - 1;
 494         }
 495
 496         for (i = lo; i <= hi; i++) {
 497                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 498                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 499                         LASSERT(!kiblnd_peer_idle(peer));
 500
 501                         if (peer->ibp_ni != ni)
 502                                 continue;
 503
 504                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 505                                 continue;
 506
 507                         if (!list_empty(&peer->ibp_tx_queue)) {
 508                                 LASSERT(list_empty(&peer->ibp_conns));
 509
 510                                 list_splice_init(&peer->ibp_tx_queue,
 511                                                  &zombies);
 512                         }
 513
 514                         kiblnd_del_peer_locked(peer);
 515                         rc = 0;         /* matched something */
 516                 }
 517         }
 518
 519         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 520
 521         kiblnd_txlist_done(ni, &zombies, -EIO);
 522
 523         return rc;
 524 }
 525
 526 static kib_conn_t *
 527 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 528 {
 529         kib_peer_t              *peer;
 530         struct list_head        *ptmp;
 531         kib_conn_t              *conn;
 532         struct list_head        *ctmp;
 533         int                     i;
 534         unsigned long           flags;
 535
 536         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 537
 538         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 539                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 540
 541                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 542                         LASSERT(!kiblnd_peer_idle(peer));
 543
 544                         if (peer->ibp_ni != ni)
 545                                 continue;
 546
 547                         list_for_each(ctmp, &peer->ibp_conns) {
 548                                 if (index-- > 0)
 549                                         continue;
 550
 551                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 552                                 kiblnd_conn_addref(conn);
 553                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 554                                                        flags);
 555                                 return conn;
 556                         }
 557                 }
 558         }
 559
 560         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 561         return NULL;
 562 }
 563
 564 static void
 565 kiblnd_debug_rx (kib_rx_t *rx)
 566 {
 567         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 568                rx, rx->rx_status, rx->rx_msg->ibm_type,
 569                rx->rx_msg->ibm_credits);
 570 }
 571
 572 static void
 573 kiblnd_debug_tx (kib_tx_t *tx)
 574 {
 575         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 576                "cookie %#llx msg %s%s type %x cred %d\n",
 577                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 578                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 579                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 580                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 581                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 582 }
 583
 584 void
 585 kiblnd_debug_conn (kib_conn_t *conn)
 586 {
 587         struct list_head        *tmp;
 588         int                     i;
 589
 590         spin_lock(&conn->ibc_lock);
 591
 592         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 593                atomic_read(&conn->ibc_refcount), conn,
 594                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 595         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 596                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 597                conn->ibc_nsends_posted, conn->ibc_credits,
 598                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 599         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 600
 601         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 602         list_for_each(tmp, &conn->ibc_early_rxs)
 603                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 604
 605         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 606         list_for_each(tmp, &conn->ibc_tx_noops)
 607                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 608
 609         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 610         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 611                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 612
 613         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 614         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 615                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 616
 617         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 618         list_for_each(tmp, &conn->ibc_tx_queue)
 619                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 620
 621         CDEBUG(D_CONSOLE, "   active_txs:\n");
 622         list_for_each(tmp, &conn->ibc_active_txs)
 623                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 624
 625         CDEBUG(D_CONSOLE, "   rxs:\n");
 626         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 627                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 628
 629         spin_unlock(&conn->ibc_lock);
 630 }
 631
 632 int
 633 kiblnd_translate_mtu(int value)
 634 {
 635         switch (value) {
 636         default:
 637                 return -1;
 638         case 0:
 639                 return 0;
 640         case 256:
 641                 return IB_MTU_256;
 642         case 512:
 643                 return IB_MTU_512;
 644         case 1024:
 645                 return IB_MTU_1024;
 646         case 2048:
 647                 return IB_MTU_2048;
 648         case 4096:
 649                 return IB_MTU_4096;
 650         }
 651 }
 652
 653 static void
 654 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 655 {
 656         int           mtu;
 657
 658         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 659         if (cmid->route.path_rec == NULL)
 660                 return;
 661
 662         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 663         LASSERT (mtu >= 0);
 664         if (mtu != 0)
 665                 cmid->route.path_rec->mtu = mtu;
 666 }
 667
 668 static int
 669 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 670 {
 671         cpumask_t       *mask;
 672         int             vectors;
 673         int             off;
 674         int             i;
 675         lnet_nid_t      ibp_nid;
 676
 677         vectors = conn->ibc_cmid->device->num_comp_vectors;
 678         if (vectors <= 1)
 679                 return 0;
 680
 681         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 682
 683         /* hash NID to CPU id in this partition... */
 684         ibp_nid = conn->ibc_peer->ibp_nid;
 685         off = do_div(ibp_nid, cpumask_weight(mask));
 686         for_each_cpu(i, mask) {
 687                 if (off-- == 0)
 688                         return i % vectors;
 689         }
 690
 691         LBUG();
 692         return 1;
 693 }
 694
 695 kib_conn_t *
 696 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 697                    int state, int version)
 698 {
 699         /* CAVEAT EMPTOR:
 700          * If the new conn is created successfully it takes over the caller's
 701          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 702          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 703          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 704          * to destroy 'cmid' here since I'm called from the CM which still has
 705          * its ref on 'cmid'). */
 706         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 707         kib_net_t              *net = peer->ibp_ni->ni_data;
 708         kib_dev_t              *dev;
 709         struct ib_qp_init_attr *init_qp_attr;
 710         struct kib_sched_info   *sched;
 711 #ifdef HAVE_IB_CQ_INIT_ATTR
 712         struct ib_cq_init_attr  cq_attr = {};
 713 #endif
 714         kib_conn_t              *conn;
 715         struct ib_cq            *cq;
 716         unsigned long           flags;
 717         int                     cpt;
 718         int                     rc;
 719         int                     i;
 720
 721         LASSERT(net != NULL);
 722         LASSERT(!in_interrupt());
 723
 724         dev = net->ibn_dev;
 725
 726         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 727         sched = kiblnd_data.kib_scheds[cpt];
 728
 729         LASSERT(sched->ibs_nthreads > 0);
 730
 731         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 732                          sizeof(*init_qp_attr));
 733         if (init_qp_attr == NULL) {
 734                 CERROR("Can't allocate qp_attr for %s\n",
 735                        libcfs_nid2str(peer->ibp_nid));
 736                 goto failed_0;
 737         }
 738
 739         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 740         if (conn == NULL) {
 741                 CERROR("Can't allocate connection for %s\n",
 742                        libcfs_nid2str(peer->ibp_nid));
 743                 goto failed_1;
 744         }
 745
 746         conn->ibc_state = IBLND_CONN_INIT;
 747         conn->ibc_version = version;
 748         conn->ibc_peer = peer;                  /* I take the caller's ref */
 749         cmid->context = conn;                   /* for future CM callbacks */
 750         conn->ibc_cmid = cmid;
 751         conn->ibc_max_frags = peer->ibp_max_frags;
 752         conn->ibc_queue_depth = peer->ibp_queue_depth;
 753
 754         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 755         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 756         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 757         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 758         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 759         INIT_LIST_HEAD(&conn->ibc_active_txs);
 760         spin_lock_init(&conn->ibc_lock);
 761
 762         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 763                          sizeof(*conn->ibc_connvars));
 764         if (conn->ibc_connvars == NULL) {
 765                 CERROR("Can't allocate in-progress connection state\n");
 766                 goto failed_2;
 767         }
 768
 769         write_lock_irqsave(glock, flags);
 770         if (dev->ibd_failover) {
 771                 write_unlock_irqrestore(glock, flags);
 772                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 773                 goto failed_2;
 774         }
 775
 776         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 777                 /* wakeup failover thread and teardown connection */
 778                 if (kiblnd_dev_can_failover(dev)) {
 779                         list_add_tail(&dev->ibd_fail_list,
 780                                       &kiblnd_data.kib_failed_devs);
 781                         wake_up(&kiblnd_data.kib_failover_waitq);
 782                 }
 783
 784                 write_unlock_irqrestore(glock, flags);
 785                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 786                        cmid->device->name, dev->ibd_ifname);
 787                 goto failed_2;
 788         }
 789
 790         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 791         conn->ibc_hdev = dev->ibd_hdev;
 792
 793         kiblnd_setup_mtu_locked(cmid);
 794
 795         write_unlock_irqrestore(glock, flags);
 796
 797         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 798                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 799         if (conn->ibc_rxs == NULL) {
 800                 CERROR("Cannot allocate RX buffers\n");
 801                 goto failed_2;
 802         }
 803
 804         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 805                                 IBLND_RX_MSG_PAGES(conn));
 806         if (rc != 0)
 807                 goto failed_2;
 808
 809         kiblnd_map_rx_descs(conn);
 810
 811 #ifdef HAVE_IB_CQ_INIT_ATTR
 812         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 813         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 814         cq = ib_create_cq(cmid->device,
 815                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 816                           &cq_attr);
 817 #else
 818         cq = ib_create_cq(cmid->device,
