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         atomic_set(&peer_ni->ibp_nconns, 0);
 343
 344         INIT_HLIST_NODE(&peer_ni->ibp_list);
 345         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 346         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer_ni;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer(struct kref *kref)
 364 {
 365         struct kib_peer_ni *peer_ni = container_of(kref, struct kib_peer_ni,
 366                                                    ibp_kref);
 367         struct kib_net *net = peer_ni->ibp_ni->ni_data;
 368
 369         LASSERT(net != NULL);
 370         LASSERT(!kiblnd_peer_active(peer_ni));
 371         LASSERT(kiblnd_peer_idle(peer_ni));
 372         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 373
 374         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 375
 376         /* NB a peer_ni's connections keep a reference on their peer_ni until
 377          * they are destroyed, so we can be assured that _all_ state to do
 378          * with this peer_ni has been cleaned up when its refcount drops to
 379          * zero.
 380          */
 381         if (atomic_dec_and_test(&net->ibn_npeers))
 382                 wake_up_var(&net->ibn_npeers);
 383 }
 384
 385 struct kib_peer_ni *
 386 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 387 {
 388         /* the caller is responsible for accounting the additional reference
 389          * that this creates
 390          */
 391         struct kib_peer_ni *peer_ni;
 392
 393         hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
 394                                ibp_list, nid) {
 395                 LASSERT(!kiblnd_peer_idle(peer_ni));
 396
 397                 /*
 398                  * Match a peer if its NID and the NID of the local NI it
 399                  * communicates over are the same. Otherwise don't match
 400                  * the peer, which will result in a new lnd peer being
 401                  * created.
 402                  */
 403                 if (peer_ni->ibp_nid != nid ||
 404                     !nid_same(&peer_ni->ibp_ni->ni_nid, &ni->ni_nid))
 405                         continue;
 406
 407                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 408                        peer_ni, libcfs_nid2str(nid),
 409                        kref_read(&peer_ni->ibp_kref),
 410                        peer_ni->ibp_version);
 411                 return peer_ni;
 412         }
 413         return NULL;
 414 }
 415
 416 void
 417 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
 418 {
 419         LASSERT(list_empty(&peer_ni->ibp_conns));
 420
 421         LASSERT(kiblnd_peer_active(peer_ni));
 422         hlist_del_init(&peer_ni->ibp_list);
 423         /* lose peerlist's ref */
 424         kiblnd_peer_decref(peer_ni);
 425 }
 426
 427
 428 static void
 429 kiblnd_debug_rx(struct kib_rx *rx)
 430 {
 431         CDEBUG(D_CONSOLE, "      %p msg_type %x cred %d\n",
 432                rx, rx->rx_msg->ibm_type,
 433                rx->rx_msg->ibm_credits);
 434 }
 435
 436 static void
 437 kiblnd_debug_tx(struct kib_tx *tx)
 438 {
 439         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 440                "cookie %#llx msg %s%s type %x cred %d\n",
 441                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 442                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 443                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 444                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 445                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 446 }
 447
 448 static void
 449 kiblnd_debug_conn(struct kib_conn *conn)
 450 {
 451         struct list_head        *tmp;
 452         int                     i;
 453
 454         spin_lock(&conn->ibc_lock);
 455
 456         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 457                atomic_read(&conn->ibc_refcount), conn,
 458                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 459         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 460                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 461                conn->ibc_nsends_posted, conn->ibc_credits,
 462                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 463         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 464
 465         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 466         list_for_each(tmp, &conn->ibc_early_rxs)
 467                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 468
 469         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 470         list_for_each(tmp, &conn->ibc_tx_noops)
 471                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 472
 473         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 474         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 475                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 476
 477         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 478         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 479                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 480
 481         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 482         list_for_each(tmp, &conn->ibc_tx_queue)
 483                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 484
 485         CDEBUG(D_CONSOLE, "   active_txs:\n");
 486         list_for_each(tmp, &conn->ibc_active_txs)
 487                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 488
 489         CDEBUG(D_CONSOLE, "   rxs:\n");
 490         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 491                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 492
 493         spin_unlock(&conn->ibc_lock);
 494 }
 495
 496 static void
 497 kiblnd_dump_peer_debug_info(struct kib_peer_ni *peer_ni)
 498 {
 499         struct kib_conn *conn;
 500         struct kib_conn *cnxt;
 501         int count = 0;
 502
 503         CDEBUG(D_CONSOLE, "[last_alive, races, reconnected, error]: %lld, %d, %d, %d\n",
 504                peer_ni->ibp_last_alive,
 505                peer_ni->ibp_races,
 506                peer_ni->ibp_reconnected,
 507                peer_ni->ibp_error);
 508         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 509                                  ibc_list) {
 510                 CDEBUG(D_CONSOLE, "Conn %d:\n", count);
 511                 kiblnd_debug_conn(conn);
