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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/klnds/o2iblnd/o2iblnd.c
  32  *
  33  * Author: Eric Barton <eric@bartonsoftware.com>
  34  */
  35
  36 #include <asm/page.h>
  37 #include <linux/ethtool.h>
  38 #include <linux/inetdevice.h>
  39
  40 #include "o2iblnd.h"
  41
  42 static const struct lnet_lnd the_o2iblnd;
  43
  44 struct kib_data kiblnd_data;
  45
  46 static __u32
  47 kiblnd_cksum (void *ptr, int nob)
  48 {
  49         char  *c  = ptr;
  50         __u32  sum = 0;
  51
  52         while (nob-- > 0)
  53                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  54
  55         /* ensure I don't return 0 (== no checksum) */
  56         return (sum == 0) ? 1 : sum;
  57 }
  58
  59 static char *
  60 kiblnd_msgtype2str(int type)
  61 {
  62         switch (type) {
  63         case IBLND_MSG_CONNREQ:
  64                 return "CONNREQ";
  65
  66         case IBLND_MSG_CONNACK:
  67                 return "CONNACK";
  68
  69         case IBLND_MSG_NOOP:
  70                 return "NOOP";
  71
  72         case IBLND_MSG_IMMEDIATE:
  73                 return "IMMEDIATE";
  74
  75         case IBLND_MSG_PUT_REQ:
  76                 return "PUT_REQ";
  77
  78         case IBLND_MSG_PUT_NAK:
  79                 return "PUT_NAK";
  80
  81         case IBLND_MSG_PUT_ACK:
  82                 return "PUT_ACK";
  83
  84         case IBLND_MSG_PUT_DONE:
  85                 return "PUT_DONE";
  86
  87         case IBLND_MSG_GET_REQ:
  88                 return "GET_REQ";
  89
  90         case IBLND_MSG_GET_DONE:
  91                 return "GET_DONE";
  92
  93         default:
  94                 return "???";
  95         }
  96 }
  97
  98 static int
  99 kiblnd_msgtype2size(int type)
 100 {
 101         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 102
 103         switch (type) {
 104         case IBLND_MSG_CONNREQ:
 105         case IBLND_MSG_CONNACK:
 106                 return hdr_size + sizeof(struct kib_connparams);
 107
 108         case IBLND_MSG_NOOP:
 109                 return hdr_size;
 110
 111         case IBLND_MSG_IMMEDIATE:
 112                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 113
 114         case IBLND_MSG_PUT_REQ:
 115                 return hdr_size + sizeof(struct kib_putreq_msg);
 116
 117         case IBLND_MSG_PUT_ACK:
 118                 return hdr_size + sizeof(struct kib_putack_msg);
 119
 120         case IBLND_MSG_GET_REQ:
 121                 return hdr_size + sizeof(struct kib_get_msg);
 122
 123         case IBLND_MSG_PUT_NAK:
 124         case IBLND_MSG_PUT_DONE:
 125         case IBLND_MSG_GET_DONE:
 126                 return hdr_size + sizeof(struct kib_completion_msg);
 127         default:
 128                 return -1;
 129         }
 130 }
 131
 132 static int kiblnd_unpack_rd(struct kib_msg *msg, bool flip)
 133 {
 134         struct kib_rdma_desc *rd;
 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
 153         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 154                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 155                        n, IBLND_MAX_RDMA_FRAGS);
 156                 return 1;
 157         }
 158
 159         nob = offsetof(struct kib_msg, ibm_u) +
 160                 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 161
 162         if (msg->ibm_nob < nob) {
 163                 CERROR("Short %s: %d(%d)\n",
 164                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 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 kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
 180                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         struct kib_net *net = ni->ni_data;
 183
 184         /* CAVEAT EMPTOR! all message fields not set here should have been
 185          * initialised previously.
 186          */
 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   = lnet_nid_to_nid4(&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 kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 205 {
 206         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 207         __u32 msg_cksum;
 208         __u16 version;
 209         int msg_nob;
 210         bool flip;
 211
 212         /* 6 bytes are enough to have received magic + version */
 213         if (nob < 6) {
 214                 CERROR("Short message: %d\n", nob);
 215                 return -EPROTO;
 216         }
 217
 218         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 219                 flip = false;
 220         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 221                 flip = true;
 222         } else {
 223                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 224                 return -EPROTO;
 225         }
 226
 227         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 228         if (version != IBLND_MSG_VERSION &&
 229             version != IBLND_MSG_VERSION_1) {
 230                 CERROR("Bad version: %x\n", version);
 231                 return -EPROTO;
 232         }
 233
 234         if (nob < hdr_size) {
 235                 CERROR("Short message: %d\n", nob);
 236                 return -EPROTO;
 237         }
 238
 239         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 240         if (msg_nob > nob) {
 241                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 242                 return -EPROTO;
 243         }
 244
 245         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 246          * gets flipped
 247          */
 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_ni endianness */
 260                 msg->ibm_version = version;
 261                 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
 262                 BUILD_BUG_ON(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(struct lnet_ni *ni, struct kib_peer_ni **peerp,
 318                    lnet_nid_t nid)
 319 {
 320         struct kib_peer_ni *peer_ni;
 321         struct kib_net *net = ni->ni_data;
 322         int cpt = lnet_cpt_of_nid(nid, ni);
 323         unsigned long flags;
 324
 325         LASSERT(net != NULL);
 326         LASSERT(nid != LNET_NID_ANY);
 327
 328         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 329         if (!peer_ni) {
 330                 CERROR("Cannot allocate peer_ni\n");
 331                 return -ENOMEM;
 332         }
 333
 334         peer_ni->ibp_ni = ni;
 335         peer_ni->ibp_nid = nid;
 336         peer_ni->ibp_error = 0;
 337         peer_ni->ibp_last_alive = 0;
 338         peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
 339         peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
 340         peer_ni->ibp_queue_depth_mod = 0;       /* try to use the default */
 341         kref_init(&peer_ni->ibp_kref);
 342
 343         INIT_HLIST_NODE(&peer_ni->ibp_list);
 344         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 345         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 346
 347         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 348
 349         /* always called with a ref on ni, which prevents ni being shutdown */
 350         LASSERT(net->ibn_shutdown == 0);
 351
 352         /* npeers only grows with the global lock held */
 353         atomic_inc(&net->ibn_npeers);
 354
 355         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 356
 357         *peerp = peer_ni;
 358         return 0;
 359 }
 360
 361 void
 362 kiblnd_destroy_peer(struct kref *kref)
 363 {
 364         struct kib_peer_ni *peer_ni = container_of(kref, struct kib_peer_ni,
 365                                                    ibp_kref);
 366         struct kib_net *net = peer_ni->ibp_ni->ni_data;
 367
 368         LASSERT(net != NULL);
 369         LASSERT(!kiblnd_peer_active(peer_ni));
 370         LASSERT(kiblnd_peer_idle(peer_ni));
 371         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 374
 375         /* NB a peer_ni's connections keep a reference on their peer_ni until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer_ni has been cleaned up when its refcount drops to
 378          * zero.
 379          */
 380         if (atomic_dec_and_test(&net->ibn_npeers))
 381                 wake_up_var(&net->ibn_npeers);
 382 }
 383
 384 struct kib_peer_ni *
 385 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 386 {
 387         /* the caller is responsible for accounting the additional reference
 388          * that this creates
 389          */
 390         struct kib_peer_ni *peer_ni;
 391
 392         hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
 393                                ibp_list, nid) {
 394                 LASSERT(!kiblnd_peer_idle(peer_ni));
 395
 396                 /*
 397                  * Match a peer if its NID and the NID of the local NI it
 398                  * communicates over are the same. Otherwise don't match
 399                  * the peer, which will result in a new lnd peer being
 400                  * created.
 401                  */
 402                 if (peer_ni->ibp_nid != nid ||
 403                     !nid_same(&peer_ni->ibp_ni->ni_nid, &ni->ni_nid))
 404                         continue;
 405
 406                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 407                        peer_ni, libcfs_nid2str(nid),
 408                        kref_read(&peer_ni->ibp_kref),
 409                        peer_ni->ibp_version);
 410                 return peer_ni;
 411         }
 412         return NULL;
 413 }
 414
 415 void
 416 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
 417 {
 418         LASSERT(list_empty(&peer_ni->ibp_conns));
 419
 420         LASSERT(kiblnd_peer_active(peer_ni));
 421         hlist_del_init(&peer_ni->ibp_list);
 422         /* lose peerlist's ref */
 423         kiblnd_peer_decref(peer_ni);
 424 }
 425
 426 static int
 427 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
 428                      lnet_nid_t *nidp, int *count)
 429 {
 430         struct kib_peer_ni              *peer_ni;
 431         int                      i;
 432         unsigned long            flags;
 433
 434         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 435
 436         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 437                 LASSERT(!kiblnd_peer_idle(peer_ni));
 438
 439                 if (peer_ni->ibp_ni != ni)
 440                         continue;
 441
 442                 if (index-- > 0)
 443                         continue;
 444
 445                 *nidp = peer_ni->ibp_nid;
 446                 *count = kref_read(&peer_ni->ibp_kref);
 447
 448                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 449                 return 0;
 450         }
 451
 452         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 453         return -ENOENT;
 454 }
 455
 456 static void
 457 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
 458 {
 459         struct kib_conn *cnxt;
 460         struct kib_conn *conn;
 461
 462         if (list_empty(&peer_ni->ibp_conns)) {
 463                 kiblnd_unlink_peer_locked(peer_ni);
 464         } else {
 465                 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 466                                          ibc_list)
 467                         kiblnd_close_conn_locked(conn, 0);
 468                 /* NB closing peer_ni's last conn unlinked it. */
 469         }
 470         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 471          * last ref on it. */
 472 }
 473
 474 static int
 475 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 476 {
 477         LIST_HEAD(zombies);
 478         struct hlist_node *pnxt;
 479         struct kib_peer_ni *peer_ni;
 480         int lo;
 481         int hi;
 482         int i;
 483         unsigned long flags;
 484         int rc = -ENOENT;
 485
 486         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 487
 488         if (nid != LNET_NID_ANY) {
 489                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
 490                 hi = lo;
 491         } else {
 492                 lo = 0;
 493                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
 494         }
 495
 496         for (i = lo; i <= hi; i++) {
 497                 hlist_for_each_entry_safe(peer_ni, pnxt,
 498                                           &kiblnd_data.kib_peers[i], ibp_list) {
 499                         LASSERT(!kiblnd_peer_idle(peer_ni));
 500
 501                         if (peer_ni->ibp_ni != ni)
 502                                 continue;
 503
 504                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 505                                 continue;
 506
 507                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 508                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 509
 510                                 list_splice_init(&peer_ni->ibp_tx_queue,
 511                                                  &zombies);
 512                         }
 513
 514                         kiblnd_del_peer_locked(peer_ni);
 515                         rc = 0;         /* matched something */
 516                 }
 517         }
 518
 519         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 520
 521         kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
 522
 523         return rc;
 524 }
 525
 526 static struct kib_conn *
 527 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 528 {
 529         struct kib_peer_ni *peer_ni;
 530         struct kib_conn *conn;
 531         int i;
 532         unsigned long flags;
 533
 534         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 535
 536         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 537                 LASSERT(!kiblnd_peer_idle(peer_ni));
 538
 539                 if (peer_ni->ibp_ni != ni)
 540                         continue;
 541
 542                 list_for_each_entry(conn, &peer_ni->ibp_conns,
 543                                     ibc_list) {
 544                         if (index-- > 0)
 545                                 continue;
 546
 547                         kiblnd_conn_addref(conn);
 548                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 549                                                flags);
 550                         return conn;
 551                 }
 552         }
 553
 554         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 555         return NULL;
 556 }
 557
 558 static void
 559 kiblnd_debug_rx(struct kib_rx *rx)
 560 {
 561         CDEBUG(D_CONSOLE, "      %p msg_type %x cred %d\n",
 562                rx, rx->rx_msg->ibm_type,
 563                rx->rx_msg->ibm_credits);
 564 }
 565
 566 static void
 567 kiblnd_debug_tx(struct kib_tx *tx)
 568 {
 569         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 570                "cookie %#llx msg %s%s type %x cred %d\n",
 571                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 572                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 573                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 574                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 575                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 576 }
 577
 578 void
 579 kiblnd_debug_conn(struct kib_conn *conn)
 580 {
 581         struct list_head        *tmp;
 582         int                     i;
 583
 584         spin_lock(&conn->ibc_lock);
 585
 586         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 587                atomic_read(&conn->ibc_refcount), conn,
 588                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 589         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 590                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 591                conn->ibc_nsends_posted, conn->ibc_credits,
 592                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 593         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 594
 595         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 596         list_for_each(tmp, &conn->ibc_early_rxs)
 597                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 598
 599         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 600         list_for_each(tmp, &conn->ibc_tx_noops)
 601                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 602
 603         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 604         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 605                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 606
 607         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 608         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 609                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 610
 611         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 612         list_for_each(tmp, &conn->ibc_tx_queue)
 613                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 614
 615         CDEBUG(D_CONSOLE, "   active_txs:\n");
 616         list_for_each(tmp, &conn->ibc_active_txs)
 617                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 618
 619         CDEBUG(D_CONSOLE, "   rxs:\n");
 620         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 621                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 622
 623         spin_unlock(&conn->ibc_lock);
 624 }
 625
 626 static void
 627 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 628 {
 629         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 630         if (cmid->route.path_rec == NULL)
 631                 return;
 632
 633         if (*kiblnd_tunables.kib_ib_mtu)
 634                 cmid->route.path_rec->mtu =
 635                         ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
 636 }
 637
 638 static int
 639 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 640 {
 641         cpumask_var_t   *mask;
 642         int             vectors;
 643         int             off;
 644         int             i;
 645         lnet_nid_t      ibp_nid;
 646
 647         vectors = conn->ibc_cmid->device->num_comp_vectors;
 648         if (vectors <= 1)
 649                 return 0;
 650
 651         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 652
 653         /* hash NID to CPU id in this partition... */
 654         ibp_nid = conn->ibc_peer->ibp_nid;
 655         off = do_div(ibp_nid, cpumask_weight(*mask));
 656         for_each_cpu(i, *mask) {
 657                 if (off-- == 0)
 658                         return i % vectors;
 659         }
 660
 661         LBUG();
 662         return 1;
 663 }
 664
 665 /*
 666  * Get the scheduler bound to this CPT. If the scheduler has no
 667  * threads, which means that the CPT has no CPUs, then grab the
 668  * next scheduler that we can use.