 819                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 820                           IBLND_CQ_ENTRIES(conn),
 821                           kiblnd_get_completion_vector(conn, cpt));
 822 #endif
 823         if (IS_ERR(cq)) {
 824                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 825                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 826                 goto failed_2;
 827         }
 828
 829         conn->ibc_cq = cq;
 830
 831         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 832         if (rc != 0) {
 833                 CERROR("Can't request completion notification: %d\n", rc);
 834                 goto failed_2;
 835         }
 836
 837         init_qp_attr->event_handler = kiblnd_qp_event;
 838         init_qp_attr->qp_context = conn;
 839         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 840         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 841         init_qp_attr->cap.max_send_sge = 1;
 842         init_qp_attr->cap.max_recv_sge = 1;
 843         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 844         init_qp_attr->qp_type = IB_QPT_RC;
 845         init_qp_attr->send_cq = cq;
 846         init_qp_attr->recv_cq = cq;
 847
 848         conn->ibc_sched = sched;
 849
 850         do {
 851                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 852                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 853                         break;
 854
 855                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 856         } while (rc);
 857
 858         if (rc) {
 859                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 860                         rc, init_qp_attr->cap.max_send_wr,
 861                         init_qp_attr->cap.max_recv_wr);
 862                 goto failed_2;
 863         }
 864
 865         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 866                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 867                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 868
 869         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 870
 871         /* 1 ref for caller and each rxmsg */
 872         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 873         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 874
 875         /* post receives */
 876         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 877                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 878                 if (rc != 0) {
 879                         CERROR("Can't post rxmsg: %d\n", rc);
 880
 881                         /* Make posted receives complete */
 882                         kiblnd_abort_receives(conn);
 883
 884                         /* correct # of posted buffers
 885                          * NB locking needed now I'm racing with completion */
 886                         spin_lock_irqsave(&sched->ibs_lock, flags);
 887                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 888                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 889
 890                         /* cmid will be destroyed by CM(ofed) after cm_callback
 891                          * returned, so we can't refer it anymore
 892                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 893                         rdma_destroy_qp(conn->ibc_cmid);
 894                         conn->ibc_cmid = NULL;
 895
 896                         /* Drop my own and unused rxbuffer refcounts */
 897                         while (i++ <= IBLND_RX_MSGS(conn))
 898                                 kiblnd_conn_decref(conn);
 899
 900                         return NULL;
 901                 }
 902         }
 903
 904         /* Init successful! */
 905         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 906                  state == IBLND_CONN_PASSIVE_WAIT);
 907         conn->ibc_state = state;
 908
 909         /* 1 more conn */
 910         atomic_inc(&net->ibn_nconns);
 911         return conn;
 912
 913  failed_2:
 914         kiblnd_destroy_conn(conn, true);
 915  failed_1:
 916         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 917  failed_0:
 918         return NULL;
 919 }
 920
 921 void
 922 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 923 {
 924         struct rdma_cm_id *cmid = conn->ibc_cmid;
 925         kib_peer_t        *peer = conn->ibc_peer;
 926         int                rc;
 927
 928         LASSERT (!in_interrupt());
 929         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 930         LASSERT(list_empty(&conn->ibc_early_rxs));
 931         LASSERT(list_empty(&conn->ibc_tx_noops));
 932         LASSERT(list_empty(&conn->ibc_tx_queue));
 933         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 934         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 935         LASSERT(list_empty(&conn->ibc_active_txs));
 936         LASSERT (conn->ibc_noops_posted == 0);
 937         LASSERT (conn->ibc_nsends_posted == 0);
 938
 939         switch (conn->ibc_state) {
 940         default:
 941                 /* conn must be completely disengaged from the network */
 942                 LBUG();
 943
 944         case IBLND_CONN_DISCONNECTED:
 945                 /* connvars should have been freed already */
 946                 LASSERT (conn->ibc_connvars == NULL);
 947                 break;
 948
 949         case IBLND_CONN_INIT:
 950                 break;
 951         }
 952
 953         /* conn->ibc_cmid might be destroyed by CM already */
 954         if (cmid != NULL && cmid->qp != NULL)
 955                 rdma_destroy_qp(cmid);
 956
 957         if (conn->ibc_cq != NULL) {
 958                 rc = ib_destroy_cq(conn->ibc_cq);
 959                 if (rc != 0)
 960                         CWARN("Error destroying CQ: %d\n", rc);
 961         }
 962
 963         if (conn->ibc_rx_pages != NULL)
 964                 kiblnd_unmap_rx_descs(conn);
 965
 966         if (conn->ibc_rxs != NULL) {
 967                 LIBCFS_FREE(conn->ibc_rxs,
 968                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 969         }
 970
 971         if (conn->ibc_connvars != NULL)
 972                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 973
 974         if (conn->ibc_hdev != NULL)
 975                 kiblnd_hdev_decref(conn->ibc_hdev);
 976
 977         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 978         if (conn->ibc_state != IBLND_CONN_INIT) {
 979                 kib_net_t *net = peer->ibp_ni->ni_data;
 980
 981                 kiblnd_peer_decref(peer);
 982                 rdma_destroy_id(cmid);
 983                 atomic_dec(&net->ibn_nconns);
 984         }
 985
 986         if (free_conn)
 987                 LIBCFS_FREE(conn, sizeof(*conn));
 988 }
 989
 990 int
 991 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 992 {
 993         kib_conn_t              *conn;
 994         struct list_head        *ctmp;
 995         struct list_head        *cnxt;
 996         int                     count = 0;
 997
 998         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 999                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1000
1001                 CDEBUG(D_NET, "Closing conn -> %s, "
1002                               "version: %x, reason: %d\n",
1003                        libcfs_nid2str(peer->ibp_nid),
1004                        conn->ibc_version, why);
1005
1006                 kiblnd_close_conn_locked(conn, why);
1007                 count++;
1008         }
1009
1010         return count;
1011 }
1012
1013 int
1014 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1015                                 int version, __u64 incarnation)
1016 {
1017         kib_conn_t              *conn;
1018         struct list_head        *ctmp;
1019         struct list_head        *cnxt;
1020         int                     count = 0;
1021
1022         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1023                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1024
1025                 if (conn->ibc_version     == version &&
1026                     conn->ibc_incarnation == incarnation)
1027                         continue;
1028
1029                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1030                               "incarnation:%#llx(%x, %#llx)\n",
1031                        libcfs_nid2str(peer->ibp_nid),
1032                        conn->ibc_version, conn->ibc_incarnation,
1033                        version, incarnation);
1034
1035                 kiblnd_close_conn_locked(conn, -ESTALE);
1036                 count++;
1037         }
1038
1039         return count;
1040 }
1041
1042 static int
1043 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1044 {
1045         kib_peer_t              *peer;
1046         struct list_head        *ptmp;
1047         struct list_head        *pnxt;
1048         int                     lo;
1049         int                     hi;
1050         int                     i;
1051         unsigned long           flags;
1052         int                     count = 0;
1053
1054         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1055
1056         if (nid != LNET_NID_ANY)
1057                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1058         else {
1059                 lo = 0;
1060                 hi = kiblnd_data.kib_peer_hash_size - 1;
1061         }
1062
1063         for (i = lo; i <= hi; i++) {
1064                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1065
1066                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1067                         LASSERT(!kiblnd_peer_idle(peer));
1068
1069                         if (peer->ibp_ni != ni)
1070                                 continue;
1071
1072                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1073                                 continue;
1074
1075                         count += kiblnd_close_peer_conns_locked(peer, 0);
1076                 }
1077         }
1078
1079         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1080
1081         /* wildcards always succeed */
1082         if (nid == LNET_NID_ANY)
1083                 return 0;
1084
1085         return (count == 0) ? -ENOENT : 0;
1086 }
1087
1088 static int
1089 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1090 {
1091         struct libcfs_ioctl_data *data = arg;
1092         int                       rc = -EINVAL;
1093
1094         switch(cmd) {
1095         case IOC_LIBCFS_GET_PEER: {
1096                 lnet_nid_t   nid = 0;
1097                 int          count = 0;
1098
1099                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1100                                           &nid, &count);
1101                 data->ioc_nid    = nid;
1102                 data->ioc_count  = count;
1103                 break;
1104         }
1105
1106         case IOC_LIBCFS_DEL_PEER: {
1107                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1108                 break;
1109         }
1110         case IOC_LIBCFS_GET_CONN: {
1111                 kib_conn_t *conn;
1112
1113                 rc = 0;
1114                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1115                 if (conn == NULL) {
1116                         rc = -ENOENT;
1117                         break;
1118                 }
1119
1120                 LASSERT (conn->ibc_cmid != NULL);
1121                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1122                 if (conn->ibc_cmid->route.path_rec == NULL)
1123                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1124                 else
1125                         data->ioc_u32[0] =
1126                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1127                 kiblnd_conn_decref(conn);
1128                 break;
1129         }
1130         case IOC_LIBCFS_CLOSE_CONNECTION: {
1131                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1132                 break;
1133         }
1134
1135         default:
1136                 break;
1137         }
1138
1139         return rc;
1140 }
1141
1142 static void
1143 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1144 {
1145         cfs_time_t      last_alive = 0;
1146         cfs_time_t      now = cfs_time_current();
1147         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1148         kib_peer_t      *peer;
1149         unsigned long   flags;
1150
1151         read_lock_irqsave(glock, flags);
1152
1153         peer = kiblnd_find_peer_locked(nid);
1154         if (peer != NULL)
1155                 last_alive = peer->ibp_last_alive;
1156
1157         read_unlock_irqrestore(glock, flags);