 512                 count++;
 513         }
 514 }
 515
 516
 517 static int
 518 kiblnd_get_peer_info(struct lnet_ni *ni, lnet_nid_t nid, int index,
 519                      lnet_nid_t *nidp, int *count)
 520 {
 521         struct kib_peer_ni              *peer_ni;
 522         int                      i;
 523         unsigned long            flags;
 524
 525         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 526
 527         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 528                 LASSERT(!kiblnd_peer_idle(peer_ni));
 529
 530                 if (peer_ni->ibp_ni != ni)
 531                         continue;
 532
 533                 if (peer_ni->ibp_nid == nid)
 534                         kiblnd_dump_peer_debug_info(peer_ni);
 535
 536                 if (index-- > 0)
 537                         continue;
 538
 539                 *nidp = peer_ni->ibp_nid;
 540                 *count = kref_read(&peer_ni->ibp_kref);
 541
 542                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 543                 return 0;
 544         }
 545
 546         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 547         return -ENOENT;
 548 }
 549
 550 static void
 551 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
 552 {
 553         struct kib_conn *cnxt;
 554         struct kib_conn *conn;
 555
 556         if (list_empty(&peer_ni->ibp_conns)) {
 557                 kiblnd_unlink_peer_locked(peer_ni);
 558         } else {
 559                 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 560                                          ibc_list)
 561                         kiblnd_close_conn_locked(conn, 0);
 562                 /* NB closing peer_ni's last conn unlinked it. */
 563         }
 564         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 565          * last ref on it. */
 566 }
 567
 568 static int
 569 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 570 {
 571         LIST_HEAD(zombies);
 572         struct hlist_node *pnxt;
 573         struct kib_peer_ni *peer_ni;
 574         int lo;
 575         int hi;
 576         int i;
 577         unsigned long flags;
 578         int rc = -ENOENT;
 579
 580         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 581
 582         if (nid != LNET_NID_ANY) {
 583                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
 584                 hi = lo;
 585         } else {
 586                 lo = 0;
 587                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
 588         }
 589
 590         for (i = lo; i <= hi; i++) {
 591                 hlist_for_each_entry_safe(peer_ni, pnxt,
 592                                           &kiblnd_data.kib_peers[i], ibp_list) {
 593                         LASSERT(!kiblnd_peer_idle(peer_ni));
 594
 595                         if (peer_ni->ibp_ni != ni)
 596                                 continue;
 597
 598                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 599                                 continue;
 600
 601                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 602                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 603
 604                                 list_splice_init(&peer_ni->ibp_tx_queue,
 605                                                  &zombies);
 606                         }
 607
 608                         kiblnd_del_peer_locked(peer_ni);
 609                         rc = 0;         /* matched something */
 610                 }
 611         }
 612
 613         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 614
 615         kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
 616
 617         return rc;
 618 }
 619
 620 static struct kib_conn *
 621 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 622 {
 623         struct kib_peer_ni *peer_ni;
 624         struct kib_conn *conn;
 625         int i;
 626         unsigned long flags;
 627
 628         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 629
 630         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 631                 LASSERT(!kiblnd_peer_idle(peer_ni));
 632
 633                 if (peer_ni->ibp_ni != ni)
 634                         continue;
 635
 636                 list_for_each_entry(conn, &peer_ni->ibp_conns,
 637                                     ibc_list) {
 638                         if (index-- > 0)
 639                                 continue;
 640
 641                         kiblnd_conn_addref(conn);
 642                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 643                                                flags);
 644                         return conn;
 645                 }
 646         }
 647
 648         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 649         return NULL;
 650 }
 651
 652 static void
 653 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 654 {
 655         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 656         if (cmid->route.path_rec == NULL)
 657                 return;
 658
 659         if (*kiblnd_tunables.kib_ib_mtu)
 660                 cmid->route.path_rec->mtu =
 661                         ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
 662 }
 663
 664 static int
 665 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 666 {
 667         cpumask_var_t   *mask;
 668         int             vectors;
 669         int             off;
 670         int             i;
 671         lnet_nid_t      ibp_nid;
 672
 673         vectors = conn->ibc_cmid->device->num_comp_vectors;
 674         if (vectors <= 1)
 675                 return 0;
 676
 677         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 678
 679         /* hash NID to CPU id in this partition... when targeting a single peer
 680          * with multiple QPs, to engage more cores in CQ processing to a single
 681          * peer, use ibp_nconns to salt the value the comp_vector value
 682          */
 683         ibp_nid = conn->ibc_peer->ibp_nid +
 684                 atomic_read(&conn->ibc_peer->ibp_nconns);
 685         off = do_div(ibp_nid, cpumask_weight(*mask));
 686         for_each_cpu(i, *mask) {
 687                 if (off-- == 0)
 688                         return i % vectors;
 689         }
 690
 691         LBUG();
 692         return 1;
 693 }
 694
 695 /*
 696  * Get the scheduler bound to this CPT. If the scheduler has no
 697  * threads, which means that the CPT has no CPUs, then grab the
 698  * next scheduler that we can use.