 669  *
 670  * This case would be triggered if a NUMA node is configured with
 671  * no associated CPUs.
 672  */
 673 static struct kib_sched_info *
 674 kiblnd_get_scheduler(int cpt)
 675 {
 676         struct kib_sched_info *sched;
 677         int i;
 678
 679         sched = kiblnd_data.kib_scheds[cpt];
 680
 681         if (sched->ibs_nthreads > 0)
 682                 return sched;
 683
 684         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 685                 if (sched->ibs_nthreads > 0) {
 686                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 687                                         cpt, sched->ibs_cpt);
 688                         return sched;
 689                 }
 690         }
 691
 692         return NULL;
 693 }
 694
 695 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 696 {
 697         /*
 698          * One WR for the LNet message
 699          * And ibc_max_frags for the transfer WRs
 700          */
 701         int ret;
 702         int multiplier = 1 + conn->ibc_max_frags;
 703
 704         /* FastReg needs two extra WRs for map and invalidate */
 705         if (IS_FAST_REG_DEV(conn->ibc_hdev->ibh_dev))
 706                 multiplier += 2;
 707
 708         /* account for a maximum of ibc_queue_depth in-flight transfers */
 709         ret = multiplier * conn->ibc_queue_depth;
 710
 711         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 712                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 713                        "request size %d larger than maximum %d device "
 714                        "can handle\n", conn->ibc_queue_depth, ret,
 715                        conn->ibc_hdev->ibh_max_qp_wr);
 716                 conn->ibc_queue_depth =
 717                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 718         }
 719
 720         /* don't go beyond the maximum the device can handle */
 721         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 722 }
 723
 724 struct kib_conn *
 725 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 726                    int state, int version)
 727 {
 728         /* CAVEAT EMPTOR:
 729          * If the new conn is created successfully it takes over the caller's
 730          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 731          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 732          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 733          * to destroy 'cmid' here since I'm called from the CM which still has
 734          * its ref on 'cmid'). */
 735         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 736         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 737         struct kib_dev *dev;
 738         struct ib_qp_init_attr init_qp_attr = {};
 739         struct kib_sched_info   *sched;
 740 #ifdef HAVE_IB_CQ_INIT_ATTR
 741         struct ib_cq_init_attr  cq_attr = {};
 742 #endif
 743         struct kib_conn *conn;
 744         struct ib_cq            *cq;
 745         unsigned long           flags;
 746         int                     cpt;
 747         int                     rc;
 748         int                     i;
 749
 750         LASSERT(net != NULL);
 751         LASSERT(!in_interrupt());
 752
 753         dev = net->ibn_dev;
 754
 755         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 756         sched = kiblnd_get_scheduler(cpt);
 757
 758         if (sched == NULL) {
 759                 CERROR("no schedulers available. node is unhealthy\n");
 760                 goto failed_0;
 761         }
 762
 763         /*
 764          * The cpt might have changed if we ended up selecting a non cpt
 765          * native scheduler. So use the scheduler's cpt instead.
 766          */
 767         cpt = sched->ibs_cpt;
 768
 769         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 770         if (conn == NULL) {
 771                 CERROR("Can't allocate connection for %s\n",
 772                        libcfs_nid2str(peer_ni->ibp_nid));
 773                 goto failed_0;
 774         }
 775
 776         conn->ibc_state = IBLND_CONN_INIT;
 777         conn->ibc_version = version;
 778         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 779         cmid->context = conn;                   /* for future CM callbacks */
 780         conn->ibc_cmid = cmid;
 781         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 782         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 783         conn->ibc_rxs = NULL;
 784         conn->ibc_rx_pages = NULL;
 785
 786         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 787         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 788         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 789         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 790         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 791         INIT_LIST_HEAD(&conn->ibc_active_txs);
 792         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 793         spin_lock_init(&conn->ibc_lock);
 794
 795         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 796                          sizeof(*conn->ibc_connvars));
 797         if (conn->ibc_connvars == NULL) {
 798                 CERROR("Can't allocate in-progress connection state\n");
 799                 goto failed_2;
 800         }
 801
 802         write_lock_irqsave(glock, flags);
 803         if (dev->ibd_failover) {
 804                 write_unlock_irqrestore(glock, flags);
 805                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 806                 goto failed_2;
 807         }
 808
 809         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 810                 /* wakeup failover thread and teardown connection */
 811                 if (kiblnd_dev_can_failover(dev)) {
 812                         list_add_tail(&dev->ibd_fail_list,
 813                                       &kiblnd_data.kib_failed_devs);
 814                         wake_up(&kiblnd_data.kib_failover_waitq);
 815                 }
 816
 817                 write_unlock_irqrestore(glock, flags);
 818                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 819                        cmid->device->name, dev->ibd_ifname);
 820                 goto failed_2;
 821         }
 822
 823         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 824         conn->ibc_hdev = dev->ibd_hdev;
 825
 826         kiblnd_setup_mtu_locked(cmid);
 827
 828         write_unlock_irqrestore(glock, flags);
 829
 830 #ifdef HAVE_IB_CQ_INIT_ATTR
 831         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 832         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 833         cq = ib_create_cq(cmid->device,
 834                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 835                           &cq_attr);
 836 #else
 837         cq = ib_create_cq(cmid->device,
 838                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 839                           IBLND_CQ_ENTRIES(conn),
 840                           kiblnd_get_completion_vector(conn, cpt));
 841 #endif
 842         if (IS_ERR(cq)) {
 843                 /*
 844                  * on MLX-5 (possibly MLX-4 as well) this error could be
 845                  * hit if the concurrent_sends and/or peer_tx_credits is set
 846                  * too high. Or due to an MLX-5 bug which tries to
 847                  * allocate 256kb via kmalloc for WR cookie array
 848                  */
 849                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 850                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 851                 goto failed_2;
 852         }
 853
 854         conn->ibc_cq = cq;
 855
 856         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 857         if (rc != 0) {
 858                 CERROR("Can't request completion notification: %d\n", rc);
 859                 goto failed_2;
 860         }
 861
 862         init_qp_attr.event_handler = kiblnd_qp_event;
 863         init_qp_attr.qp_context = conn;
 864         init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 865         init_qp_attr.cap.max_recv_sge = 1;
 866         init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 867         init_qp_attr.qp_type = IB_QPT_RC;
 868         init_qp_attr.send_cq = cq;
 869         init_qp_attr.recv_cq = cq;
 870
 871         if (peer_ni->ibp_queue_depth_mod &&
 872             peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
 873                 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
 874                 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
 875                        peer_ni->ibp_queue_depth_mod,
 876                        peer_ni->ibp_queue_depth);
 877         }
 878
 879         do {
 880                 /* kiblnd_send_wrs() can change the connection's queue depth if
 881                  * the maximum work requests for the device is maxed out
 882                  */
 883                 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
 884                 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
 885                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
 886                                     &init_qp_attr);
 887                 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
 888                         break;
 889                 conn->ibc_queue_depth--;
 890         } while (rc);
 891
 892         if (rc) {
 893                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 894                        "send_sge: %d, recv_sge: %d\n",
 895                        rc, init_qp_attr.cap.max_send_wr,
 896                        init_qp_attr.cap.max_recv_wr,
 897                        init_qp_attr.cap.max_send_sge,
 898                        init_qp_attr.cap.max_recv_sge);
 899                 goto failed_2;
 900         }
 901
 902         conn->ibc_sched = sched;
 903
 904         if (!peer_ni->ibp_queue_depth_mod &&
 905             conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
 906                 CWARN("peer %s - queue depth reduced from %u to %u"
 907                       "  to allow for qp creation\n",
 908                       libcfs_nid2str(peer_ni->ibp_nid),
 909                       peer_ni->ibp_queue_depth,
 910                       conn->ibc_queue_depth);
 911                 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
 912         }
 913
 914         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 915                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 916         if (conn->ibc_rxs == NULL) {
 917                 CERROR("Cannot allocate RX buffers\n");
 918                 goto failed_2;
 919         }
 920
 921         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 922                                 IBLND_RX_MSG_PAGES(conn));
 923         if (rc != 0)
 924                 goto failed_2;
 925
 926         kiblnd_map_rx_descs(conn);
 927
 928         /* 1 ref for caller and each rxmsg */
 929         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 930         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 931
 932         /* post receives */
 933         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 934                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 935                 if (rc != 0) {
 936                         CERROR("Can't post rxmsg: %d\n", rc);
 937
 938                         /* Make posted receives complete */
 939                         kiblnd_abort_receives(conn);
 940
 941                         /* correct # of posted buffers
 942                          * NB locking needed now I'm racing with completion */
 943                         spin_lock_irqsave(&sched->ibs_lock, flags);
 944                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 945                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 946
 947                         /* cmid will be destroyed by CM(ofed) after cm_callback
 948                          * returned, so we can't refer it anymore
 949                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 950                         rdma_destroy_qp(conn->ibc_cmid);
 951                         conn->ibc_cmid = NULL;
 952
 953                         /* Drop my own and unused rxbuffer refcounts */
 954                         while (i++ <= IBLND_RX_MSGS(conn))
 955                                 kiblnd_conn_decref(conn);
 956
 957                         return NULL;
 958                 }
 959         }
 960
 961         /* Init successful! */
 962         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 963                  state == IBLND_CONN_PASSIVE_WAIT);
 964         conn->ibc_state = state;
 965
 966         /* 1 more conn */
 967         atomic_inc(&net->ibn_nconns);
 968         return conn;
 969
 970  failed_2:
 971         kiblnd_destroy_conn(conn);
 972         LIBCFS_FREE(conn, sizeof(*conn));
 973  failed_0:
 974         return NULL;
 975 }
 976
 977 void
 978 kiblnd_destroy_conn(struct kib_conn *conn)
 979 {
 980         struct rdma_cm_id *cmid = conn->ibc_cmid;
 981         struct kib_peer_ni *peer_ni = conn->ibc_peer;
 982
 983         LASSERT (!in_interrupt());
 984         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 985         LASSERT(list_empty(&conn->ibc_early_rxs));
 986         LASSERT(list_empty(&conn->ibc_tx_noops));
 987         LASSERT(list_empty(&conn->ibc_tx_queue));
 988         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 989         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 990         LASSERT(list_empty(&conn->ibc_active_txs));
 991         LASSERT (conn->ibc_noops_posted == 0);
 992         LASSERT (conn->ibc_nsends_posted == 0);
 993
 994         switch (conn->ibc_state) {
 995         default:
 996                 /* conn must be completely disengaged from the network */
 997                 LBUG();
 998
 999         case IBLND_CONN_DISCONNECTED:
1000                 /* connvars should have been freed already */
1001                 LASSERT (conn->ibc_connvars == NULL);
1002                 break;
1003
1004         case IBLND_CONN_INIT:
1005                 break;
1006         }
1007
1008         /* conn->ibc_cmid might be destroyed by CM already */
1009         if (cmid != NULL && cmid->qp != NULL)
1010                 rdma_destroy_qp(cmid);
1011
1012         if (conn->ibc_cq)
1013                 ib_destroy_cq(conn->ibc_cq);
1014
1015         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1016                            LNET_MSG_STATUS_OK);
1017
1018         if (conn->ibc_rx_pages != NULL)
1019                 kiblnd_unmap_rx_descs(conn);
1020
1021         if (conn->ibc_rxs != NULL)
1022                 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1023
1024         if (conn->ibc_connvars != NULL)
1025                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1026
1027         if (conn->ibc_hdev != NULL)
1028                 kiblnd_hdev_decref(conn->ibc_hdev);
1029
1030         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1031         if (conn->ibc_state != IBLND_CONN_INIT) {
1032                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1033
1034                 kiblnd_peer_decref(peer_ni);
1035                 rdma_destroy_id(cmid);
1036                 atomic_dec(&net->ibn_nconns);
1037         }
1038 }
1039
1040 int
1041 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1042 {
1043         struct kib_conn *conn;
1044         struct kib_conn *cnxt;
1045         int count = 0;
1046
1047         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1048                                  ibc_list) {
1049                 CDEBUG(D_NET, "Closing conn -> %s, "
1050                               "version: %x, reason: %d\n",
1051                        libcfs_nid2str(peer_ni->ibp_nid),
1052                        conn->ibc_version, why);
1053
1054                 kiblnd_close_conn_locked(conn, why);
1055                 count++;
1056         }
1057
1058         return count;
1059 }
1060
1061 int
1062 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1063                                 int version, __u64 incarnation)
1064 {
1065         struct kib_conn *conn;
1066         struct kib_conn *cnxt;
1067         int count = 0;
1068
1069         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1070                                  ibc_list) {
1071                 if (conn->ibc_version     == version &&
1072                     conn->ibc_incarnation == incarnation)
1073                         continue;
1074
1075                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1076                               "incarnation:%#llx(%x, %#llx)\n",