1158
1159         if (last_alive != 0)
1160                 *when = last_alive;
1161
1162         /* peer is not persistent in hash, trigger peer creation
1163          * and connection establishment with a NULL tx */
1164         if (peer == NULL)
1165                 kiblnd_launch_tx(ni, NULL, nid);
1166
1167         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1168                libcfs_nid2str(nid), peer,
1169                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1170         return;
1171 }
1172
1173 static void
1174 kiblnd_free_pages(kib_pages_t *p)
1175 {
1176         int     npages = p->ibp_npages;
1177         int     i;
1178
1179         for (i = 0; i < npages; i++) {
1180                 if (p->ibp_pages[i] != NULL)
1181                         __free_page(p->ibp_pages[i]);
1182         }
1183
1184         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1185 }
1186
1187 int
1188 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1189 {
1190         kib_pages_t     *p;
1191         int             i;
1192
1193         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1194                          offsetof(kib_pages_t, ibp_pages[npages]));
1195         if (p == NULL) {
1196                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1197                 return -ENOMEM;
1198         }
1199
1200         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1201         p->ibp_npages = npages;
1202
1203         for (i = 0; i < npages; i++) {
1204                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1205                                                      GFP_NOFS);
1206                 if (p->ibp_pages[i] == NULL) {
1207                         CERROR("Can't allocate page %d of %d\n", i, npages);
1208                         kiblnd_free_pages(p);
1209                         return -ENOMEM;
1210                 }
1211         }
1212
1213         *pp = p;
1214         return 0;
1215 }
1216
1217 void
1218 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1219 {
1220         kib_rx_t *rx;
1221         int       i;
1222
1223         LASSERT (conn->ibc_rxs != NULL);
1224         LASSERT (conn->ibc_hdev != NULL);
1225
1226         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1227                 rx = &conn->ibc_rxs[i];
1228
1229                 LASSERT(rx->rx_nob >= 0); /* not posted */
1230
1231                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1232                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1233                                                           rx->rx_msgaddr),
1234                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1235         }
1236
1237         kiblnd_free_pages(conn->ibc_rx_pages);
1238
1239         conn->ibc_rx_pages = NULL;
1240 }
1241
1242 void
1243 kiblnd_map_rx_descs(kib_conn_t *conn)
1244 {
1245         kib_rx_t       *rx;
1246         struct page    *pg;
1247         int             pg_off;
1248         int             ipg;
1249         int             i;
1250
1251         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1252                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1253                 rx = &conn->ibc_rxs[i];
1254
1255                 rx->rx_conn = conn;
1256                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1257
1258                 rx->rx_msgaddr =
1259                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1260                                               rx->rx_msg, IBLND_MSG_SIZE,
1261                                               DMA_FROM_DEVICE);
1262                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1263                                                   rx->rx_msgaddr));
1264                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1265
1266                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1267                        i, rx->rx_msg, rx->rx_msgaddr,
1268                        (__u64)(page_to_phys(pg) + pg_off));
1269
1270                 pg_off += IBLND_MSG_SIZE;
1271                 LASSERT(pg_off <= PAGE_SIZE);
1272
1273                 if (pg_off == PAGE_SIZE) {
1274                         pg_off = 0;
1275                         ipg++;
1276                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1277                 }
1278         }
1279 }
1280
1281 static void
1282 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1283 {
1284         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1285         kib_tx_t       *tx;
1286         int             i;
1287
1288         LASSERT (tpo->tpo_pool.po_allocated == 0);
1289
1290         if (hdev == NULL)
1291                 return;
1292
1293         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1294                 tx = &tpo->tpo_tx_descs[i];
1295                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1296                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1297                                                           tx->tx_msgaddr),
1298                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1299         }
1300
1301         kiblnd_hdev_decref(hdev);
1302         tpo->tpo_hdev = NULL;
1303 }
1304
1305 static kib_hca_dev_t *
1306 kiblnd_current_hdev(kib_dev_t *dev)
1307 {
1308         kib_hca_dev_t *hdev;
1309         unsigned long  flags;
1310         int            i = 0;
1311
1312         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1313         while (dev->ibd_failover) {
1314                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1315                 if (i++ % 50 == 0)
1316                         CDEBUG(D_NET, "%s: Wait for failover\n",
1317                                dev->ibd_ifname);
1318                 set_current_state(TASK_INTERRUPTIBLE);
1319                 schedule_timeout(cfs_time_seconds(1) / 100);
1320
1321                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1322         }
1323
1324         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1325         hdev = dev->ibd_hdev;
1326
1327         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1328
1329         return hdev;
1330 }
1331
1332 static void
1333 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1334 {
1335         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1336         kib_pool_t     *pool  = &tpo->tpo_pool;
1337         kib_net_t      *net   = pool->po_owner->ps_net;
1338         kib_dev_t      *dev;
1339         struct page    *page;
1340         kib_tx_t       *tx;
1341         int             page_offset;
1342         int             ipage;
1343         int             i;
1344
1345         LASSERT (net != NULL);
1346
1347         dev = net->ibn_dev;
1348
1349         /* pre-mapped messages are not bigger than 1 page */
1350         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1351
1352         /* No fancy arithmetic when we do the buffer calculations */
1353         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1354
1355         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1356
1357         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1358                 page = txpgs->ibp_pages[ipage];
1359                 tx = &tpo->tpo_tx_descs[i];
1360
1361                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1362                                            page_offset);
1363
1364                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1365                                                        tx->tx_msg,
1366                                                        IBLND_MSG_SIZE,
1367                                                        DMA_TO_DEVICE);
1368                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1369                                                   tx->tx_msgaddr));
1370                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1371
1372                 list_add(&tx->tx_list, &pool->po_free_list);
1373
1374                 page_offset += IBLND_MSG_SIZE;
1375                 LASSERT(page_offset <= PAGE_SIZE);
1376
1377                 if (page_offset == PAGE_SIZE) {
1378                         page_offset = 0;
1379                         ipage++;
1380                         LASSERT(ipage <= txpgs->ibp_npages);
1381                 }
1382         }
1383 }
1384
1385 struct ib_mr *
1386 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1387                       int negotiated_nfrags)
1388 {
1389         kib_net_t     *net   = ni->ni_data;
1390         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
1391         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1392         int     mod;
1393         __u16   nfrags;
1394
1395         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1396         mod = tunables->lnd_map_on_demand;
1397         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1398
1399         LASSERT(hdev->ibh_mrs != NULL);
1400
1401         if (mod > 0 && nfrags <= rd->rd_nfrags)
1402                 return NULL;
1403
1404         return hdev->ibh_mrs;
1405 }
1406
1407 static void
1408 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1409 {
1410         LASSERT(fpo->fpo_map_count == 0);
1411
1412         if (fpo->fpo_is_fmr) {
1413                 if (fpo->fmr.fpo_fmr_pool)
1414                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1415         } else {
1416                 struct kib_fast_reg_descriptor *frd, *tmp;
1417                 int i = 0;
1418
1419                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1420                                          frd_list) {
1421                         list_del(&frd->frd_list);
1422 #ifndef HAVE_IB_MAP_MR_SG
1423                         ib_free_fast_reg_page_list(frd->frd_frpl);
1424 #endif
1425                         ib_dereg_mr(frd->frd_mr);
1426                         LIBCFS_FREE(frd, sizeof(*frd));
1427                         i++;
1428                 }
1429                 if (i < fpo->fast_reg.fpo_pool_size)
1430                         CERROR("FastReg pool still has %d regions registered\n",
1431                                 fpo->fast_reg.fpo_pool_size - i);
1432         }
1433
1434         if (fpo->fpo_hdev)
1435                 kiblnd_hdev_decref(fpo->fpo_hdev);
1436
1437         LIBCFS_FREE(fpo, sizeof(*fpo));
1438 }
1439
1440 static void
1441 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1442 {
1443         kib_fmr_pool_t *fpo, *tmp;
1444
1445         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1446                 list_del(&fpo->fpo_list);
1447                 kiblnd_destroy_fmr_pool(fpo);
1448         }
1449 }
1450
1451 static int
1452 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1453                      int ncpts)
1454 {
1455         int size = tunables->lnd_fmr_pool_size / ncpts;
1456
1457         return max(IBLND_FMR_POOL, size);
1458 }
1459
1460 static int
1461 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1462                          int ncpts)
1463 {
1464         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1465
1466         return max(IBLND_FMR_POOL_FLUSH, size);
1467 }
1468
1469 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1470 {
1471         struct ib_fmr_pool_param param = {
1472                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1473                 .page_shift        = PAGE_SHIFT,
1474                 .access            = (IB_ACCESS_LOCAL_WRITE |
1475                                       IB_ACCESS_REMOTE_WRITE),
1476                 .pool_size         = fps->fps_pool_size,
1477                 .dirty_watermark   = fps->fps_flush_trigger,
1478                 .flush_function    = NULL,
1479                 .flush_arg         = NULL,
1480                 .cache             = !!fps->fps_cache };
1481         int rc = 0;
1482
1483         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1484                                                    &param);
1485         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1486                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1487                 if (rc != -ENOSYS)
1488                         CERROR("Failed to create FMR pool: %d\n", rc);
1489                 else
1490                         CERROR("FMRs are not supported\n");
1491         }
1492
1493         return rc;
1494 }
1495
1496 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1497 {
1498         struct kib_fast_reg_descriptor *frd, *tmp;
1499         int i, rc;
1500
1501         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1502         fpo->fast_reg.fpo_pool_size = 0;
1503         for (i = 0; i < fps->fps_pool_size; i++) {