 699  *
 700  * This case would be triggered if a NUMA node is configured with
 701  * no associated CPUs.
 702  */
 703 static struct kib_sched_info *
 704 kiblnd_get_scheduler(int cpt)
 705 {
 706         struct kib_sched_info *sched;
 707         int i;
 708
 709         sched = kiblnd_data.kib_scheds[cpt];
 710
 711         if (sched->ibs_nthreads > 0)
 712                 return sched;
 713
 714         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 715                 if (sched->ibs_nthreads > 0) {
 716                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 717                                         cpt, sched->ibs_cpt);
 718                         return sched;
 719                 }
 720         }
 721
 722         return NULL;
 723 }
 724
 725 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 726 {
 727         /*
 728          * One WR for the LNet message
 729          * And ibc_max_frags for the transfer WRs
 730          */
 731         int ret;
 732         int multiplier = 1 + conn->ibc_max_frags;
 733
 734         /* FastReg needs two extra WRs for map and invalidate */
 735         if (IS_FAST_REG_DEV(conn->ibc_hdev->ibh_dev))
 736                 multiplier += 2;
 737
 738         /* account for a maximum of ibc_queue_depth in-flight transfers */
 739         ret = multiplier * conn->ibc_queue_depth;
 740
 741         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 742                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 743                        "request size %d larger than maximum %d device "
 744                        "can handle\n", conn->ibc_queue_depth, ret,
 745                        conn->ibc_hdev->ibh_max_qp_wr);
 746                 conn->ibc_queue_depth =
 747                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 748         }
 749
 750         /* don't go beyond the maximum the device can handle */
 751         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 752 }
 753
 754 struct kib_conn *
 755 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 756                    int state, int version)
 757 {
 758         /* CAVEAT EMPTOR:
 759          * If the new conn is created successfully it takes over the caller's
 760          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 761          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 762          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 763          * to destroy 'cmid' here since I'm called from the CM which still has
 764          * its ref on 'cmid'). */
 765         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 766         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 767         struct kib_dev *dev;
 768         struct ib_qp_init_attr init_qp_attr = {};
 769         struct kib_sched_info   *sched;
 770 #ifdef HAVE_OFED_IB_CQ_INIT_ATTR
 771         struct ib_cq_init_attr  cq_attr = {};
 772 #endif
 773         struct kib_conn *conn;
 774         struct ib_cq            *cq;
 775         unsigned long           flags;
 776         int                     cpt;
 777         int                     rc;
 778         int                     i;
 779
 780         LASSERT(net != NULL);
 781         LASSERT(!in_interrupt());
 782
 783         dev = net->ibn_dev;
 784
 785         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 786         sched = kiblnd_get_scheduler(cpt);
 787
 788         if (sched == NULL) {
 789                 CERROR("no schedulers available. node is unhealthy\n");
 790                 goto failed_0;
 791         }
 792
 793         /*
 794          * The cpt might have changed if we ended up selecting a non cpt
 795          * native scheduler. So use the scheduler's cpt instead.