1077                        libcfs_nid2str(peer_ni->ibp_nid),
1078                        conn->ibc_version, conn->ibc_incarnation,
1079                        version, incarnation);
1080
1081                 kiblnd_close_conn_locked(conn, -ESTALE);
1082                 count++;
1083         }
1084
1085         return count;
1086 }
1087
1088 static int
1089 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1090 {
1091         struct kib_peer_ni *peer_ni;
1092         struct hlist_node *pnxt;
1093         int lo;
1094         int hi;
1095         int i;
1096         unsigned long flags;
1097         int count = 0;
1098
1099         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1100
1101         if (nid != LNET_NID_ANY) {
1102                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1103                 hi = lo;
1104         } else {
1105                 lo = 0;
1106                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1107         }
1108
1109         for (i = lo; i <= hi; i++) {
1110                 hlist_for_each_entry_safe(peer_ni, pnxt,
1111                                           &kiblnd_data.kib_peers[i], ibp_list) {
1112                         LASSERT(!kiblnd_peer_idle(peer_ni));
1113
1114                         if (peer_ni->ibp_ni != ni)
1115                                 continue;
1116
1117                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1118                                 continue;
1119
1120                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1121                 }
1122         }
1123
1124         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1125
1126         /* wildcards always succeed */
1127         if (nid == LNET_NID_ANY)
1128                 return 0;
1129
1130         return (count == 0) ? -ENOENT : 0;
1131 }
1132
1133 static int
1134 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1135 {
1136         struct libcfs_ioctl_data *data = arg;
1137         int                       rc = -EINVAL;
1138
1139         switch(cmd) {
1140         case IOC_LIBCFS_GET_PEER: {
1141                 lnet_nid_t   nid = 0;
1142                 int          count = 0;
1143
1144                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1145                                           &nid, &count);
1146                 data->ioc_nid    = nid;
1147                 data->ioc_count  = count;
1148                 break;
1149         }
1150
1151         case IOC_LIBCFS_DEL_PEER: {
1152                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1153                 break;
1154         }
1155         case IOC_LIBCFS_GET_CONN: {
1156                 struct kib_conn *conn;
1157
1158                 rc = 0;
1159                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1160                 if (conn == NULL) {
1161                         rc = -ENOENT;
1162                         break;
1163                 }
1164
1165                 LASSERT(conn->ibc_cmid != NULL);
1166                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1167                 if (conn->ibc_cmid->route.path_rec == NULL)
1168                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1169                 else
1170                         data->ioc_u32[0] =
1171                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1172                 kiblnd_conn_decref(conn);
1173                 break;
1174         }
1175         case IOC_LIBCFS_CLOSE_CONNECTION: {
1176                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1177                 break;
1178         }
1179
1180         default:
1181                 break;
1182         }
1183
1184         return rc;
1185 }
1186
1187 static void
1188 kiblnd_free_pages(struct kib_pages *p)
1189 {
1190         int     npages = p->ibp_npages;
1191         int     i;
1192
1193         for (i = 0; i < npages; i++) {
1194                 if (p->ibp_pages[i] != NULL)
1195                         __free_page(p->ibp_pages[i]);
1196         }
1197
1198         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1199 }
1200
1201 int
1202 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1203 {
1204         struct kib_pages *p;
1205         int i;
1206
1207         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1208                          offsetof(struct kib_pages, ibp_pages[npages]));
1209         if (p == NULL) {
1210                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1211                 return -ENOMEM;
1212         }
1213
1214         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1215         p->ibp_npages = npages;
1216
1217         for (i = 0; i < npages; i++) {
1218                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1219                                                      GFP_NOFS);
1220                 if (p->ibp_pages[i] == NULL) {
1221                         CERROR("Can't allocate page %d of %d\n", i, npages);
1222                         kiblnd_free_pages(p);
1223                         return -ENOMEM;
1224                 }
1225         }
1226
1227         *pp = p;
1228         return 0;
1229 }
1230
1231 void
1232 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1233 {
1234         struct kib_rx *rx;
1235         int       i;
1236
1237         LASSERT (conn->ibc_rxs != NULL);
1238         LASSERT (conn->ibc_hdev != NULL);
1239
1240         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1241                 rx = &conn->ibc_rxs[i];
1242
1243                 LASSERT(rx->rx_nob >= 0); /* not posted */
1244
1245                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1246                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1247                                                           rx->rx_msgaddr),
1248                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1249         }
1250
1251         kiblnd_free_pages(conn->ibc_rx_pages);
1252
1253         conn->ibc_rx_pages = NULL;
1254 }
1255
1256 void
1257 kiblnd_map_rx_descs(struct kib_conn *conn)
1258 {
1259         struct kib_rx *rx;
1260         struct page    *pg;
1261         int             pg_off;
1262         int             ipg;
1263         int             i;
1264
1265         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1266                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1267                 rx = &conn->ibc_rxs[i];
1268
1269                 rx->rx_conn = conn;
1270                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1271
1272                 rx->rx_msgaddr =
1273                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1274                                               rx->rx_msg, IBLND_MSG_SIZE,
1275                                               DMA_FROM_DEVICE);
1276                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1277                                                   rx->rx_msgaddr));
1278                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1279
1280                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1281                        i, rx->rx_msg, rx->rx_msgaddr,
1282                        (__u64)(page_to_phys(pg) + pg_off));
1283
1284                 pg_off += IBLND_MSG_SIZE;
1285                 LASSERT(pg_off <= PAGE_SIZE);
1286
1287                 if (pg_off == PAGE_SIZE) {
1288                         pg_off = 0;
1289                         ipg++;
1290                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1291                 }
1292         }
1293 }
1294
1295 static void
1296 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1297 {
1298         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1299         struct kib_tx *tx;
1300         int i;
1301
1302         LASSERT (tpo->tpo_pool.po_allocated == 0);
1303
1304         if (hdev == NULL)
1305                 return;
1306
1307         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1308                 tx = &tpo->tpo_tx_descs[i];
1309                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1310                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1311                                                           tx->tx_msgaddr),
1312                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1313         }
1314
1315         kiblnd_hdev_decref(hdev);
1316         tpo->tpo_hdev = NULL;
1317 }
1318
1319 static struct kib_hca_dev *
1320 kiblnd_current_hdev(struct kib_dev *dev)
1321 {
1322         struct kib_hca_dev *hdev;
1323         unsigned long  flags;
1324         int            i = 0;
1325
1326         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1327         while (dev->ibd_failover) {
1328                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1329                 if (i++ % 50 == 0)
1330                         CDEBUG(D_NET, "%s: Wait for failover\n",
1331                                dev->ibd_ifname);
1332                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1333
1334                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1335         }
1336
1337         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1338         hdev = dev->ibd_hdev;
1339
1340         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1341
1342         return hdev;
1343 }
1344
1345 static void
1346 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1347 {
1348         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1349         struct kib_pool *pool = &tpo->tpo_pool;
1350         struct kib_net      *net   = pool->po_owner->ps_net;
1351         struct kib_dev *dev;
1352         struct page *page;
1353         struct kib_tx *tx;
1354         int             page_offset;
1355         int             ipage;
1356         int             i;
1357
1358         LASSERT (net != NULL);
1359
1360         dev = net->ibn_dev;
1361
1362         /* pre-mapped messages are not bigger than 1 page */
1363         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1364
1365         /* No fancy arithmetic when we do the buffer calculations */
1366         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1367
1368         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1369
1370         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1371                 page = txpgs->ibp_pages[ipage];
1372                 tx = &tpo->tpo_tx_descs[i];
1373
1374                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1375                                                 page_offset);
1376
1377                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1378                                                        tx->tx_msg,
1379                                                        IBLND_MSG_SIZE,
1380                                                        DMA_TO_DEVICE);
1381                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1382                                                   tx->tx_msgaddr));
1383                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1384
1385                 list_add(&tx->tx_list, &pool->po_free_list);
1386
1387                 page_offset += IBLND_MSG_SIZE;
1388                 LASSERT(page_offset <= PAGE_SIZE);
1389
1390                 if (page_offset == PAGE_SIZE) {
1391                         page_offset = 0;
1392                         ipage++;
1393                         LASSERT(ipage <= txpgs->ibp_npages);
1394                 }
1395         }
1396 }
1397
1398 static void
1399 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1400 {
1401         LASSERT(fpo->fpo_map_count == 0);
1402
1403 #ifdef HAVE_FMR_POOL_API
1404         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1405                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1406         } else
1407 #endif /* HAVE_FMR_POOL_API */
1408         {
1409                 struct kib_fast_reg_descriptor *frd, *tmp;
1410                 int i = 0;
1411
1412                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1413                                          frd_list) {
1414                         list_del(&frd->frd_list);
1415 #ifndef HAVE_IB_MAP_MR_SG
1416                         ib_free_fast_reg_page_list(frd->frd_frpl);
1417 #endif
1418                         ib_dereg_mr(frd->frd_mr);
1419                         LIBCFS_FREE(frd, sizeof(*frd));
1420                         i++;
1421                 }
1422                 if (i < fpo->fast_reg.fpo_pool_size)
1423                         CERROR("FastReg pool still has %d regions registered\n",
1424                                 fpo->fast_reg.fpo_pool_size - i);
1425         }
1426
1427         if (fpo->fpo_hdev)
1428                 kiblnd_hdev_decref(fpo->fpo_hdev);
1429
1430         LIBCFS_FREE(fpo, sizeof(*fpo));
1431 }
1432
1433 static void
1434 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1435 {
1436         struct kib_fmr_pool *fpo, *tmp;
1437
1438         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1439                 list_del(&fpo->fpo_list);
1440                 kiblnd_destroy_fmr_pool(fpo);
1441         }
1442 }
1443
1444 static int
1445 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1446                      int ncpts)
1447 {
1448         int size = tunables->lnd_fmr_pool_size / ncpts;
1449
1450         return max(IBLND_FMR_POOL, size);
1451 }
1452
1453 static int
1454 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1455                          int ncpts)
1456 {
1457         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1458
1459         return max(IBLND_FMR_POOL_FLUSH, size);
1460 }
1461
1462 #ifdef HAVE_FMR_POOL_API
1463 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1464                                  struct kib_fmr_pool *fpo)
1465 {
1466         struct ib_fmr_pool_param param = {
1467                 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1468                 .page_shift        = PAGE_SHIFT,
1469                 .access            = (IB_ACCESS_LOCAL_WRITE |
1470                                       IB_ACCESS_REMOTE_WRITE),
1471                 .pool_size         = fps->fps_pool_size,
1472                 .dirty_watermark   = fps->fps_flush_trigger,
1473                 .flush_function    = NULL,
1474                 .flush_arg         = NULL,
1475                 .cache             = !!fps->fps_cache };
1476         int rc = 0;
1477
1478         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1479                                                    &param);
1480         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1481                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1482                 if (rc != -ENOSYS)
1483                         CERROR("Failed to create FMR pool: %d\n", rc);
1484                 else
1485                         CERROR("FMRs are not supported\n");
1486         }
1487         fpo->fpo_is_fmr = true;
1488
1489         return rc;
1490 }
1491 #endif /* HAVE_FMR_POOL_API */
1492
1493 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1494                                   struct kib_fmr_pool *fpo,
1495                                   enum kib_dev_caps dev_caps)
1496 {
1497         struct kib_fast_reg_descriptor *frd, *tmp;
1498         int i, rc;
1499
1500 #ifdef HAVE_FMR_POOL_API
1501         fpo->fpo_is_fmr = false;
1502 #endif
1503
1504         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1505         fpo->fast_reg.fpo_pool_size = 0;
1506         for (i = 0; i < fps->fps_pool_size; i++) {
1507                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1508                                  sizeof(*frd));
1509                 if (!frd) {
1510                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1511                         rc = -ENOMEM;
1512                         goto out;
1513                 }
1514                 frd->frd_mr = NULL;
1515
1516 #ifndef HAVE_IB_MAP_MR_SG
1517                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1518                                                             IBLND_MAX_RDMA_FRAGS);
1519                 if (IS_ERR(frd->frd_frpl)) {
1520                         rc = PTR_ERR(frd->frd_frpl);
1521                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1522                                 rc);
1523                         frd->frd_frpl = NULL;
1524                         goto out_middle;
1525                 }
1526 #endif
1527
1528 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1529                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1530                                                    IBLND_MAX_RDMA_FRAGS);
1531 #else
1532                 /*
1533                  * it is expected to get here if this is an MLX-5 card.