1504                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1505                                  sizeof(*frd));
1506                 if (!frd) {
1507                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1508                         rc = -ENOMEM;
1509                         goto out;
1510                 }
1511                 frd->frd_mr = NULL;
1512
1513 #ifndef HAVE_IB_MAP_MR_SG
1514                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1515                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1516                 if (IS_ERR(frd->frd_frpl)) {
1517                         rc = PTR_ERR(frd->frd_frpl);
1518                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1519                                 rc);
1520                         frd->frd_frpl = NULL;
1521                         goto out_middle;
1522                 }
1523 #endif
1524
1525 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1526                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1527                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1528 #else
1529                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1530                                           IB_MR_TYPE_MEM_REG,
1531                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1532 #endif
1533                 if (IS_ERR(frd->frd_mr)) {
1534                         rc = PTR_ERR(frd->frd_mr);
1535                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1536                         frd->frd_mr = NULL;
1537                         goto out_middle;
1538                 }
1539
1540                 frd->frd_valid = true;
1541
1542                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1543                 fpo->fast_reg.fpo_pool_size++;
1544         }
1545
1546         return 0;
1547
1548 out_middle:
1549         if (frd->frd_mr)
1550                 ib_dereg_mr(frd->frd_mr);
1551 #ifndef HAVE_IB_MAP_MR_SG
1552         if (frd->frd_frpl)
1553                 ib_free_fast_reg_page_list(frd->frd_frpl);
1554 #endif
1555         LIBCFS_FREE(frd, sizeof(*frd));
1556
1557 out:
1558         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1559                                  frd_list) {
1560                 list_del(&frd->frd_list);
1561 #ifndef HAVE_IB_MAP_MR_SG
1562                 ib_free_fast_reg_page_list(frd->frd_frpl);
1563 #endif
1564                 ib_dereg_mr(frd->frd_mr);
1565                 LIBCFS_FREE(frd, sizeof(*frd));
1566         }
1567
1568         return rc;
1569 }
1570
1571 static int
1572 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1573 {
1574         struct ib_device_attr *dev_attr;
1575         kib_dev_t *dev = fps->fps_net->ibn_dev;
1576         kib_fmr_pool_t *fpo;
1577         int rc;
1578
1579 #ifndef HAVE_IB_DEVICE_ATTRS
1580         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1581         if (!dev_attr)
1582                 return -ENOMEM;
1583 #endif
1584
1585         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1586         if (!fpo) {
1587                 rc = -ENOMEM;
1588                 goto out_dev_attr;
1589         }
1590
1591         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1592
1593 #ifdef HAVE_IB_DEVICE_ATTRS
1594         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1595 #else
1596         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1597         if (rc) {
1598                 CERROR("Query device failed for %s: %d\n",
1599                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1600                 goto out_dev_attr;
1601         }
1602 #endif
1603
1604         /* Check for FMR or FastReg support */
1605         fpo->fpo_is_fmr = 0;
1606         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1607             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1608             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1609             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1610                 LCONSOLE_INFO("Using FMR for registration\n");
1611                 fpo->fpo_is_fmr = 1;
1612         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1613                 LCONSOLE_INFO("Using FastReg for registration\n");
1614         } else {
1615                 rc = -ENOSYS;
1616                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1617                 goto out_dev_attr;
1618         }
1619
1620         if (fpo->fpo_is_fmr)
1621                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1622         else
1623                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1624         if (rc)
1625                 goto out_fpo;
1626
1627 #ifndef HAVE_IB_DEVICE_ATTRS
1628         kfree(dev_attr);
1629 #endif
1630         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1631         fpo->fpo_owner    = fps;
1632         *pp_fpo = fpo;
1633
1634         return 0;
1635
1636 out_fpo:
1637         kiblnd_hdev_decref(fpo->fpo_hdev);
1638         LIBCFS_FREE(fpo, sizeof(*fpo));
1639
1640 out_dev_attr:
1641 #ifndef HAVE_IB_DEVICE_ATTRS
1642         kfree(dev_attr);
1643 #endif
1644
1645         return rc;
1646 }
1647
1648 static void
1649 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1650 {
1651         if (fps->fps_net == NULL) /* intialized? */
1652                 return;
1653
1654         spin_lock(&fps->fps_lock);
1655
1656         while (!list_empty(&fps->fps_pool_list)) {
1657                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1658                                                  kib_fmr_pool_t, fpo_list);
1659                 fpo->fpo_failed = 1;
1660                 list_del(&fpo->fpo_list);
1661                 if (fpo->fpo_map_count == 0)
1662                         list_add(&fpo->fpo_list, zombies);
1663                 else
1664                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1665         }
1666
1667         spin_unlock(&fps->fps_lock);
1668 }
1669
1670 static void
1671 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1672 {
1673         if (fps->fps_net != NULL) { /* initialized? */
1674                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1675                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1676         }
1677 }
1678
1679 static int
1680 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1681                         kib_net_t *net,
1682                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1683 {
1684         kib_fmr_pool_t *fpo;
1685         int             rc;
1686
1687         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1688
1689         fps->fps_net = net;
1690         fps->fps_cpt = cpt;
1691
1692         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1693         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1694         fps->fps_cache = tunables->lnd_fmr_cache;
1695
1696         spin_lock_init(&fps->fps_lock);
1697         INIT_LIST_HEAD(&fps->fps_pool_list);
1698         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1699
1700         rc = kiblnd_create_fmr_pool(fps, &fpo);
1701         if (rc == 0)
1702                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1703
1704         return rc;
1705 }
1706
1707 static int
1708 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1709 {
1710         if (fpo->fpo_map_count != 0) /* still in use */
1711                 return 0;
1712         if (fpo->fpo_failed)
1713                 return 1;
1714         return cfs_time_aftereq(now, fpo->fpo_deadline);
1715 }
1716
1717 static int
1718 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1719 {
1720         kib_hca_dev_t   *hdev;
1721         __u64           *pages = tx->tx_pages;
1722         int             npages;
1723         int             size;
1724         int             i;
1725
1726         hdev = tx->tx_pool->tpo_hdev;
1727
1728         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1729                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1730                         size += hdev->ibh_page_size) {
1731                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1732                                            hdev->ibh_page_mask) + size;
1733                 }
1734         }
1735
1736         return npages;
1737 }
1738
1739 void
1740 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1741 {
1742         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1743         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1744         kib_fmr_poolset_t *fps;
1745         cfs_time_t         now = cfs_time_current();
1746         kib_fmr_pool_t    *tmp;
1747         int                rc;
1748
1749         if (!fpo)
1750                 return;
1751
1752         fps = fpo->fpo_owner;
1753         if (fpo->fpo_is_fmr) {
1754                 if (fmr->fmr_pfmr) {
1755                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1756                         LASSERT(!rc);
1757                         fmr->fmr_pfmr = NULL;
1758                 }
1759
1760                 if (status) {
1761                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1762                         LASSERT(!rc);
1763                 }
1764         } else {
1765                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1766
1767                 if (frd) {
1768                         frd->frd_valid = false;
1769                         spin_lock(&fps->fps_lock);
1770                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1771                         spin_unlock(&fps->fps_lock);
1772                         fmr->fmr_frd = NULL;
1773                 }
1774         }
1775         fmr->fmr_pool = NULL;
1776
1777         spin_lock(&fps->fps_lock);
1778         fpo->fpo_map_count--;   /* decref the pool */
1779
1780         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1781                 /* the first pool is persistent */
1782                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1783                         continue;
1784
1785                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1786                         list_move(&fpo->fpo_list, &zombies);
1787                         fps->fps_version++;
1788                 }
1789         }
1790         spin_unlock(&fps->fps_lock);
1791
1792         if (!list_empty(&zombies))
1793                 kiblnd_destroy_fmr_pool_list(&zombies);
1794 }
1795
1796 int
1797 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1798                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1799 {
1800         kib_fmr_pool_t *fpo;
1801         __u64 *pages = tx->tx_pages;
1802         __u64 version;
1803         bool is_rx = (rd != tx->tx_rd);
1804         bool tx_pages_mapped = 0;
1805         int npages = 0;
1806         int rc;
1807
1808 again:
1809         spin_lock(&fps->fps_lock);
1810         version = fps->fps_version;
1811         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1812                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1813                 fpo->fpo_map_count++;
1814
1815                 if (fpo->fpo_is_fmr) {
1816                         struct ib_pool_fmr *pfmr;
1817
1818                         spin_unlock(&fps->fps_lock);
1819
1820                         if (!tx_pages_mapped) {
1821                                 npages = kiblnd_map_tx_pages(tx, rd);
1822                                 tx_pages_mapped = 1;
1823                         }
1824
1825                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1826                                                     pages, npages, iov);
1827                         if (likely(!IS_ERR(pfmr))) {
1828                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1829                                                       : pfmr->fmr->lkey;
1830                                 fmr->fmr_frd  = NULL;
1831                                 fmr->fmr_pfmr = pfmr;
1832                                 fmr->fmr_pool = fpo;
1833                                 return 0;
1834                         }
1835                         rc = PTR_ERR(pfmr);
1836                 } else {
1837                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1838                                 struct kib_fast_reg_descriptor *frd;
1839 #ifdef HAVE_IB_MAP_MR_SG
1840                                 struct ib_reg_wr *wr;
1841                                 int n;
1842 #else
1843                                 struct ib_rdma_wr *wr;
1844                                 struct ib_fast_reg_page_list *frpl;
1845 #endif
1846                                 struct ib_mr *mr;
1847
1848                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1849                                                         struct kib_fast_reg_descriptor,
1850                                                         frd_list);
1851                                 list_del(&frd->frd_list);
1852                                 spin_unlock(&fps->fps_lock);
1853
1854 #ifndef HAVE_IB_MAP_MR_SG
1855                                 frpl = frd->frd_frpl;
1856 #endif
1857                                 mr   = frd->frd_mr;
1858
1859                                 if (!frd->frd_valid) {
1860                                         struct ib_rdma_wr *inv_wr;
1861                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1862
1863                                         inv_wr = &frd->frd_inv_wr;
1864                                         memset(inv_wr, 0, sizeof(*inv_wr));
1865
1866                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1867                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1868                                         inv_wr->wr.ex.invalidate_rkey = key;
1869
1870                                         /* Bump the key */
1871                                         key = ib_inc_rkey(key);
1872                                         ib_update_fast_reg_key(mr, key);
1873                                 }
1874
1875 #ifdef HAVE_IB_MAP_MR_SG
1876 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1877                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1878                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1879 #else
1880                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1881                                                  tx->tx_nfrags, PAGE_SIZE);
1882 #endif
1883                                 if (unlikely(n != tx->tx_nfrags)) {
1884                                         CERROR("Failed to map mr %d/%d "
1885                                                "elements\n", n, tx->tx_nfrags);
1886                                         return n < 0 ? n : -EINVAL;
1887                                 }
1888
1889                                 mr->iova = iov;
1890
1891                                 wr = &frd->frd_fastreg_wr;
1892                                 memset(wr, 0, sizeof(*wr));
1893
1894                                 wr->wr.opcode = IB_WR_REG_MR;
1895                                 wr->wr.wr_id  = IBLND_WID_MR;
1896                                 wr->wr.num_sge = 0;
1897                                 wr->wr.send_flags = 0;
1898                                 wr->mr = mr;
1899                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1900                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1901                                               IB_ACCESS_REMOTE_WRITE);
1902 #else
1903                                 if (!tx_pages_mapped) {
1904                                         npages = kiblnd_map_tx_pages(tx, rd);
1905                                         tx_pages_mapped = 1;
1906                                 }
1907
1908                                 LASSERT(npages <= frpl->max_page_list_len);
1909                                 memcpy(frpl->page_list, pages,
1910                                         sizeof(*pages) * npages);
1911
1912                                 /* Prepare FastReg WR */
1913                                 wr = &frd->frd_fastreg_wr;
1914                                 memset(wr, 0, sizeof(*wr));
1915
1916                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1917                                 wr->wr.wr_id  = IBLND_WID_MR;
1918
1919                                 wr->wr.wr.fast_reg.iova_start = iov;
1920                                 wr->wr.wr.fast_reg.page_list  = frpl;
1921                                 wr->wr.wr.fast_reg.page_list_len = npages;
1922                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1923                                 wr->wr.wr.fast_reg.length = nob;
1924                                 wr->wr.wr.fast_reg.rkey =
1925                                                 is_rx ? mr->rkey : mr->lkey;
1926                                 wr->wr.wr.fast_reg.access_flags =
1927                                                 (IB_ACCESS_LOCAL_WRITE |
1928                                                  IB_ACCESS_REMOTE_WRITE);
1929 #endif
1930
1931                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1932                                 fmr->fmr_frd  = frd;
1933                                 fmr->fmr_pfmr = NULL;
1934                                 fmr->fmr_pool = fpo;
1935                                 return 0;
1936                         }
1937                         spin_unlock(&fps->fps_lock);
1938                         rc = -EBUSY;
1939                 }
1940
1941                 spin_lock(&fps->fps_lock);
1942                 fpo->fpo_map_count--;
1943                 if (rc != -EAGAIN) {
1944                         spin_unlock(&fps->fps_lock);
1945                         return rc;
1946                 }
1947
1948                 /* EAGAIN and ... */
1949                 if (version != fps->fps_version) {
1950                         spin_unlock(&fps->fps_lock);
1951                         goto again;
1952                 }
1953         }
1954
1955         if (fps->fps_increasing) {
1956                 spin_unlock(&fps->fps_lock);
1957                 CDEBUG(D_NET, "Another thread is allocating new "
1958                        "FMR pool, waiting for her to complete\n");
1959                 schedule();
1960                 goto again;
1961
1962         }
1963
1964         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1965                 /* someone failed recently */
1966                 spin_unlock(&fps->fps_lock);
1967                 return -EAGAIN;
1968         }
1969
1970         fps->fps_increasing = 1;
1971         spin_unlock(&fps->fps_lock);
1972
1973         CDEBUG(D_NET, "Allocate new FMR pool\n");
1974         rc = kiblnd_create_fmr_pool(fps, &fpo);
1975         spin_lock(&fps->fps_lock);
1976         fps->fps_increasing = 0;
1977         if (rc == 0) {
1978                 fps->fps_version++;
1979                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1980         } else {
1981                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1982         }
1983         spin_unlock(&fps->fps_lock);
1984
1985         goto again;
1986 }
1987
1988 static void
1989 kiblnd_fini_pool(kib_pool_t *pool)
1990 {
1991         LASSERT(list_empty(&pool->po_free_list));
1992         LASSERT(pool->po_allocated == 0);
1993
1994         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1995 }
1996
1997 static void
1998 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1999 {
2000         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2001
2002         memset(pool, 0, sizeof(kib_pool_t));
2003         INIT_LIST_HEAD(&pool->po_free_list);
2004         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2005         pool->po_owner    = ps;
2006         pool->po_size     = size;
2007 }
2008
2009 static void
2010 kiblnd_destroy_pool_list(struct list_head *head)
2011 {
2012         kib_pool_t *pool;
2013
2014         while (!list_empty(head)) {
2015                 pool = list_entry(head->next, kib_pool_t, po_list);
2016                 list_del(&pool->po_list);
2017
2018                 LASSERT(pool->po_owner != NULL);
2019                 pool->po_owner->ps_pool_destroy(pool);
2020         }
2021 }
2022
2023 static void
2024 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2025 {
2026         if (ps->ps_net == NULL) /* intialized? */
2027                 return;
2028
2029         spin_lock(&ps->ps_lock);
2030         while (!list_empty(&ps->ps_pool_list)) {
2031                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2032                                             kib_pool_t, po_list);
2033                 po->po_failed = 1;
2034                 list_del(&po->po_list);
2035                 if (po->po_allocated == 0)
2036                         list_add(&po->po_list, zombies);
2037                 else
2038                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2039         }
2040         spin_unlock(&ps->ps_lock);
2041 }
2042
2043 static void
2044 kiblnd_fini_poolset(kib_poolset_t *ps)
2045 {
2046         if (ps->ps_net != NULL) { /* initialized? */
2047                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2048                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2049         }
2050 }
2051
2052 static int
2053 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2054                     kib_net_t *net, char *name, int size,
2055                     kib_ps_pool_create_t po_create,
2056                     kib_ps_pool_destroy_t po_destroy,
2057                     kib_ps_node_init_t nd_init,
2058                     kib_ps_node_fini_t nd_fini)
2059 {
2060         kib_pool_t      *pool;
2061         int             rc;
2062
2063         memset(ps, 0, sizeof(kib_poolset_t));
2064
2065         ps->ps_cpt          = cpt;
2066         ps->ps_net          = net;
2067         ps->ps_pool_create  = po_create;
2068         ps->ps_pool_destroy = po_destroy;
2069         ps->ps_node_init    = nd_init;
2070         ps->ps_node_fini    = nd_fini;
2071         ps->ps_pool_size    = size;
2072         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2073             >= sizeof(ps->ps_name))
2074                 return -E2BIG;
2075         spin_lock_init(&ps->ps_lock);
2076         INIT_LIST_HEAD(&ps->ps_pool_list);
2077         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2078
2079         rc = ps->ps_pool_create(ps, size, &pool);
2080         if (rc == 0)
2081                 list_add(&pool->po_list, &ps->ps_pool_list);
2082         else
2083                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2084
2085         return rc;
2086 }
2087
2088 static int
2089 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2090 {
2091         if (pool->po_allocated != 0) /* still in use */
2092                 return 0;
2093         if (pool->po_failed)
2094                 return 1;
2095         return cfs_time_aftereq(now, pool->po_deadline);
2096 }
2097
2098 void
2099 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2100 {
2101         struct list_head zombies = LIST_HEAD_INIT(zombies);
2102         kib_poolset_t   *ps = pool->po_owner;
2103         kib_pool_t      *tmp;
2104         cfs_time_t       now = cfs_time_current();
2105
2106         spin_lock(&ps->ps_lock);
2107
2108         if (ps->ps_node_fini != NULL)
2109                 ps->ps_node_fini(pool, node);
2110
2111         LASSERT(pool->po_allocated > 0);
2112         list_add(node, &pool->po_free_list);
2113         pool->po_allocated--;
2114
2115         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2116                 /* the first pool is persistent */
2117                 if (ps->ps_pool_list.next == &pool->po_list)
2118                         continue;
2119
2120                 if (kiblnd_pool_is_idle(pool, now))
2121                         list_move(&pool->po_list, &zombies);
2122         }
2123         spin_unlock(&ps->ps_lock);
2124
2125         if (!list_empty(&zombies))
2126                 kiblnd_destroy_pool_list(&zombies);
2127 }
2128
2129 struct list_head *
2130 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2131 {
2132         struct list_head        *node;
2133         kib_pool_t              *pool;
2134         int                     rc;
2135         unsigned int            interval = 1;
2136         cfs_time_t              time_before;
2137         unsigned int            trips = 0;
2138
2139 again:
2140         spin_lock(&ps->ps_lock);
2141         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2142                 if (list_empty(&pool->po_free_list))
2143                         continue;
2144
2145                 pool->po_allocated++;
2146                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2147                 node = pool->po_free_list.next;
2148                 list_del(node);
2149
2150                 if (ps->ps_node_init != NULL) {
2151                         /* still hold the lock */
2152                         ps->ps_node_init(pool, node);
2153                 }
2154                 spin_unlock(&ps->ps_lock);
2155                 return node;
2156         }
2157
2158         /* no available tx pool and ... */
2159         if (ps->ps_increasing) {
2160                 /* another thread is allocating a new pool */
2161                 spin_unlock(&ps->ps_lock);
2162                 trips++;
2163                 CDEBUG(D_NET, "Another thread is allocating new "
2164                        "%s pool, waiting %d HZs for her to complete."