1534                  * MLX-4 cards will always use FMR and MLX-5 cards will
1535                  * always use fast_reg. It turns out that some MLX-5 cards
1536                  * (possibly due to older FW versions) do not natively support
1537                  * gaps. So we will need to track them here.
1538                  */
1539                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1540 #ifdef IB_MR_TYPE_SG_GAPS
1541                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1542                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1543                                                 IB_MR_TYPE_SG_GAPS :
1544                                                 IB_MR_TYPE_MEM_REG,
1545 #else
1546                                                 IB_MR_TYPE_MEM_REG,
1547 #endif
1548                                           IBLND_MAX_RDMA_FRAGS);
1549                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1550                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1551                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1552 #endif
1553                 if (IS_ERR(frd->frd_mr)) {
1554                         rc = PTR_ERR(frd->frd_mr);
1555                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1556                         frd->frd_mr = NULL;
1557                         goto out_middle;
1558                 }
1559
1560                 /* There appears to be a bug in MLX5 code where you must
1561                  * invalidate the rkey of a new FastReg pool before first
1562                  * using it. Thus, I am marking the FRD invalid here. */
1563                 frd->frd_valid = false;
1564
1565                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1566                 fpo->fast_reg.fpo_pool_size++;
1567         }
1568
1569         return 0;
1570
1571 out_middle:
1572         if (frd->frd_mr)
1573                 ib_dereg_mr(frd->frd_mr);
1574 #ifndef HAVE_IB_MAP_MR_SG
1575         if (frd->frd_frpl)
1576                 ib_free_fast_reg_page_list(frd->frd_frpl);
1577 #endif
1578         LIBCFS_FREE(frd, sizeof(*frd));
1579
1580 out:
1581         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1582                                  frd_list) {
1583                 list_del(&frd->frd_list);
1584 #ifndef HAVE_IB_MAP_MR_SG
1585                 ib_free_fast_reg_page_list(frd->frd_frpl);
1586 #endif
1587                 ib_dereg_mr(frd->frd_mr);
1588                 LIBCFS_FREE(frd, sizeof(*frd));
1589         }
1590
1591         return rc;
1592 }
1593
1594 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1595                                   struct kib_fmr_pool **pp_fpo)
1596 {
1597         struct kib_dev *dev = fps->fps_net->ibn_dev;
1598         struct kib_fmr_pool *fpo;
1599         int rc;
1600
1601         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1602         if (!fpo) {
1603                 return -ENOMEM;
1604         }
1605         memset(fpo, 0, sizeof(*fpo));
1606
1607         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1608
1609 #ifdef HAVE_FMR_POOL_API
1610         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1611                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1612         else
1613 #endif /* HAVE_FMR_POOL_API */
1614                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1615         if (rc)
1616                 goto out_fpo;
1617
1618         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1619         fpo->fpo_owner = fps;
1620         *pp_fpo = fpo;
1621
1622         return 0;
1623
1624 out_fpo:
1625         kiblnd_hdev_decref(fpo->fpo_hdev);
1626         LIBCFS_FREE(fpo, sizeof(*fpo));
1627         return rc;
1628 }
1629
1630 static void
1631 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1632 {
1633         struct kib_fmr_pool *fpo;
1634
1635         if (fps->fps_net == NULL) /* intialized? */
1636                 return;
1637
1638         spin_lock(&fps->fps_lock);
1639
1640         while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1641                                                struct kib_fmr_pool,
1642                                                fpo_list)) != NULL) {
1643                 fpo->fpo_failed = 1;
1644                 if (fpo->fpo_map_count == 0)
1645                         list_move(&fpo->fpo_list, zombies);
1646                 else
1647                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1648         }
1649
1650         spin_unlock(&fps->fps_lock);
1651 }
1652
1653 static void
1654 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1655 {
1656         if (fps->fps_net != NULL) { /* initialized? */
1657                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1658                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1659         }
1660 }
1661
1662 static int
1663 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1664                         struct kib_net *net,
1665                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1666 {
1667         struct kib_fmr_pool *fpo;
1668         int rc;
1669
1670         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1671
1672         fps->fps_net = net;
1673         fps->fps_cpt = cpt;
1674
1675         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1676         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1677         fps->fps_cache = tunables->lnd_fmr_cache;
1678
1679         spin_lock_init(&fps->fps_lock);
1680         INIT_LIST_HEAD(&fps->fps_pool_list);
1681         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1682
1683         rc = kiblnd_create_fmr_pool(fps, &fpo);
1684         if (rc == 0)
1685                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1686
1687         return rc;
1688 }
1689
1690 static int
1691 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1692 {
1693         if (fpo->fpo_map_count != 0) /* still in use */
1694                 return 0;
1695         if (fpo->fpo_failed)
1696                 return 1;
1697         return now >= fpo->fpo_deadline;
1698 }
1699
1700 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1701 static int
1702 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1703 {
1704         struct kib_hca_dev *hdev;
1705         __u64           *pages = tx->tx_pages;
1706         int             npages;
1707         int             size;
1708         int             i;
1709
1710         hdev = tx->tx_pool->tpo_hdev;
1711
1712         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1713                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1714                         size += hdev->ibh_page_size) {
1715                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1716                                            hdev->ibh_page_mask) + size;
1717                 }
1718         }
1719
1720         return npages;
1721 }
1722 #endif
1723
1724 void
1725 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1726 {
1727         LIST_HEAD(zombies);
1728         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1729         struct kib_fmr_poolset *fps;
1730         time64_t now = ktime_get_seconds();
1731         struct kib_fmr_pool *tmp;
1732
1733         if (!fpo)
1734                 return;
1735
1736         fps = fpo->fpo_owner;
1737
1738 #ifdef HAVE_FMR_POOL_API
1739         if (fpo->fpo_is_fmr) {
1740                 if (fmr->fmr_pfmr) {
1741                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1742                         fmr->fmr_pfmr = NULL;
1743                 }
1744
1745                 if (status) {
1746                         int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1747                         LASSERT(!rc);
1748                 }
1749         } else
1750 #endif /* HAVE_FMR_POOL_API */
1751         {
1752                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1753
1754                 if (frd) {
1755                         frd->frd_valid = false;
1756                         frd->frd_posted = false;
1757                         fmr->fmr_frd = NULL;
1758                         spin_lock(&fps->fps_lock);
1759                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1760                         spin_unlock(&fps->fps_lock);
1761                 }
1762         }
1763         fmr->fmr_pool = NULL;
1764
1765         spin_lock(&fps->fps_lock);
1766         fpo->fpo_map_count--;   /* decref the pool */
1767
1768         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1769                 /* the first pool is persistent */
1770                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1771                         continue;
1772
1773                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1774                         list_move(&fpo->fpo_list, &zombies);
1775                         fps->fps_version++;
1776                 }
1777         }
1778         spin_unlock(&fps->fps_lock);
1779
1780         if (!list_empty(&zombies))
1781                 kiblnd_destroy_fmr_pool_list(&zombies);
1782 }
1783
1784 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1785                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1786                         struct kib_fmr *fmr)
1787 {
1788         struct kib_fmr_pool *fpo;
1789         __u64 version;
1790         bool is_rx = (rd != tx->tx_rd);
1791 #ifdef HAVE_FMR_POOL_API
1792         __u64 *pages = tx->tx_pages;
1793         bool tx_pages_mapped = false;
1794         int npages = 0;
1795 #endif
1796         int rc;
1797
1798 again:
1799         spin_lock(&fps->fps_lock);
1800         version = fps->fps_version;
1801         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1802                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1803                 fpo->fpo_map_count++;
1804
1805 #ifdef HAVE_FMR_POOL_API
1806                 fmr->fmr_pfmr = NULL;
1807                 if (fpo->fpo_is_fmr) {
1808                         struct ib_pool_fmr *pfmr;
1809
1810                         spin_unlock(&fps->fps_lock);
1811
1812                         if (!tx_pages_mapped) {
1813                                 npages = kiblnd_map_tx_pages(tx, rd);
1814                                 tx_pages_mapped = true;
1815                         }
1816
1817                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1818                                                 pages, npages, iov);
1819                         if (likely(!IS_ERR(pfmr))) {
1820                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1821                                         : pfmr->fmr->lkey;
1822                                 fmr->fmr_frd  = NULL;
1823                                 fmr->fmr_pfmr = pfmr;
1824                                 fmr->fmr_pool = fpo;
1825                                 return 0;
1826                         }
1827                         rc = PTR_ERR(pfmr);
1828                 } else
1829 #endif /* HAVE_FMR_POOL_API */
1830                 {
1831                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1832                                 struct kib_fast_reg_descriptor *frd;
1833 #ifdef HAVE_IB_MAP_MR_SG
1834                                 struct ib_reg_wr *wr;
1835                                 int n;
1836 #else
1837                                 struct ib_rdma_wr *wr;
1838                                 struct ib_fast_reg_page_list *frpl;
1839 #endif
1840                                 struct ib_mr *mr;
1841
1842                                 frd = list_first_entry(
1843                                         &fpo->fast_reg.fpo_pool_list,
1844                                         struct kib_fast_reg_descriptor,
1845                                         frd_list);
1846                                 list_del(&frd->frd_list);
1847                                 spin_unlock(&fps->fps_lock);
1848
1849 #ifndef HAVE_IB_MAP_MR_SG
1850                                 frpl = frd->frd_frpl;
1851 #endif
1852                                 mr   = frd->frd_mr;
1853
1854                                 if (!frd->frd_valid) {
1855                                         struct ib_rdma_wr *inv_wr;
1856                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1857
1858                                         inv_wr = &frd->frd_inv_wr;
1859                                         memset(inv_wr, 0, sizeof(*inv_wr));
1860
1861                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1862                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1863                                         inv_wr->wr.ex.invalidate_rkey = key;
1864
1865                                         /* Bump the key */
1866                                         key = ib_inc_rkey(key);
1867                                         ib_update_fast_reg_key(mr, key);
1868                                 }
1869
1870 #ifdef HAVE_IB_MAP_MR_SG
1871 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1872                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1873                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
1874 #else
1875                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1876                                                  rd->rd_nfrags, PAGE_SIZE);
1877 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1878                                 if (unlikely(n != rd->rd_nfrags)) {
1879                                         CERROR("Failed to map mr %d/%d elements\n",
1880                                                n, rd->rd_nfrags);
1881                                         return n < 0 ? n : -EINVAL;
1882                                 }
1883
1884                                 wr = &frd->frd_fastreg_wr;
1885                                 memset(wr, 0, sizeof(*wr));
1886
1887                                 wr->wr.opcode = IB_WR_REG_MR;
1888                                 wr->wr.wr_id  = IBLND_WID_MR;
1889                                 wr->wr.num_sge = 0;
1890                                 wr->wr.send_flags = 0;
1891                                 wr->mr = mr;
1892                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1893                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1894                                               IB_ACCESS_REMOTE_WRITE);
1895 #else /* HAVE_IB_MAP_MR_SG */
1896                                 if (!tx_pages_mapped) {
1897                                         npages = kiblnd_map_tx_pages(tx, rd);
1898                                         tx_pages_mapped = true;
1899                                 }
1900
1901                                 LASSERT(npages <= frpl->max_page_list_len);
1902                                 memcpy(frpl->page_list, pages,
1903                                        sizeof(*pages) * npages);
1904
1905                                 /* Prepare FastReg WR */
1906                                 wr = &frd->frd_fastreg_wr;
1907                                 memset(wr, 0, sizeof(*wr));
1908
1909                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1910                                 wr->wr.wr_id  = IBLND_WID_MR;
1911
1912                                 wr->wr.wr.fast_reg.iova_start = iov;
1913                                 wr->wr.wr.fast_reg.page_list  = frpl;
1914                                 wr->wr.wr.fast_reg.page_list_len = npages;
1915                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1916                                 wr->wr.wr.fast_reg.length = nob;
1917                                 wr->wr.wr.fast_reg.rkey =
1918                                         is_rx ? mr->rkey : mr->lkey;
1919                                 wr->wr.wr.fast_reg.access_flags =
1920                                         (IB_ACCESS_LOCAL_WRITE |
1921                                          IB_ACCESS_REMOTE_WRITE);
1922 #endif /* HAVE_IB_MAP_MR_SG */
1923
1924                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1925                                 fmr->fmr_frd  = frd;
1926                                 fmr->fmr_pool = fpo;
1927                                 frd->frd_posted = false;
1928                                 return 0;
1929                         }
1930                         spin_unlock(&fps->fps_lock);
1931                         rc = -EAGAIN;
1932                 }
1933
1934                 spin_lock(&fps->fps_lock);
1935                 fpo->fpo_map_count--;
1936                 if (rc != -EAGAIN) {
1937                         spin_unlock(&fps->fps_lock);
1938                         return rc;
1939                 }
1940
1941                 /* EAGAIN and ... */
1942                 if (version != fps->fps_version) {
1943                         spin_unlock(&fps->fps_lock);
1944                         goto again;
1945                 }
1946         }
1947
1948         if (fps->fps_increasing) {
1949                 spin_unlock(&fps->fps_lock);
1950                 CDEBUG(D_NET, "Another thread is allocating new "
1951                        "FMR pool, waiting for her to complete\n");
1952                 wait_var_event(fps, !fps->fps_increasing);
1953                 goto again;
1954
1955         }
1956
1957         if (ktime_get_seconds() < fps->fps_next_retry) {
1958                 /* someone failed recently */
1959                 spin_unlock(&fps->fps_lock);
1960                 return -EAGAIN;
1961         }
1962
1963         fps->fps_increasing = 1;
1964         spin_unlock(&fps->fps_lock);
1965
1966         CDEBUG(D_NET, "Allocate new FMR pool\n");
1967         rc = kiblnd_create_fmr_pool(fps, &fpo);
1968         spin_lock(&fps->fps_lock);
1969         fps->fps_increasing = 0;
1970         wake_up_var(fps);
1971         if (rc == 0) {
1972                 fps->fps_version++;
1973                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1974         } else {
1975                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1976         }
1977         spin_unlock(&fps->fps_lock);
1978
1979         goto again;
1980 }
1981
1982 static void
1983 kiblnd_fini_pool(struct kib_pool *pool)
1984 {
1985         LASSERT(list_empty(&pool->po_free_list));
1986         LASSERT(pool->po_allocated == 0);
1987
1988         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1989 }
1990
1991 static void
1992 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1993 {
1994         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1995
1996         memset(pool, 0, sizeof(struct kib_pool));
1997         INIT_LIST_HEAD(&pool->po_free_list);
1998         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1999         pool->po_owner = ps;
2000         pool->po_size = size;
2001 }
2002
2003 static void
2004 kiblnd_destroy_pool_list(struct list_head *head)
2005 {
2006         struct kib_pool *pool;
2007
2008         while ((pool = list_first_entry_or_null(head,
2009                                                 struct kib_pool,
2010                                                 po_list)) != NULL) {
2011                 list_del(&pool->po_list);
2012
2013                 LASSERT(pool->po_owner != NULL);
2014                 pool->po_owner->ps_pool_destroy(pool);
2015         }
2016 }
2017
2018 static void
2019 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2020 {
2021         struct kib_pool *po;
2022
2023         if (ps->ps_net == NULL) /* intialized? */
2024                 return;
2025
2026         spin_lock(&ps->ps_lock);
2027         while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2028                                               struct kib_pool,
2029                                               po_list)) != NULL) {
2030                 po->po_failed = 1;
2031                 if (po->po_allocated == 0)
2032                         list_move(&po->po_list, zombies);
2033                 else
2034                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2035         }
2036         spin_unlock(&ps->ps_lock);
2037 }
2038
2039 static void
2040 kiblnd_fini_poolset(struct kib_poolset *ps)
2041 {
2042         if (ps->ps_net != NULL) { /* initialized? */
2043                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2044                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2045         }
2046 }
2047
2048 static int
2049 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2050                     struct kib_net *net, char *name, int size,
2051                     kib_ps_pool_create_t po_create,
2052                     kib_ps_pool_destroy_t po_destroy,
2053                     kib_ps_node_init_t nd_init,
2054                     kib_ps_node_fini_t nd_fini)
2055 {
2056         struct kib_pool *pool;
2057         int rc;
2058
2059         memset(ps, 0, sizeof(struct kib_poolset));
2060
2061         ps->ps_cpt          = cpt;
2062         ps->ps_net          = net;
2063         ps->ps_pool_create  = po_create;
2064         ps->ps_pool_destroy = po_destroy;
2065         ps->ps_node_init    = nd_init;
2066         ps->ps_node_fini    = nd_fini;
2067         ps->ps_pool_size    = size;
2068         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2069             >= sizeof(ps->ps_name))
2070                 return -E2BIG;
2071         spin_lock_init(&ps->ps_lock);
2072         INIT_LIST_HEAD(&ps->ps_pool_list);
2073         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2074
2075         rc = ps->ps_pool_create(ps, size, &pool);
2076         if (rc == 0)
2077                 list_add(&pool->po_list, &ps->ps_pool_list);
2078         else
2079                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2080
2081         return rc;
2082 }
2083
2084 static int
2085 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2086 {
2087         if (pool->po_allocated != 0) /* still in use */
2088                 return 0;
2089         if (pool->po_failed)
2090                 return 1;
2091         return now >= pool->po_deadline;
2092 }
2093
2094 void
2095 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2096 {
2097         LIST_HEAD(zombies);
2098         struct kib_poolset *ps = pool->po_owner;
2099         struct kib_pool *tmp;
2100         time64_t now = ktime_get_seconds();
2101
2102         spin_lock(&ps->ps_lock);
2103
2104         if (ps->ps_node_fini != NULL)
2105                 ps->ps_node_fini(pool, node);
2106
2107         LASSERT(pool->po_allocated > 0);
2108         list_add(node, &pool->po_free_list);
2109         pool->po_allocated--;
2110
2111         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2112                 /* the first pool is persistent */
2113                 if (ps->ps_pool_list.next == &pool->po_list)
2114                         continue;
2115
2116                 if (kiblnd_pool_is_idle(pool, now))
2117                         list_move(&pool->po_list, &zombies);
2118         }
2119         spin_unlock(&ps->ps_lock);
2120
2121         if (!list_empty(&zombies))
2122                 kiblnd_destroy_pool_list(&zombies);
2123 }
2124
2125 struct list_head *
2126 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2127 {
2128         struct list_head        *node;
2129         struct kib_pool *pool;
2130         int                     rc;
2131         unsigned int            interval = 1;
2132         ktime_t time_before;
2133         unsigned int trips = 0;
2134
2135 again:
2136         spin_lock(&ps->ps_lock);
2137         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2138                 if (list_empty(&pool->po_free_list))
2139                         continue;
2140
2141                 pool->po_allocated++;
2142                 pool->po_deadline = ktime_get_seconds() +
2143                                     IBLND_POOL_DEADLINE;
2144                 node = pool->po_free_list.next;
2145                 list_del(node);
2146
2147                 if (ps->ps_node_init != NULL) {
2148                         /* still hold the lock */
2149                         ps->ps_node_init(pool, node);
2150                 }
2151                 spin_unlock(&ps->ps_lock);
2152                 return node;
2153         }
2154
2155         /* no available tx pool and ... */
2156         if (ps->ps_increasing) {
2157                 /* another thread is allocating a new pool */
2158                 spin_unlock(&ps->ps_lock);
2159                 trips++;
2160                 CDEBUG(D_NET,
2161                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2162                        ps->ps_name, interval, trips);
2163
2164                 schedule_timeout_interruptible(interval);
2165                 if (interval < cfs_time_seconds(1))
2166                         interval *= 2;
2167
2168                 goto again;
2169         }
2170
2171         if (ktime_get_seconds() < ps->ps_next_retry) {
2172                 /* someone failed recently */
2173                 spin_unlock(&ps->ps_lock);
2174                 return NULL;
2175         }
2176
2177         ps->ps_increasing = 1;
2178         spin_unlock(&ps->ps_lock);
2179
2180         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2181         time_before = ktime_get();
2182         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2183         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2184                ktime_ms_delta(ktime_get(), time_before));
2185
2186         spin_lock(&ps->ps_lock);
2187         ps->ps_increasing = 0;
2188         if (rc == 0) {
2189                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2190         } else {
2191                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2192                 CERROR("Can't allocate new %s pool because out of memory\n",
2193                        ps->ps_name);
2194         }
2195         spin_unlock(&ps->ps_lock);
2196
2197         goto again;
2198 }
2199
2200 static void
2201 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2202 {
2203         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2204                                                tpo_pool);
2205         int i;
2206
2207         LASSERT (pool->po_allocated == 0);
2208
2209         if (tpo->tpo_tx_pages != NULL) {
2210                 kiblnd_unmap_tx_pool(tpo);
2211                 kiblnd_free_pages(tpo->tpo_tx_pages);
2212         }
2213
2214         if (tpo->tpo_tx_descs == NULL)
2215                 goto out;
2216
2217         for (i = 0; i < pool->po_size; i++) {
2218                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2219                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2220
2221                 list_del(&tx->tx_list);
2222                 if (tx->tx_pages != NULL)
2223                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2224                 if (tx->tx_frags != NULL)
2225                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2226                                            IBLND_MAX_RDMA_FRAGS);
2227                 if (tx->tx_wrq != NULL)
2228                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2229                                            IBLND_MAX_RDMA_FRAGS);
2230                 if (tx->tx_sge != NULL) {
2231                         /* +1 is for the lnet header/message itself */
2232                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2233                                            (IBLND_MAX_RDMA_FRAGS *
2234                                            wrq_sge + 1));
2235                 }
2236                 if (tx->tx_rd != NULL)
2237                         LIBCFS_FREE(tx->tx_rd,
2238                                     offsetof(struct kib_rdma_desc,
2239                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2240         }
2241
2242         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2243 out:
2244         kiblnd_fini_pool(pool);
2245         CFS_FREE_PTR(tpo);
2246 }
2247
2248 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2249 {
2250         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2251         int ntx;
2252
2253         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2254         ntx = tunables->lnd_ntx / ncpts;
2255
2256         return max(IBLND_TX_POOL, ntx);
2257 }
2258
2259 static int
2260 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2261 {
2262         int            i;
2263         int            npg;
2264         struct kib_pool *pool;
2265         struct kib_tx_pool *tpo;
2266
2267         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2268         if (tpo == NULL) {
2269                 CERROR("Failed to allocate TX pool\n");
2270                 return -ENOMEM;
2271         }
2272
2273         pool = &tpo->tpo_pool;
2274         kiblnd_init_pool(ps, pool, size);
2275         tpo->tpo_tx_descs = NULL;
2276         tpo->tpo_tx_pages = NULL;
2277
2278         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2279         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2280                 CERROR("Can't allocate tx pages: %d\n", npg);
2281                 CFS_FREE_PTR(tpo);
2282                 return -ENOMEM;
2283         }
2284
2285         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2286                          size * sizeof(struct kib_tx));
2287         if (tpo->tpo_tx_descs == NULL) {
2288                 CERROR("Can't allocate %d tx descriptors\n", size);
2289                 ps->ps_pool_destroy(pool);
2290                 return -ENOMEM;
2291         }
2292
2293         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2294
2295         for (i = 0; i < size; i++) {
2296                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2297                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2298
2299                 tx->tx_pool = tpo;
2300                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2301                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2302                                          lnet_cpt_table(), ps->ps_cpt,
2303                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2304                         if (tx->tx_pages == NULL)
2305                                 break;
2306                 }