2165                        "trips = %d\n",
2166                        ps->ps_name, interval, trips);
2167
2168                 set_current_state(TASK_INTERRUPTIBLE);
2169                 schedule_timeout(interval);
2170                 if (interval < cfs_time_seconds(1))
2171                         interval *= 2;
2172
2173                 goto again;
2174         }
2175
2176         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2177                 /* someone failed recently */
2178                 spin_unlock(&ps->ps_lock);
2179                 return NULL;
2180         }
2181
2182         ps->ps_increasing = 1;
2183         spin_unlock(&ps->ps_lock);
2184
2185         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2186         time_before = cfs_time_current();
2187         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2188         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2189                cfs_time_current() - time_before);
2190
2191         spin_lock(&ps->ps_lock);
2192         ps->ps_increasing = 0;
2193         if (rc == 0) {
2194                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2195         } else {
2196                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2197                 CERROR("Can't allocate new %s pool because out of memory\n",
2198                        ps->ps_name);
2199         }
2200         spin_unlock(&ps->ps_lock);
2201
2202         goto again;
2203 }
2204
2205 static void
2206 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2207 {
2208         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2209         int             i;
2210
2211         LASSERT (pool->po_allocated == 0);
2212
2213         if (tpo->tpo_tx_pages != NULL) {
2214                 kiblnd_unmap_tx_pool(tpo);
2215                 kiblnd_free_pages(tpo->tpo_tx_pages);
2216         }
2217
2218         if (tpo->tpo_tx_descs == NULL)
2219                 goto out;
2220
2221         for (i = 0; i < pool->po_size; i++) {
2222                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2223
2224                 list_del(&tx->tx_list);
2225                 if (tx->tx_pages != NULL)
2226                         LIBCFS_FREE(tx->tx_pages,
2227                                     LNET_MAX_IOV *
2228                                     sizeof(*tx->tx_pages));
2229                 if (tx->tx_frags != NULL)
2230                         LIBCFS_FREE(tx->tx_frags,
2231                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2232                                     sizeof(*tx->tx_frags));
2233                 if (tx->tx_wrq != NULL)
2234                         LIBCFS_FREE(tx->tx_wrq,
2235                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2236                                     sizeof(*tx->tx_wrq));
2237                 if (tx->tx_sge != NULL)
2238                         LIBCFS_FREE(tx->tx_sge,
2239                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2240                                     sizeof(*tx->tx_sge));
2241                 if (tx->tx_rd != NULL)
2242                         LIBCFS_FREE(tx->tx_rd,
2243                                     offsetof(kib_rdma_desc_t,
2244                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2245         }
2246
2247         LIBCFS_FREE(tpo->tpo_tx_descs,
2248                     pool->po_size * sizeof(kib_tx_t));
2249 out:
2250         kiblnd_fini_pool(pool);
2251         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2252 }
2253
2254 static int kiblnd_tx_pool_size(int ncpts)
2255 {
2256         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2257
2258         return max(IBLND_TX_POOL, ntx);
2259 }
2260
2261 static int
2262 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2263 {
2264         int            i;
2265         int            npg;
2266         kib_pool_t    *pool;
2267         kib_tx_pool_t *tpo;
2268
2269         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2270         if (tpo == NULL) {
2271                 CERROR("Failed to allocate TX pool\n");
2272                 return -ENOMEM;
2273         }
2274
2275         pool = &tpo->tpo_pool;
2276         kiblnd_init_pool(ps, pool, size);
2277         tpo->tpo_tx_descs = NULL;
2278         tpo->tpo_tx_pages = NULL;
2279
2280         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2281         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2282                 CERROR("Can't allocate tx pages: %d\n", npg);
2283                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2284                 return -ENOMEM;
2285         }
2286
2287         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2288                          size * sizeof(kib_tx_t));
2289         if (tpo->tpo_tx_descs == NULL) {
2290                 CERROR("Can't allocate %d tx descriptors\n", size);
2291                 ps->ps_pool_destroy(pool);
2292                 return -ENOMEM;
2293         }
2294
2295         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2296
2297         for (i = 0; i < size; i++) {
2298                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2299
2300                 tx->tx_pool = tpo;
2301                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2302                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2303                                          lnet_cpt_table(), ps->ps_cpt,
2304                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2305                         if (tx->tx_pages == NULL)
2306                                 break;
2307                 }
2308
2309                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2310                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2311                                  sizeof(*tx->tx_frags));
2312                 if (tx->tx_frags == NULL)
2313                         break;
2314
2315                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2316
2317                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2318                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2319                                  sizeof(*tx->tx_wrq));
2320                 if (tx->tx_wrq == NULL)
2321                         break;
2322
2323                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2324                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2325                                  sizeof(*tx->tx_sge));
2326                 if (tx->tx_sge == NULL)
2327                         break;
2328
2329                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2330                                  offsetof(kib_rdma_desc_t,
2331                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2332                 if (tx->tx_rd == NULL)
2333                         break;
2334         }
2335
2336         if (i == size) {
2337                 kiblnd_map_tx_pool(tpo);
2338                 *pp_po = pool;
2339                 return 0;
2340         }
2341
2342         ps->ps_pool_destroy(pool);
2343         return -ENOMEM;
2344 }
2345
2346 static void
2347 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2348 {
2349         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2350                                              tps_poolset);
2351         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2352
2353         tx->tx_cookie = tps->tps_next_tx_cookie++;
2354 }
2355
2356 static void
2357 kiblnd_net_fini_pools(kib_net_t *net)
2358 {
2359         int     i;
2360
2361         cfs_cpt_for_each(i, lnet_cpt_table()) {
2362                 kib_tx_poolset_t        *tps;
2363                 kib_fmr_poolset_t       *fps;
2364
2365                 if (net->ibn_tx_ps != NULL) {
2366                         tps = net->ibn_tx_ps[i];
2367                         kiblnd_fini_poolset(&tps->tps_poolset);
2368                 }
2369
2370                 if (net->ibn_fmr_ps != NULL) {
2371                         fps = net->ibn_fmr_ps[i];
2372                         kiblnd_fini_fmr_poolset(fps);
2373                 }
2374         }
2375
2376         if (net->ibn_tx_ps != NULL) {
2377                 cfs_percpt_free(net->ibn_tx_ps);
2378                 net->ibn_tx_ps = NULL;
2379         }
2380
2381         if (net->ibn_fmr_ps != NULL) {
2382                 cfs_percpt_free(net->ibn_fmr_ps);
2383                 net->ibn_fmr_ps = NULL;
2384         }
2385 }
2386
2387 static int
2388 kiblnd_net_init_pools(kib_net_t *net, lnet_ni_t *ni, __u32 *cpts, int ncpts)
2389 {
2390         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2391         unsigned long   flags;
2392         int             cpt;
2393         int             rc;
2394         int             i;
2395
2396         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2397
2398         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2399         if (tunables->lnd_map_on_demand == 0) {
2400                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2401                                            flags);
2402                 goto create_tx_pool;
2403         }
2404
2405         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2406
2407         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2408                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2409                        tunables->lnd_fmr_pool_size,
2410                        *kiblnd_tunables.kib_ntx / 4);
2411                 rc = -EINVAL;
2412                 goto failed;
2413         }
2414
2415         /* TX pool must be created later than FMR, see LU-2268
2416          * for details */
2417         LASSERT(net->ibn_tx_ps == NULL);
2418
2419         /* premapping can fail if ibd_nmr > 1, so we always create
2420          * FMR pool and map-on-demand if premapping failed */
2421
2422         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2423                                            sizeof(kib_fmr_poolset_t));
2424         if (net->ibn_fmr_ps == NULL) {
2425                 CERROR("Failed to allocate FMR pool array\n");
2426                 rc = -ENOMEM;
2427                 goto failed;
2428         }
2429
2430         for (i = 0; i < ncpts; i++) {
2431                 cpt = (cpts == NULL) ? i : cpts[i];
2432                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2433                                              net, tunables);
2434                 if (rc != 0) {
2435                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2436                                cpt, rc);
2437                         goto failed;
2438                 }
2439         }
2440
2441         if (i > 0)
2442                 LASSERT(i == ncpts);
2443
2444  create_tx_pool:
2445         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2446                                           sizeof(kib_tx_poolset_t));
2447         if (net->ibn_tx_ps == NULL) {
2448                 CERROR("Failed to allocate tx pool array\n");
2449                 rc = -ENOMEM;
2450                 goto failed;
2451         }
2452
2453         for (i = 0; i < ncpts; i++) {
2454                 cpt = (cpts == NULL) ? i : cpts[i];
2455                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2456                                          cpt, net, "TX",
2457                                          kiblnd_tx_pool_size(ncpts),
2458                                          kiblnd_create_tx_pool,
2459                                          kiblnd_destroy_tx_pool,
2460                                          kiblnd_tx_init, NULL);
2461                 if (rc != 0) {
2462                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2463                                cpt, rc);
2464                         goto failed;
2465                 }
2466         }
2467
2468         return 0;
2469  failed:
2470         kiblnd_net_fini_pools(net);
2471         LASSERT(rc != 0);
2472         return rc;
2473 }
2474
2475 static int
2476 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2477 {
2478 #ifndef HAVE_IB_DEVICE_ATTRS
2479         struct ib_device_attr *attr;
2480         int                    rc;
2481 #endif
2482
2483         /* It's safe to assume a HCA can handle a page size
2484          * matching that of the native system */
2485         hdev->ibh_page_shift = PAGE_SHIFT;
2486         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2487         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2488
2489 #ifdef HAVE_IB_DEVICE_ATTRS
2490         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2491 #else
2492         LIBCFS_ALLOC(attr, sizeof(*attr));
2493         if (attr == NULL) {
2494                 CERROR("Out of memory\n");
2495                 return -ENOMEM;
2496         }
2497
2498         rc = ib_query_device(hdev->ibh_ibdev, attr);
2499         if (rc == 0)
2500                 hdev->ibh_mr_size = attr->max_mr_size;
2501
2502         LIBCFS_FREE(attr, sizeof(*attr));
2503
2504         if (rc != 0) {
2505                 CERROR("Failed to query IB device: %d\n", rc);
2506                 return rc;
2507         }
2508 #endif
2509
2510         if (hdev->ibh_mr_size == ~0ULL) {
2511                 hdev->ibh_mr_shift = 64;
2512                 return 0;
2513         }
2514
2515         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2516         return -EINVAL;
2517 }
2518
2519 static void
2520 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2521 {
2522         if (hdev->ibh_mrs == NULL)
2523                 return;
2524
2525         ib_dereg_mr(hdev->ibh_mrs);
2526
2527         hdev->ibh_mrs = NULL;
2528 }
2529
2530 void
2531 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2532 {
2533         kiblnd_hdev_cleanup_mrs(hdev);
2534
2535         if (hdev->ibh_pd != NULL)
2536                 ib_dealloc_pd(hdev->ibh_pd);
2537
2538         if (hdev->ibh_cmid != NULL)
2539                 rdma_destroy_id(hdev->ibh_cmid);
2540
2541         LIBCFS_FREE(hdev, sizeof(*hdev));
2542 }
2543
2544 static int
2545 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2546 {
2547         struct ib_mr *mr;
2548         int           rc;
2549         int           acflags = IB_ACCESS_LOCAL_WRITE |
2550                                 IB_ACCESS_REMOTE_WRITE;
2551
2552         rc = kiblnd_hdev_get_attr(hdev);
2553         if (rc != 0)
2554                 return rc;
2555
2556         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2557         if (IS_ERR(mr)) {
2558                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2559                 kiblnd_hdev_cleanup_mrs(hdev);
2560                 return PTR_ERR(mr);
2561         }
2562
2563         hdev->ibh_mrs = mr;
2564
2565         return 0;
2566 }
2567
2568 static int
2569 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2570 {       /* DUMMY */
2571         return 0;
2572 }
2573
2574 static int
2575 kiblnd_dev_need_failover(kib_dev_t *dev)
2576 {
2577         struct rdma_cm_id  *cmid;
2578         struct sockaddr_in  srcaddr;
2579         struct sockaddr_in  dstaddr;
2580         int                 rc;
2581
2582         if (dev->ibd_hdev == NULL || /* initializing */
2583             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2584             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2585                 return 1;
2586
2587         /* XXX: it's UGLY, but I don't have better way to find
2588          * ib-bonding HCA failover because:
2589          *
2590          * a. no reliable CM event for HCA failover...
2591          * b. no OFED API to get ib_device for current net_device...