2307
2308                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2309                                  IBLND_MAX_RDMA_FRAGS *
2310                                  sizeof(*tx->tx_frags));
2311                 if (tx->tx_frags == NULL)
2312                         break;
2313
2314                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2315
2316                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2317                                  IBLND_MAX_RDMA_FRAGS *
2318                                  sizeof(*tx->tx_wrq));
2319                 if (tx->tx_wrq == NULL)
2320                         break;
2321
2322                 /* +1 is for the lnet header/message itself */
2323                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2324                                  (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
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(struct kib_rdma_desc,
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(struct kib_pool *pool, struct list_head *node)
2348 {
2349         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2350                                                   struct kib_tx_poolset,
2351                                                   tps_poolset);
2352         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2353
2354         tx->tx_cookie = tps->tps_next_tx_cookie++;
2355 }
2356
2357 static void
2358 kiblnd_net_fini_pools(struct kib_net *net)
2359 {
2360         int     i;
2361
2362         cfs_cpt_for_each(i, lnet_cpt_table()) {
2363                 struct kib_tx_poolset *tps;
2364                 struct kib_fmr_poolset *fps;
2365
2366                 if (net->ibn_tx_ps != NULL) {
2367                         tps = net->ibn_tx_ps[i];
2368                         kiblnd_fini_poolset(&tps->tps_poolset);
2369                 }
2370
2371                 if (net->ibn_fmr_ps != NULL) {
2372                         fps = net->ibn_fmr_ps[i];
2373                         kiblnd_fini_fmr_poolset(fps);
2374                 }
2375         }
2376
2377         if (net->ibn_tx_ps != NULL) {
2378                 cfs_percpt_free(net->ibn_tx_ps);
2379                 net->ibn_tx_ps = NULL;
2380         }
2381
2382         if (net->ibn_fmr_ps != NULL) {
2383                 cfs_percpt_free(net->ibn_fmr_ps);
2384                 net->ibn_fmr_ps = NULL;
2385         }
2386 }
2387
2388 static int
2389 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2390                       int ncpts)
2391 {
2392         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2393 #ifdef HAVE_IB_GET_DMA_MR
2394         unsigned long   flags;
2395 #endif
2396         int             cpt;
2397         int             rc;
2398         int             i;
2399
2400         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2401
2402 #ifdef HAVE_IB_GET_DMA_MR
2403         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2404         /*
2405          * if lnd_map_on_demand is zero then we have effectively disabled
2406          * FMR or FastReg and we're using global memory regions
2407          * exclusively.
2408          */
2409         if (!tunables->lnd_map_on_demand) {
2410                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2411                                            flags);
2412                 goto create_tx_pool;
2413         }
2414
2415         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2416 #endif
2417
2418         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2419                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2420                        tunables->lnd_fmr_pool_size,
2421                        tunables->lnd_ntx / 4);
2422                 rc = -EINVAL;
2423                 goto failed;
2424         }
2425
2426         /* TX pool must be created later than FMR, see LU-2268
2427          * for details */
2428         LASSERT(net->ibn_tx_ps == NULL);
2429
2430         /* premapping can fail if ibd_nmr > 1, so we always create
2431          * FMR pool and map-on-demand if premapping failed */
2432
2433         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2434                                            sizeof(struct kib_fmr_poolset));
2435         if (net->ibn_fmr_ps == NULL) {
2436                 CERROR("Failed to allocate FMR pool array\n");
2437                 rc = -ENOMEM;
2438                 goto failed;
2439         }
2440
2441         for (i = 0; i < ncpts; i++) {
2442                 cpt = (cpts == NULL) ? i : cpts[i];
2443                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2444                                              net, tunables);
2445                 if (rc != 0) {
2446                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2447                                cpt, rc);
2448                         goto failed;
2449                 }
2450         }
2451
2452         if (i > 0)
2453                 LASSERT(i == ncpts);
2454
2455 #ifdef HAVE_IB_GET_DMA_MR
2456  create_tx_pool:
2457 #endif
2458         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2459                                           sizeof(struct kib_tx_poolset));
2460         if (net->ibn_tx_ps == NULL) {
2461                 CERROR("Failed to allocate tx pool array\n");
2462                 rc = -ENOMEM;
2463                 goto failed;
2464         }
2465
2466         for (i = 0; i < ncpts; i++) {
2467                 cpt = (cpts == NULL) ? i : cpts[i];
2468                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2469                                          cpt, net, "TX",
2470                                          kiblnd_tx_pool_size(ni, ncpts),
2471                                          kiblnd_create_tx_pool,
2472                                          kiblnd_destroy_tx_pool,
2473                                          kiblnd_tx_init, NULL);
2474                 if (rc != 0) {
2475                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2476                                cpt, rc);
2477                         goto failed;
2478                 }
2479         }
2480
2481         return 0;
2482  failed:
2483         kiblnd_net_fini_pools(net);
2484         LASSERT(rc != 0);
2485         return rc;
2486 }
2487
2488 static int
2489 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2490 {
2491         struct ib_port_attr *port_attr;
2492         int rc;
2493         unsigned long flags;
2494         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2495
2496         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2497         if (port_attr == NULL) {
2498                 CDEBUG(D_NETERROR, "Out of memory\n");
2499                 return -ENOMEM;
2500         }
2501
2502         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2503
2504         write_lock_irqsave(g_lock, flags);
2505
2506         if (rc == 0)
2507                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2508                                  ? IBLND_DEV_PORT_ACTIVE
2509                                  : IBLND_DEV_PORT_DOWN;
2510
2511         write_unlock_irqrestore(g_lock, flags);
2512         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2513
2514         if (rc != 0) {
2515                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2516                 return rc;
2517         }
2518         return 0;
2519 }
2520
2521 static inline void
2522 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2523 {
2524         struct kib_net  *net;
2525
2526         /* for health check */
2527         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2528                 if (val)
2529                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2530                                         libcfs_nidstr(&net->ibn_ni->ni_nid));
2531                 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2532         }
2533 }
2534
2535 void
2536 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2537 {
2538         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2539         struct kib_hca_dev  *hdev;
2540         unsigned long flags;
2541
2542         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2543
2544         write_lock_irqsave(g_lock, flags);
2545
2546         switch (event->event) {
2547         case IB_EVENT_DEVICE_FATAL:
2548                 CDEBUG(D_NET, "IB device fatal\n");
2549                 hdev->ibh_state = IBLND_DEV_FATAL;
2550                 kiblnd_set_ni_fatal_on(hdev, 1);
2551                 break;
2552         case IB_EVENT_PORT_ACTIVE:
2553                 CDEBUG(D_NET, "IB port active\n");
2554                 if (event->element.port_num == hdev->ibh_port) {
2555                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2556                         kiblnd_set_ni_fatal_on(hdev, 0);
2557                 }
2558                 break;
2559         case IB_EVENT_PORT_ERR:
2560                 CDEBUG(D_NET, "IB port err\n");
2561                 if (event->element.port_num == hdev->ibh_port) {
2562                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2563                         kiblnd_set_ni_fatal_on(hdev, 1);
2564                 }
2565                 break;
2566         default:
2567                 break;
2568         }
2569         write_unlock_irqrestore(g_lock, flags);
2570 }
2571
2572 static int
2573 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2574 {
2575         struct ib_device_attr *dev_attr;
2576         int rc = 0;
2577         int rc2 = 0;
2578
2579         /* It's safe to assume a HCA can handle a page size
2580          * matching that of the native system */
2581         hdev->ibh_page_shift = PAGE_SHIFT;
2582         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2583         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2584
2585 #ifndef HAVE_IB_DEVICE_ATTRS
2586         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2587         if (dev_attr == NULL) {
2588                 CERROR("Out of memory\n");
2589                 return -ENOMEM;
2590         }
2591
2592         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2593         if (rc != 0) {
2594                 CERROR("Failed to query IB device: %d\n", rc);
2595                 goto out_clean_attr;
2596         }
2597 #else
2598         dev_attr = &hdev->ibh_ibdev->attrs;
2599 #endif
2600
2601         hdev->ibh_mr_size = dev_attr->max_mr_size;
2602         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2603
2604         /* Setup device Memory Registration capabilities */
2605 #ifdef HAVE_FMR_POOL_API
2606 #ifdef HAVE_IB_DEVICE_OPS
2607         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2608             hdev->ibh_ibdev->ops.dealloc_fmr &&
2609             hdev->ibh_ibdev->ops.map_phys_fmr &&
2610             hdev->ibh_ibdev->ops.unmap_fmr) {
2611 #else
2612         if (hdev->ibh_ibdev->alloc_fmr &&
2613             hdev->ibh_ibdev->dealloc_fmr &&
2614             hdev->ibh_ibdev->map_phys_fmr &&
2615             hdev->ibh_ibdev->unmap_fmr) {
2616 #endif
2617                 LCONSOLE_INFO("Using FMR for registration\n");
2618                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2619         } else
2620 #endif /* HAVE_FMR_POOL_API */
2621         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2622                 LCONSOLE_INFO("Using FastReg for registration\n");
2623                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2624 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2625 #ifdef IB_DEVICE_SG_GAPS_REG
2626                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2627                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2628 #endif
2629 #endif
2630         } else {
2631                 rc = -ENOSYS;
2632         }
2633
2634         rc2 = kiblnd_port_get_attr(hdev);
2635         if (rc2 != 0)
2636                 return rc2;
2637
2638         if (rc != 0)
2639                 rc = -EINVAL;
2640
2641 #ifndef HAVE_IB_DEVICE_ATTRS
2642 out_clean_attr:
2643         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2644 #endif
2645
2646         if (rc == -ENOSYS)
2647                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2648                        "register memory: %d\n", rc);
2649         else if (rc == -EINVAL)
2650                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2651         return rc;
2652 }
2653
2654 #ifdef HAVE_IB_GET_DMA_MR
2655 static void
2656 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2657 {
2658         if (hdev->ibh_mrs == NULL)
2659                 return;
2660
2661         ib_dereg_mr(hdev->ibh_mrs);
2662
2663         hdev->ibh_mrs = NULL;
2664 }
2665 #endif
2666
2667 void
2668 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2669 {
2670         if (hdev->ibh_event_handler.device != NULL)
2671                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2672
2673 #ifdef HAVE_IB_GET_DMA_MR
2674         kiblnd_hdev_cleanup_mrs(hdev);
2675 #endif
2676
2677         if (hdev->ibh_pd != NULL)
2678                 ib_dealloc_pd(hdev->ibh_pd);
2679
2680         if (hdev->ibh_cmid != NULL)
2681                 rdma_destroy_id(hdev->ibh_cmid);
2682
2683         LIBCFS_FREE(hdev, sizeof(*hdev));
2684 }
2685
2686 #ifdef HAVE_IB_GET_DMA_MR
2687 static int
2688 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2689 {
2690         struct ib_mr *mr;
2691         int           acflags = IB_ACCESS_LOCAL_WRITE |
2692                                 IB_ACCESS_REMOTE_WRITE;
2693
2694         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2695         if (IS_ERR(mr)) {
2696                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2697                 kiblnd_hdev_cleanup_mrs(hdev);
2698                 return PTR_ERR(mr);
2699         }
2700
2701         hdev->ibh_mrs = mr;
2702
2703         return 0;
2704 }
2705 #endif
2706
2707 static int
2708 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2709 {       /* DUMMY */
2710         return 0;
2711 }
2712
2713 static int kiblnd_get_link_status(struct net_device *dev)
2714 {
2715         int ret = -1;
2716
2717         LASSERT(dev);
2718
2719         if (!netif_running(dev))
2720                 ret = 0;
2721         /* Some devices may not be providing link settings */
2722         else if (dev->ethtool_ops->get_link)
2723                 ret = dev->ethtool_ops->get_link(dev);
2724
2725         return ret;
2726 }
2727
2728 static int
2729 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2730 {
2731         struct rdma_cm_id  *cmid;
2732         struct sockaddr_in  srcaddr;
2733         struct sockaddr_in  dstaddr;
2734         int                 rc;
2735
2736         if (dev->ibd_hdev == NULL || /* initializing */
2737             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2738             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2739                 return 1;
2740
2741         /* XXX: it's UGLY, but I don't have better way to find
2742          * ib-bonding HCA failover because:
2743          *
2744          * a. no reliable CM event for HCA failover...
2745          * b. no OFED API to get ib_device for current net_device...
2746          *
2747          * We have only two choices at this point:
2748          *
2749          * a. rdma_bind_addr(), it will conflict with listener cmid
2750          * b. rdma_resolve_addr() to zero addr */
2751         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2752                                      RDMA_PS_TCP, IB_QPT_RC);
2753         if (IS_ERR(cmid)) {
2754                 rc = PTR_ERR(cmid);
2755                 CERROR("Failed to create cmid for failover: %d\n", rc);
2756                 return rc;
2757         }
2758
2759         memset(&srcaddr, 0, sizeof(srcaddr));
2760         srcaddr.sin_family      = AF_INET;
2761         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2762
2763         memset(&dstaddr, 0, sizeof(dstaddr));
2764         dstaddr.sin_family = AF_INET;
2765         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2766                                (struct sockaddr *)&dstaddr, 1);
2767         if (rc != 0 || cmid->device == NULL) {
2768                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2769                        dev->ibd_ifname, &dev->ibd_ifip,
2770                        cmid->device, rc);
2771                 rdma_destroy_id(cmid);
2772                 return rc;
2773         }
2774
2775         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2776         rdma_destroy_id(cmid);
2777         return rc;
2778 }
2779
2780 int
2781 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2782 {
2783         LIST_HEAD(zombie_tpo);
2784         LIST_HEAD(zombie_ppo);
2785         LIST_HEAD(zombie_fpo);
2786         struct rdma_cm_id  *cmid  = NULL;
2787         struct kib_hca_dev *hdev  = NULL;
2788         struct kib_hca_dev *old;
2789         struct ib_pd       *pd;
2790         struct kib_net *net;
2791         struct sockaddr_in  addr;
2792         struct net_device *netdev;
2793         unsigned long       flags;
2794         int                 rc = 0;
2795         int                 i;
2796         bool                set_fatal = true;
2797
2798         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2799                 dev->ibd_can_failover ||
2800                 dev->ibd_hdev == NULL);
2801
2802         rc = kiblnd_dev_need_failover(dev, ns);
2803         if (rc <= 0)
2804                 goto out;
2805
2806         if (dev->ibd_hdev != NULL &&
2807             dev->ibd_hdev->ibh_cmid != NULL) {
2808                 /* XXX it's not good to close old listener at here,
2809                  * because we can fail to create new listener.
2810                  * But we have to close it now, otherwise rdma_bind_addr
2811                  * will return EADDRINUSE... How crap! */
2812                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2813
2814                 cmid = dev->ibd_hdev->ibh_cmid;
2815                 /* make next schedule of kiblnd_dev_need_failover()
2816                  * return 1 for me */
2817                 dev->ibd_hdev->ibh_cmid  = NULL;
2818                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2819
2820                 rdma_destroy_id(cmid);
2821         }
2822
2823         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2824                                      IB_QPT_RC);
2825         if (IS_ERR(cmid)) {
2826                 rc = PTR_ERR(cmid);
2827                 CERROR("Failed to create cmid for failover: %d\n", rc);
2828                 goto out;
2829         }
2830
2831         memset(&addr, 0, sizeof(addr));
2832         addr.sin_family      = AF_INET;
2833         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2834         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2835
2836         /* Bind to failover device or port */
2837         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2838         if (rc != 0 || cmid->device == NULL) {
2839                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2840                        dev->ibd_ifname, &dev->ibd_ifip,
2841                        cmid->device, rc);
2842                 if (!rc && !cmid->device)
2843                         set_fatal = false;
2844                 rdma_destroy_id(cmid);
2845                 goto out;
2846         }
2847
2848         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2849         if (hdev == NULL) {
2850                 CERROR("Failed to allocate kib_hca_dev\n");
2851                 rdma_destroy_id(cmid);
2852                 rc = -ENOMEM;
2853                 goto out;
2854         }
2855
2856         atomic_set(&hdev->ibh_ref, 1);
2857         hdev->ibh_dev   = dev;
2858         hdev->ibh_cmid  = cmid;
2859         hdev->ibh_ibdev = cmid->device;
2860         hdev->ibh_port  = cmid->port_num;
2861
2862 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2863         pd = ib_alloc_pd(cmid->device, 0);
2864 #else
2865         pd = ib_alloc_pd(cmid->device);
2866 #endif
2867         if (IS_ERR(pd)) {
2868                 rc = PTR_ERR(pd);
2869                 CERROR("Can't allocate PD: %d\n", rc);
2870                 goto out;
2871         }
2872
2873         hdev->ibh_pd = pd;
2874
2875         rc = rdma_listen(cmid, 0);
2876         if (rc != 0) {
2877                 CERROR("Can't start new listener: %d\n", rc);
2878                 goto out;
2879         }
2880
2881         rc = kiblnd_hdev_get_attr(hdev);
2882         if (rc != 0) {
2883                 CERROR("Can't get device attributes: %d\n", rc);
2884                 goto out;
2885         }
2886
2887 #ifdef HAVE_IB_GET_DMA_MR
2888         rc = kiblnd_hdev_setup_mrs(hdev);
2889         if (rc != 0) {
2890                 CERROR("Can't setup device: %d\n", rc);
2891                 goto out;
2892         }
2893 #endif
2894
2895         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2896                                 hdev->ibh_ibdev, kiblnd_event_handler);
2897         ib_register_event_handler(&hdev->ibh_event_handler);
2898
2899         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2900
2901         old = dev->ibd_hdev;
2902         dev->ibd_hdev = hdev;   /* take over the refcount */
2903         hdev = old;
2904
2905         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2906                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2907                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2908                                             &zombie_tpo);
2909
2910                         if (net->ibn_fmr_ps != NULL)
2911                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2912                                                         &zombie_fpo);
2913                 }
2914         }
2915
2916         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2917  out:
2918         if (!list_empty(&zombie_tpo))
2919                 kiblnd_destroy_pool_list(&zombie_tpo);
2920         if (!list_empty(&zombie_ppo))
2921                 kiblnd_destroy_pool_list(&zombie_ppo);
2922         if (!list_empty(&zombie_fpo))
2923                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2924         if (hdev != NULL)
2925                 kiblnd_hdev_decref(hdev);
2926
2927         if (rc != 0) {
2928                 dev->ibd_failed_failover++;
2929         } else {
2930                 dev->ibd_failed_failover = 0;
2931
2932                 if (set_fatal) {
2933                         rcu_read_lock();
2934                         netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
2935                         if (netdev && (kiblnd_get_link_status(netdev) == 1))
2936                                 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
2937                         rcu_read_unlock();
2938                 }
2939         }
2940
2941         return rc;
2942 }
2943
2944 void
2945 kiblnd_destroy_dev(struct kib_dev *dev)
2946 {
2947         LASSERT(dev->ibd_nnets == 0);
2948         LASSERT(list_empty(&dev->ibd_nets));
2949
2950         list_del(&dev->ibd_fail_list);
2951         list_del(&dev->ibd_list);
2952
2953         if (dev->ibd_hdev != NULL)
2954                 kiblnd_hdev_decref(dev->ibd_hdev);
2955
2956         LIBCFS_FREE(dev, sizeof(*dev));
2957 }
2958
2959 static void
2960 kiblnd_base_shutdown(void)
2961 {
2962         struct kib_sched_info *sched;
2963         struct kib_peer_ni *peer_ni;
2964         int i;
2965
2966         LASSERT(list_empty(&kiblnd_data.kib_devs));
2967
2968         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
2969                libcfs_kmem_read());
2970
2971         switch (kiblnd_data.kib_init) {
2972         default:
2973                 LBUG();
2974
2975         case IBLND_INIT_ALL:
2976         case IBLND_INIT_DATA:
2977                 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
2978                         LASSERT(0);
2979                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2980                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2981                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2982                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2983
2984                 /* flag threads to terminate; wake and wait for them to die */
2985                 kiblnd_data.kib_shutdown = 1;
2986
2987                 /* NB: we really want to stop scheduler threads net by net
2988                  * instead of the whole module, this should be improved
2989                  * with dynamic configuration LNet.