2592          *
2593          * We have only two choices at this point:
2594          *
2595          * a. rdma_bind_addr(), it will conflict with listener cmid
2596          * b. rdma_resolve_addr() to zero addr */
2597         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2598                                      IB_QPT_RC);
2599         if (IS_ERR(cmid)) {
2600                 rc = PTR_ERR(cmid);
2601                 CERROR("Failed to create cmid for failover: %d\n", rc);
2602                 return rc;
2603         }
2604
2605         memset(&srcaddr, 0, sizeof(srcaddr));
2606         srcaddr.sin_family      = AF_INET;
2607         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2608
2609         memset(&dstaddr, 0, sizeof(dstaddr));
2610         dstaddr.sin_family = AF_INET;
2611         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2612                                (struct sockaddr *)&dstaddr, 1);
2613         if (rc != 0 || cmid->device == NULL) {
2614                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2615                        dev->ibd_ifname, &dev->ibd_ifip,
2616                        cmid->device, rc);
2617                 rdma_destroy_id(cmid);
2618                 return rc;
2619         }
2620
2621         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2622         rdma_destroy_id(cmid);
2623         return rc;
2624 }
2625
2626 int
2627 kiblnd_dev_failover(kib_dev_t *dev)
2628 {
2629         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2630         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2631         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2632         struct rdma_cm_id  *cmid  = NULL;
2633         kib_hca_dev_t      *hdev  = NULL;
2634         kib_hca_dev_t      *old;
2635         struct ib_pd       *pd;
2636         kib_net_t          *net;
2637         struct sockaddr_in  addr;
2638         unsigned long       flags;
2639         int                 rc = 0;
2640         int                 i;
2641
2642         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2643                  dev->ibd_can_failover ||
2644                  dev->ibd_hdev == NULL);
2645
2646         rc = kiblnd_dev_need_failover(dev);
2647         if (rc <= 0)
2648                 goto out;
2649
2650         if (dev->ibd_hdev != NULL &&
2651             dev->ibd_hdev->ibh_cmid != NULL) {
2652                 /* XXX it's not good to close old listener at here,
2653                  * because we can fail to create new listener.
2654                  * But we have to close it now, otherwise rdma_bind_addr
2655                  * will return EADDRINUSE... How crap! */
2656                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2657
2658                 cmid = dev->ibd_hdev->ibh_cmid;
2659                 /* make next schedule of kiblnd_dev_need_failover()
2660                  * return 1 for me */
2661                 dev->ibd_hdev->ibh_cmid  = NULL;
2662                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2663
2664                 rdma_destroy_id(cmid);
2665         }
2666
2667         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2668                                      IB_QPT_RC);
2669         if (IS_ERR(cmid)) {
2670                 rc = PTR_ERR(cmid);
2671                 CERROR("Failed to create cmid for failover: %d\n", rc);
2672                 goto out;
2673         }
2674
2675         memset(&addr, 0, sizeof(addr));
2676         addr.sin_family      = AF_INET;
2677         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2678         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2679
2680         /* Bind to failover device or port */
2681         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2682         if (rc != 0 || cmid->device == NULL) {
2683                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2684                        dev->ibd_ifname, &dev->ibd_ifip,
2685                        cmid->device, rc);
2686                 rdma_destroy_id(cmid);
2687                 goto out;
2688         }
2689
2690         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2691         if (hdev == NULL) {
2692                 CERROR("Failed to allocate kib_hca_dev\n");
2693                 rdma_destroy_id(cmid);
2694                 rc = -ENOMEM;
2695                 goto out;
2696         }
2697
2698         atomic_set(&hdev->ibh_ref, 1);
2699         hdev->ibh_dev   = dev;
2700         hdev->ibh_cmid  = cmid;
2701         hdev->ibh_ibdev = cmid->device;
2702
2703         pd = ib_alloc_pd(cmid->device);
2704         if (IS_ERR(pd)) {
2705                 rc = PTR_ERR(pd);
2706                 CERROR("Can't allocate PD: %d\n", rc);
2707                 goto out;
2708         }
2709
2710         hdev->ibh_pd = pd;
2711
2712         rc = rdma_listen(cmid, 0);
2713         if (rc != 0) {
2714                 CERROR("Can't start new listener: %d\n", rc);
2715                 goto out;
2716         }
2717
2718         rc = kiblnd_hdev_setup_mrs(hdev);
2719         if (rc != 0) {
2720                 CERROR("Can't setup device: %d\n", rc);
2721                 goto out;
2722         }
2723
2724         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2725
2726         old = dev->ibd_hdev;
2727         dev->ibd_hdev = hdev;   /* take over the refcount */
2728         hdev = old;
2729
2730         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2731                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2732                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2733                                             &zombie_tpo);
2734
2735                         if (net->ibn_fmr_ps != NULL)
2736                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2737                                                         &zombie_fpo);
2738                 }
2739         }
2740
2741         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2742  out:
2743         if (!list_empty(&zombie_tpo))
2744                 kiblnd_destroy_pool_list(&zombie_tpo);
2745         if (!list_empty(&zombie_ppo))
2746                 kiblnd_destroy_pool_list(&zombie_ppo);
2747         if (!list_empty(&zombie_fpo))
2748                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2749         if (hdev != NULL)
2750                 kiblnd_hdev_decref(hdev);
2751
2752         if (rc != 0)
2753                 dev->ibd_failed_failover++;
2754         else
2755                 dev->ibd_failed_failover = 0;
2756
2757         return rc;
2758 }
2759
2760 void
2761 kiblnd_destroy_dev (kib_dev_t *dev)
2762 {
2763         LASSERT (dev->ibd_nnets == 0);
2764         LASSERT(list_empty(&dev->ibd_nets));
2765
2766         list_del(&dev->ibd_fail_list);
2767         list_del(&dev->ibd_list);
2768
2769         if (dev->ibd_hdev != NULL)
2770                 kiblnd_hdev_decref(dev->ibd_hdev);
2771
2772         LIBCFS_FREE(dev, sizeof(*dev));
2773 }
2774
2775 static kib_dev_t *
2776 kiblnd_create_dev(char *ifname)
2777 {
2778         struct net_device *netdev;
2779         kib_dev_t         *dev;
2780         __u32              netmask;
2781         __u32              ip;
2782         int                up;
2783         int                rc;
2784
2785         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2786         if (rc != 0) {
2787                 CERROR("Can't query IPoIB interface %s: %d\n",
2788                        ifname, rc);
2789                 return NULL;
2790         }
2791
2792         if (!up) {
2793                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2794                 return NULL;
2795         }
2796
2797         LIBCFS_ALLOC(dev, sizeof(*dev));
2798         if (dev == NULL)
2799                 return NULL;
2800
2801         netdev = dev_get_by_name(&init_net, ifname);
2802         if (netdev == NULL) {
2803                 dev->ibd_can_failover = 0;
2804         } else {
2805                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2806                 dev_put(netdev);
2807         }
2808
2809         INIT_LIST_HEAD(&dev->ibd_nets);
2810         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2811         INIT_LIST_HEAD(&dev->ibd_fail_list);
2812         dev->ibd_ifip = ip;
2813         strcpy(&dev->ibd_ifname[0], ifname);
2814
2815         /* initialize the device */
2816         rc = kiblnd_dev_failover(dev);
2817         if (rc != 0) {
2818                 CERROR("Can't initialize device: %d\n", rc);
2819                 LIBCFS_FREE(dev, sizeof(*dev));
2820                 return NULL;
2821         }
2822
2823         list_add_tail(&dev->ibd_list,
2824                           &kiblnd_data.kib_devs);
2825         return dev;
2826 }
2827
2828 static void
2829 kiblnd_base_shutdown(void)
2830 {
2831         struct kib_sched_info   *sched;
2832         int                     i;
2833
2834         LASSERT(list_empty(&kiblnd_data.kib_devs));
2835
2836         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2837                atomic_read(&libcfs_kmemory));
2838
2839         switch (kiblnd_data.kib_init) {
2840         default:
2841                 LBUG();
2842
2843         case IBLND_INIT_ALL:
2844         case IBLND_INIT_DATA:
2845                 LASSERT (kiblnd_data.kib_peers != NULL);
2846                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2847                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2848                 }
2849                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2850                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2851                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2852                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2853
2854                 /* flag threads to terminate; wake and wait for them to die */
2855                 kiblnd_data.kib_shutdown = 1;
2856
2857                 /* NB: we really want to stop scheduler threads net by net
2858                  * instead of the whole module, this should be improved
2859                  * with dynamic configuration LNet */
2860                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2861                         wake_up_all(&sched->ibs_waitq);
2862
2863                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2864                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2865
2866                 i = 2;
2867                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2868                         i++;
2869                         /* power of 2? */
2870                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2871                                "Waiting for %d threads to terminate\n",
2872                                atomic_read(&kiblnd_data.kib_nthreads));
2873                         set_current_state(TASK_UNINTERRUPTIBLE);
2874                         schedule_timeout(cfs_time_seconds(1));
2875                 }
2876
2877                 /* fall through */
2878
2879         case IBLND_INIT_NOTHING:
2880                 break;
2881         }
2882
2883         if (kiblnd_data.kib_peers != NULL) {
2884                 LIBCFS_FREE(kiblnd_data.kib_peers,
2885                             sizeof(struct list_head) *
2886                             kiblnd_data.kib_peer_hash_size);
2887         }
2888
2889         if (kiblnd_data.kib_scheds != NULL)
2890                 cfs_percpt_free(kiblnd_data.kib_scheds);
2891
2892         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2893                atomic_read(&libcfs_kmemory));
2894
2895         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2896         module_put(THIS_MODULE);
2897 }
2898
2899 static void
2900 kiblnd_shutdown (lnet_ni_t *ni)
2901 {
2902         kib_net_t        *net = ni->ni_data;
2903         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2904         int               i;
2905         unsigned long     flags;
2906
2907         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2908
2909         if (net == NULL)
2910                 goto out;
2911
2912         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2913                atomic_read(&libcfs_kmemory));
2914
2915         write_lock_irqsave(g_lock, flags);
2916         net->ibn_shutdown = 1;
2917         write_unlock_irqrestore(g_lock, flags);
2918
2919         switch (net->ibn_init) {
2920         default:
2921                 LBUG();
2922
2923         case IBLND_INIT_ALL:
2924                 /* nuke all existing peers within this net */
2925                 kiblnd_del_peer(ni, LNET_NID_ANY);
2926
2927                 /* Wait for all peer state to clean up */
2928                 i = 2;
2929                 while (atomic_read(&net->ibn_npeers) != 0) {
2930                         i++;
2931                         /* power of 2? */
2932                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2933                                "%s: waiting for %d peers to disconnect\n",
2934                                libcfs_nid2str(ni->ni_nid),
2935                                atomic_read(&net->ibn_npeers));
2936                         set_current_state(TASK_UNINTERRUPTIBLE);
2937                         schedule_timeout(cfs_time_seconds(1));
2938                 }
2939
2940                 kiblnd_net_fini_pools(net);
2941
2942                 write_lock_irqsave(g_lock, flags);
2943                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2944                 net->ibn_dev->ibd_nnets--;
2945                 list_del(&net->ibn_list);
2946                 write_unlock_irqrestore(g_lock, flags);
2947
2948                 /* fall through */
2949
2950         case IBLND_INIT_NOTHING:
2951                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2952
2953                 if (net->ibn_dev != NULL &&
2954                     net->ibn_dev->ibd_nnets == 0)
2955                         kiblnd_destroy_dev(net->ibn_dev);
2956
2957                 break;
2958         }
2959
2960         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2961                atomic_read(&libcfs_kmemory));
2962
2963         net->ibn_init = IBLND_INIT_NOTHING;
2964         ni->ni_data = NULL;
2965
2966         LIBCFS_FREE(net, sizeof(*net));
2967
2968 out:
2969         if (list_empty(&kiblnd_data.kib_devs))
2970                 kiblnd_base_shutdown();
2971         return;
2972 }
2973
2974 static int
2975 kiblnd_base_startup(void)
2976 {
2977         struct kib_sched_info   *sched;
2978         int                     rc;
2979         int                     i;
2980
2981         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2982
2983         try_module_get(THIS_MODULE);
2984         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2985
2986         rwlock_init(&kiblnd_data.kib_global_lock);
2987
2988         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2989         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2990
2991         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2992         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2993                      sizeof(struct list_head) *
2994                      kiblnd_data.kib_peer_hash_size);
2995         if (kiblnd_data.kib_peers == NULL)
2996                 goto failed;
2997
2998         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2999                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3000
3001         spin_lock_init(&kiblnd_data.kib_connd_lock);
3002         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3003         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3004         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3005         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3006
3007         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3008         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3009
3010         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3011                                                   sizeof(*sched));
3012         if (kiblnd_data.kib_scheds == NULL)
3013                 goto failed;
3014
3015         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3016                 int     nthrs;
3017
3018                 spin_lock_init(&sched->ibs_lock);
3019                 INIT_LIST_HEAD(&sched->ibs_conns);
3020                 init_waitqueue_head(&sched->ibs_waitq);
3021
3022                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3023                 if (*kiblnd_tunables.kib_nscheds > 0) {
3024                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3025                 } else {
3026                         /* max to half of CPUs, another half is reserved for
3027                          * upper layer modules */
3028                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3029                 }
3030
3031                 sched->ibs_nthreads_max = nthrs;
3032                 sched->ibs_cpt = i;
3033         }
3034
3035         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3036
3037         /* lists/ptrs/locks initialised */
3038         kiblnd_data.kib_init = IBLND_INIT_DATA;
3039         /*****************************************************/
3040
3041         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3042         if (rc != 0) {
3043                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3044                 goto failed;
3045         }
3046
3047         if (*kiblnd_tunables.kib_dev_failover != 0)
3048                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3049                                          "kiblnd_failover");
3050
3051         if (rc != 0) {
3052                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3053                 goto failed;
3054         }
3055
3056         /* flag everything initialised */
3057         kiblnd_data.kib_init = IBLND_INIT_ALL;
3058         /*****************************************************/
3059
3060         return 0;
3061
3062  failed:
3063         kiblnd_base_shutdown();
3064         return -ENETDOWN;
3065 }
3066
3067 static int
3068 kiblnd_start_schedulers(struct kib_sched_info *sched)
3069 {
3070         int     rc = 0;
3071         int     nthrs;
3072         int     i;
3073
3074         if (sched->ibs_nthreads == 0) {
3075                 if (*kiblnd_tunables.kib_nscheds > 0) {
3076                         nthrs = sched->ibs_nthreads_max;
3077                 } else {
3078                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3079                                                sched->ibs_cpt);
3080                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3081                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3082                 }
3083         } else {
3084                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3085                 /* increase one thread if there is new interface */
3086                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3087         }
3088
3089         for (i = 0; i < nthrs; i++) {
3090                 long    id;
3091                 char    name[20];
3092                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3093                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3094                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3095                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3096                 if (rc == 0)
3097                         continue;
3098
3099                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3100                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3101                 break;
3102         }
3103
3104         sched->ibs_nthreads += i;
3105         return rc;
3106 }
3107
3108 static int
3109 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3110 {
3111         int     cpt;
3112         int     rc;
3113         int     i;
3114
3115         for (i = 0; i < ncpts; i++) {
3116                 struct kib_sched_info *sched;
3117
3118                 cpt = (cpts == NULL) ? i : cpts[i];
3119                 sched = kiblnd_data.kib_scheds[cpt];
3120
3121                 if (!newdev && sched->ibs_nthreads > 0)
3122                         continue;
3123
3124                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3125                 if (rc != 0) {
3126                         CERROR("Failed to start scheduler threads for %s\n",
3127                                dev->ibd_ifname);
3128                         return rc;
3129                 }
3130         }
3131         return 0;
3132 }
3133
3134 static kib_dev_t *
3135 kiblnd_dev_search(char *ifname)
3136 {
3137         kib_dev_t       *alias = NULL;
3138         kib_dev_t       *dev;
3139         char            *colon;
3140         char            *colon2;
3141
3142         colon = strchr(ifname, ':');
3143         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3144                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3145                         return dev;
3146
3147                 if (alias != NULL)
3148                         continue;
3149
3150                 colon2 = strchr(dev->ibd_ifname, ':');
3151                 if (colon != NULL)
3152                         *colon = 0;
3153                 if (colon2 != NULL)
3154                         *colon2 = 0;
3155
3156                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3157                         alias = dev;
3158
3159                 if (colon != NULL)
3160                         *colon = ':';
3161                 if (colon2 != NULL)
3162                         *colon2 = ':';
3163         }
3164         return alias;
3165 }
3166
3167 static int
3168 kiblnd_startup (lnet_ni_t *ni)
3169 {
3170         char                     *ifname;
3171         kib_dev_t                *ibdev = NULL;
3172         kib_net_t                *net;
3173         struct timeval            tv;
3174         unsigned long             flags;
3175         int                       rc;
3176         int                       newdev;
3177
3178         LASSERT (ni->ni_lnd == &the_o2iblnd);
3179
3180         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3181                 rc = kiblnd_base_startup();
3182                 if (rc != 0)
3183                         return rc;
3184         }
3185
3186         LIBCFS_ALLOC(net, sizeof(*net));
3187         ni->ni_data = net;
3188         if (net == NULL)
3189                 goto failed;
3190
3191         do_gettimeofday(&tv);
3192         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3193
3194         kiblnd_tunables_setup(ni);
3195
3196         if (ni->ni_interfaces[0] != NULL) {
3197                 /* Use the IPoIB interface specified in 'networks=' */
3198
3199                 CLASSERT (LNET_MAX_INTERFACES > 1);
3200                 if (ni->ni_interfaces[1] != NULL) {
3201                         CERROR("Multiple interfaces not supported\n");
3202                         goto failed;
3203                 }
3204
3205                 ifname = ni->ni_interfaces[0];
3206         } else {
3207                 ifname = *kiblnd_tunables.kib_default_ipif;
3208         }
3209
3210         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3211                 CERROR("IPoIB interface name too long: %s\n", ifname);
3212                 goto failed;
3213         }
3214
3215         ibdev = kiblnd_dev_search(ifname);
3216
3217         newdev = ibdev == NULL;
3218         /* hmm...create kib_dev even for alias */
3219         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3220                 ibdev = kiblnd_create_dev(ifname);
3221
3222         if (ibdev == NULL)
3223                 goto failed;
3224
3225         net->ibn_dev = ibdev;
3226         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3227
3228         rc = kiblnd_dev_start_threads(ibdev, newdev,
3229                                       ni->ni_cpts, ni->ni_ncpts);
3230         if (rc != 0)
3231                 goto failed;
3232
3233         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3234         if (rc != 0) {
3235                 CERROR("Failed to initialize NI pools: %d\n", rc);
3236                 goto failed;
3237         }
3238
3239         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3240         ibdev->ibd_nnets++;
3241         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3242         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3243
3244         net->ibn_init = IBLND_INIT_ALL;
3245
3246         return 0;
3247
3248 failed:
3249         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3250                 kiblnd_destroy_dev(ibdev);
3251
3252         kiblnd_shutdown(ni);
3253
3254         CDEBUG(D_NET, "kiblnd_startup failed\n");
3255         return -ENETDOWN;
3256 }
3257
3258 static lnd_t the_o2iblnd = {
3259         .lnd_type       = O2IBLND,
3260         .lnd_startup    = kiblnd_startup,
3261         .lnd_shutdown   = kiblnd_shutdown,
3262         .lnd_ctl        = kiblnd_ctl,
3263         .lnd_query      = kiblnd_query,
3264         .lnd_send       = kiblnd_send,
3265         .lnd_recv       = kiblnd_recv,
3266 };
3267
3268 static void __exit ko2iblnd_exit(void)
3269 {
3270         lnet_unregister_lnd(&the_o2iblnd);
3271 }
3272
3273 static int __init ko2iblnd_init(void)
3274 {
3275         int rc;
3276
3277         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3278         CLASSERT(offsetof(kib_msg_t,
3279                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3280                  IBLND_MSG_SIZE);
3281         CLASSERT(offsetof(kib_msg_t,
3282                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3283                  <= IBLND_MSG_SIZE);
3284
3285         rc = kiblnd_tunables_init();
3286         if (rc != 0)
3287                 return rc;
3288
3289         lnet_register_lnd(&the_o2iblnd);
3290
3291         return 0;
3292 }
3293
3294 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3295 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3296 MODULE_VERSION("2.8.0");
3297 MODULE_LICENSE("GPL");
3298
3299 module_init(ko2iblnd_init);
3300 module_exit(ko2iblnd_exit);