2990                  */
2991                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2992                         wake_up_all(&sched->ibs_waitq);
2993
2994                 wake_up(&kiblnd_data.kib_connd_waitq);
2995                 wake_up(&kiblnd_data.kib_failover_waitq);
2996
2997                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2998                                        !atomic_read(&kiblnd_data.kib_nthreads),
2999                                        "Waiting for %d threads to terminate\n",
3000                                        atomic_read(&kiblnd_data.kib_nthreads));
3001                 fallthrough;
3002
3003         case IBLND_INIT_NOTHING:
3004                 break;
3005         }
3006
3007         if (kiblnd_data.kib_scheds != NULL)
3008                 cfs_percpt_free(kiblnd_data.kib_scheds);
3009
3010         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3011                libcfs_kmem_read());
3012
3013         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3014         module_put(THIS_MODULE);
3015 }
3016
3017 static void
3018 kiblnd_shutdown(struct lnet_ni *ni)
3019 {
3020         struct kib_net *net = ni->ni_data;
3021         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3022         unsigned long     flags;
3023
3024         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3025
3026         if (net == NULL)
3027                 goto out;
3028
3029         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3030                libcfs_kmem_read());
3031
3032         write_lock_irqsave(g_lock, flags);
3033         net->ibn_shutdown = 1;
3034         write_unlock_irqrestore(g_lock, flags);
3035
3036         switch (net->ibn_init) {
3037         default:
3038                 LBUG();
3039
3040         case IBLND_INIT_ALL:
3041                 /* nuke all existing peers within this net */
3042                 kiblnd_del_peer(ni, LNET_NID_ANY);
3043
3044                 /* Wait for all peer_ni state to clean up */
3045                 wait_var_event_warning(&net->ibn_npeers,
3046                                        atomic_read(&net->ibn_npeers) == 0,
3047                                        "%s: waiting for %d peers to disconnect\n",
3048                                        libcfs_nidstr(&ni->ni_nid),
3049                                        atomic_read(&net->ibn_npeers));
3050
3051                 kiblnd_net_fini_pools(net);
3052
3053                 write_lock_irqsave(g_lock, flags);
3054                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3055                 net->ibn_dev->ibd_nnets--;
3056                 list_del(&net->ibn_list);
3057                 write_unlock_irqrestore(g_lock, flags);
3058
3059                 wake_up_all(&kiblnd_data.kib_connd_waitq);
3060                 wait_var_event_warning(&net->ibn_nconns,
3061                                        atomic_read(&net->ibn_nconns) == 0,
3062                                        "%s: waiting for %d conns to clean\n",
3063                                        libcfs_nidstr(&ni->ni_nid),
3064                                        atomic_read(&net->ibn_nconns));
3065                 fallthrough;
3066
3067         case IBLND_INIT_NOTHING:
3068                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3069
3070                 if (net->ibn_dev != NULL &&
3071                     net->ibn_dev->ibd_nnets == 0)
3072                         kiblnd_destroy_dev(net->ibn_dev);
3073
3074                 break;
3075         }
3076
3077         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3078                libcfs_kmem_read());
3079
3080         net->ibn_init = IBLND_INIT_NOTHING;
3081         ni->ni_data = NULL;
3082
3083         LIBCFS_FREE(net, sizeof(*net));
3084
3085 out:
3086         if (list_empty(&kiblnd_data.kib_devs))
3087                 kiblnd_base_shutdown();
3088 }
3089
3090 static int
3091 kiblnd_base_startup(struct net *ns)
3092 {
3093         struct kib_sched_info *sched;
3094         int rc;
3095         int i;
3096
3097         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3098
3099         if (!try_module_get(THIS_MODULE))
3100                 goto failed;
3101
3102         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3103
3104         rwlock_init(&kiblnd_data.kib_global_lock);
3105
3106         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3107         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3108
3109         hash_init(kiblnd_data.kib_peers);
3110
3111         spin_lock_init(&kiblnd_data.kib_connd_lock);
3112         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3113         INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3114         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3115         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3116         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3117
3118         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3119         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3120
3121         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3122                                                   sizeof(*sched));
3123         if (kiblnd_data.kib_scheds == NULL)
3124                 goto failed;
3125
3126         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3127                 int     nthrs;
3128
3129                 spin_lock_init(&sched->ibs_lock);
3130                 INIT_LIST_HEAD(&sched->ibs_conns);
3131                 init_waitqueue_head(&sched->ibs_waitq);
3132
3133                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3134                 if (*kiblnd_tunables.kib_nscheds > 0) {
3135                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3136                 } else {
3137                         /* max to half of CPUs, another half is reserved for
3138                          * upper layer modules */
3139                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3140                 }
3141
3142                 sched->ibs_nthreads_max = nthrs;
3143                 sched->ibs_cpt = i;
3144         }
3145
3146         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3147
3148         /* lists/ptrs/locks initialised */
3149         kiblnd_data.kib_init = IBLND_INIT_DATA;
3150         /*****************************************************/
3151
3152         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3153         if (rc != 0) {
3154                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3155                 goto failed;
3156         }
3157
3158         if (*kiblnd_tunables.kib_dev_failover != 0)
3159                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3160                                          "kiblnd_failover");
3161
3162         if (rc != 0) {
3163                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3164                 goto failed;
3165         }
3166
3167         /* flag everything initialised */
3168         kiblnd_data.kib_init = IBLND_INIT_ALL;
3169         /*****************************************************/
3170
3171         return 0;
3172
3173  failed:
3174         kiblnd_base_shutdown();
3175         return -ENETDOWN;
3176 }
3177
3178 static int
3179 kiblnd_start_schedulers(struct kib_sched_info *sched)
3180 {
3181         int     rc = 0;
3182         int     nthrs;
3183         int     i;
3184
3185         if (sched->ibs_nthreads == 0) {
3186                 if (*kiblnd_tunables.kib_nscheds > 0) {
3187                         nthrs = sched->ibs_nthreads_max;
3188                 } else {
3189                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3190                                                sched->ibs_cpt);
3191                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3192                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3193                 }
3194         } else {
3195                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3196                 /* increase one thread if there is new interface */
3197                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3198         }
3199
3200         for (i = 0; i < nthrs; i++) {
3201                 long    id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3202
3203                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3204                                          "kiblnd_sd_%02ld_%02ld",
3205                                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3206                 if (rc == 0)
3207                         continue;
3208
3209                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3210                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3211                 break;
3212         }
3213
3214         sched->ibs_nthreads += i;
3215         return rc;
3216 }
3217
3218 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3219                                     int ncpts)
3220 {
3221         int     cpt;
3222         int     rc;
3223         int     i;
3224
3225         for (i = 0; i < ncpts; i++) {
3226                 struct kib_sched_info *sched;
3227
3228                 cpt = (cpts == NULL) ? i : cpts[i];
3229                 sched = kiblnd_data.kib_scheds[cpt];
3230
3231                 if (!newdev && sched->ibs_nthreads > 0)
3232                         continue;
3233
3234                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3235                 if (rc != 0) {
3236                         CERROR("Failed to start scheduler threads for %s\n",
3237                                dev->ibd_ifname);
3238                         return rc;
3239                 }
3240         }
3241         return 0;
3242 }
3243
3244 static struct kib_dev *
3245 kiblnd_dev_search(char *ifname)
3246 {
3247         struct kib_dev *alias = NULL;
3248         struct kib_dev *dev;
3249         char            *colon;
3250         char            *colon2;
3251
3252         colon = strchr(ifname, ':');
3253         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3254                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3255                         return dev;
3256
3257                 if (alias != NULL)
3258                         continue;
3259
3260                 colon2 = strchr(dev->ibd_ifname, ':');
3261                 if (colon != NULL)
3262                         *colon = 0;
3263                 if (colon2 != NULL)
3264                         *colon2 = 0;
3265
3266                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3267                         alias = dev;
3268
3269                 if (colon != NULL)
3270                         *colon = ':';
3271                 if (colon2 != NULL)
3272                         *colon2 = ':';
3273         }
3274         return alias;
3275 }
3276
3277 static int
3278 kiblnd_startup(struct lnet_ni *ni)
3279 {
3280         char *ifname = NULL;
3281         struct lnet_inetdev *ifaces = NULL;
3282         struct kib_dev *ibdev = NULL;
3283         struct kib_net *net = NULL;
3284         unsigned long flags;
3285         int rc;
3286         int i;
3287         bool newdev;
3288
3289         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3290
3291         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3292                 rc = kiblnd_base_startup(ni->ni_net_ns);
3293                 if (rc != 0)
3294                         return rc;
3295         }
3296
3297         LIBCFS_ALLOC(net, sizeof(*net));
3298         ni->ni_data = net;
3299         if (net == NULL) {
3300                 rc = -ENOMEM;
3301                 goto failed;
3302         }
3303
3304         net->ibn_ni = ni;
3305         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3306
3307         kiblnd_tunables_setup(ni);
3308
3309         /*
3310          * Multi-Rail wants each secondary
3311          * IP to be treated as an unique 'struct ni' interface.
3312          */
3313         if (ni->ni_interface != NULL) {
3314                 /* Use the IPoIB interface specified in 'networks=' */
3315                 ifname = ni->ni_interface;
3316         } else {
3317                 ifname = *kiblnd_tunables.kib_default_ipif;
3318         }
3319
3320         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3321                 CERROR("IPoIB interface name too long: %s\n", ifname);
3322                 rc = -E2BIG;
3323                 goto failed;
3324         }
3325
3326         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3327         if (rc < 0)
3328                 goto failed;
3329
3330         for (i = 0; i < rc; i++) {
3331                 if (strcmp(ifname, ifaces[i].li_name) == 0)
3332                         break;
3333         }
3334
3335         if (i == rc) {
3336                 CERROR("ko2iblnd: No matching interfaces\n");
3337                 rc = -ENOENT;
3338                 goto failed;
3339         }
3340
3341         ibdev = kiblnd_dev_search(ifname);
3342         newdev = ibdev == NULL;
3343         /* hmm...create kib_dev even for alias */
3344         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3345                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3346                 if (!ibdev) {
3347                         rc = -ENOMEM;
3348                         goto failed;
3349                 }
3350
3351                 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3352                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3353                         sizeof(ibdev->ibd_ifname));
3354                 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3355
3356                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3357                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3358                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3359
3360                 /* initialize the device */
3361                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3362                 if (rc) {
3363                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3364                                rc);
3365                         goto failed;
3366                 }
3367
3368                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3369         }
3370
3371         net->ibn_dev = ibdev;
3372         ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3373
3374         ni->ni_dev_cpt = ifaces[i].li_cpt;
3375
3376         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3377         if (rc != 0)
3378                 goto failed;
3379
3380         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3381         if (rc != 0) {
3382                 CERROR("Failed to initialize NI pools: %d\n", rc);
3383                 goto failed;
3384         }
3385
3386         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3387         ibdev->ibd_nnets++;
3388         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3389         /* for health check */
3390         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3391                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3392         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3393
3394         net->ibn_init = IBLND_INIT_ALL;
3395
3396         return 0;
3397
3398 failed:
3399         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3400                 kiblnd_destroy_dev(ibdev);
3401
3402         kfree(ifaces);
3403         kiblnd_shutdown(ni);
3404
3405         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3406                ifname ? ifname : "", rc);
3407
3408         return -ENETDOWN;
3409 }
3410
3411 static const struct lnet_lnd the_o2iblnd = {
3412         .lnd_type       = O2IBLND,
3413         .lnd_startup    = kiblnd_startup,
3414         .lnd_shutdown   = kiblnd_shutdown,
3415         .lnd_ctl        = kiblnd_ctl,
3416         .lnd_send       = kiblnd_send,
3417         .lnd_recv       = kiblnd_recv,
3418         .lnd_get_dev_prio = kiblnd_get_dev_prio,
3419 };
3420
3421 static void ko2inlnd_assert_wire_constants(void)
3422 {
3423         BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3424         BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3425         BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3426         BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3427
3428         BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3429         BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3430         BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3431         BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3432         BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3433         BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3434         BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3435         BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3436         BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3437         BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3438
3439         BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3440         BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3441         BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3442         BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3443         BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3444         BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3445         BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3446         BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3447
3448         BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3449         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3450         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3451         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3452         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3453         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3454         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3455
3456         BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3457         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3458         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3459         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3460         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3461
3462         BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3463         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3464         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3465         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3466         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3467
3468         BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3469         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3470         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3471         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3472         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3473         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3474         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3475
3476         BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3477         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3478         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3479         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3480         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3481
3482         BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3483         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3484         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3485         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3486         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3487         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3488         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3489
3490         BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3491         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3492         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3493         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3494         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3495         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3496         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3497
3498         BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3499         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3500         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3501         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3502         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3503
3504         /* Checks for struct kib_msg */
3505         //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3506         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3507         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3508         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3509         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3510         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3511         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3512         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3513         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3514         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3515         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3516         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3517         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3518         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3519         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3520         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3521         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3522         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3523         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3524         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3525         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3526
3527         /* Connparams */
3528         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3529         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3530         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3531         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3532         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3533         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3534
3535         /* Immediate message */
3536         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3537         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3538         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3539         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3540
3541         /* PUT req message */
3542         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3543         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3544         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3545         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3546
3547         /* Put ACK */
3548         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3549         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3550         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3551         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3552         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3553         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3554
3555         /* GET message */
3556         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3557         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3558         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3559         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3560         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3561         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3562
3563         /* Completion message */
3564         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3565         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3566         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3567         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3568
3569         /* Sanity checks */
3570         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3571         BUILD_BUG_ON(offsetof(struct kib_msg,
3572                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3573                      IBLND_MSG_SIZE);
3574         BUILD_BUG_ON(offsetof(struct kib_msg,
3575                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3576                      IBLND_MSG_SIZE);
3577 }
3578
3579 static void __exit ko2iblnd_exit(void)
3580 {
3581         lnet_unregister_lnd(&the_o2iblnd);
3582 }
3583
3584 static int __init ko2iblnd_init(void)
3585 {
3586         int rc;
3587
3588         ko2inlnd_assert_wire_constants();
3589
3590         rc = kiblnd_tunables_init();
3591         if (rc != 0)
3592                 return rc;
3593
3594         lnet_register_lnd(&the_o2iblnd);
3595
3596         return 0;
3597 }
3598
3599 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3600 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3601 MODULE_VERSION("2.8.0");
3602 MODULE_LICENSE("GPL");
3603
3604 module_init(ko2iblnd_init);
3605 module_exit(ko2iblnd_exit);