 796          */
 797         cpt = sched->ibs_cpt;
 798
 799         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 800         if (conn == NULL) {
 801                 CERROR("Can't allocate connection for %s\n",
 802                        libcfs_nid2str(peer_ni->ibp_nid));
 803                 goto failed_0;
 804         }
 805
 806         conn->ibc_state = IBLND_CONN_INIT;
 807         conn->ibc_version = version;
 808         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 809         cmid->context = conn;                   /* for future CM callbacks */
 810         conn->ibc_cmid = cmid;
 811         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 812         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 813         conn->ibc_rxs = NULL;
 814         conn->ibc_rx_pages = NULL;
 815
 816         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 817         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 818         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 819         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 820         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 821         INIT_LIST_HEAD(&conn->ibc_active_txs);
 822         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 823         spin_lock_init(&conn->ibc_lock);
 824
 825         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 826                          sizeof(*conn->ibc_connvars));
 827         if (conn->ibc_connvars == NULL) {
 828                 CERROR("Can't allocate in-progress connection state\n");
 829                 goto failed_2;
 830         }
 831
 832         write_lock_irqsave(glock, flags);
 833         if (dev->ibd_failover) {
 834                 write_unlock_irqrestore(glock, flags);
 835                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 836                 goto failed_2;
 837         }
 838
 839         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 840                 /* wakeup failover thread and teardown connection */
 841                 if (kiblnd_dev_can_failover(dev)) {
 842                         list_add_tail(&dev->ibd_fail_list,
 843                                       &kiblnd_data.kib_failed_devs);
 844                         wake_up(&kiblnd_data.kib_failover_waitq);
 845                 }
 846
 847                 write_unlock_irqrestore(glock, flags);
 848                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 849                        cmid->device->name, dev->ibd_ifname);
 850                 goto failed_2;
 851         }
 852
 853         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 854         conn->ibc_hdev = dev->ibd_hdev;
 855
 856         kiblnd_setup_mtu_locked(cmid);
 857
 858         write_unlock_irqrestore(glock, flags);
 859
 860 #ifdef HAVE_OFED_IB_CQ_INIT_ATTR
 861         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 862         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 863         cq = ib_create_cq(cmid->device,
 864                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 865                           &cq_attr);
 866 #else
 867         cq = ib_create_cq(cmid->device,
 868                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 869                           IBLND_CQ_ENTRIES(conn),
 870                           kiblnd_get_completion_vector(conn, cpt));
 871 #endif
 872         if (IS_ERR(cq)) {
 873                 /*
 874                  * on MLX-5 (possibly MLX-4 as well) this error could be
 875                  * hit if the concurrent_sends and/or peer_tx_credits is set
 876                  * too high. Or due to an MLX-5 bug which tries to
 877                  * allocate 256kb via kmalloc for WR cookie array
 878                  */
 879                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 880                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 881                 goto failed_2;
 882         }
 883
 884         conn->ibc_cq = cq;
 885
 886         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 887         if (rc != 0) {
 888                 CERROR("Can't request completion notification: %d\n", rc);
 889                 goto failed_2;
 890         }
 891
 892         init_qp_attr.event_handler = kiblnd_qp_event;
 893         init_qp_attr.qp_context = conn;
 894         init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 895         init_qp_attr.cap.max_recv_sge = 1;
 896         init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 897         init_qp_attr.qp_type = IB_QPT_RC;
 898         init_qp_attr.send_cq = cq;
 899         init_qp_attr.recv_cq = cq;
 900
 901         if (peer_ni->ibp_queue_depth_mod &&
 902             peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
 903                 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
 904                 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
 905                        peer_ni->ibp_queue_depth_mod,
 906                        peer_ni->ibp_queue_depth);
 907         }
 908
 909         do {
 910                 /* kiblnd_send_wrs() can change the connection's queue depth if
 911                  * the maximum work requests for the device is maxed out
 912                  */
 913                 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
 914                 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
 915                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
 916                                     &init_qp_attr);
 917                 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
 918                         break;
 919                 conn->ibc_queue_depth--;
 920         } while (rc);
 921
 922         if (rc) {
 923                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 924                        "send_sge: %d, recv_sge: %d\n",
 925                        rc, init_qp_attr.cap.max_send_wr,
 926                        init_qp_attr.cap.max_recv_wr,
 927                        init_qp_attr.cap.max_send_sge,
 928                        init_qp_attr.cap.max_recv_sge);
 929                 goto failed_2;
 930         }
 931
 932         conn->ibc_sched = sched;
 933
 934         if (!peer_ni->ibp_queue_depth_mod &&
 935             conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
 936                 CWARN("peer %s - queue depth reduced from %u to %u"
 937                       "  to allow for qp creation\n",
 938                       libcfs_nid2str(peer_ni->ibp_nid),
 939                       peer_ni->ibp_queue_depth,
 940                       conn->ibc_queue_depth);
 941                 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
 942         }
 943
 944         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 945                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 946         if (conn->ibc_rxs == NULL) {
 947                 CERROR("Cannot allocate RX buffers\n");
 948                 goto failed_2;
 949         }
 950
 951         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 952                                 IBLND_RX_MSG_PAGES(conn));
 953         if (rc != 0)
 954                 goto failed_2;
 955
 956         kiblnd_map_rx_descs(conn);
 957
 958         /* 1 ref for caller and each rxmsg */
 959         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 960         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 961
 962         /* post receives */
 963         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 964                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 965                 if (rc != 0) {
 966                         CERROR("Can't post rxmsg: %d\n", rc);
 967
 968                         /* Make posted receives complete */
 969                         kiblnd_abort_receives(conn);
 970
 971                         /* correct # of posted buffers
 972                          * NB locking needed now I'm racing with completion */
 973                         spin_lock_irqsave(&sched->ibs_lock, flags);
 974                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 975                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 976
 977                         /* cmid will be destroyed by CM(ofed) after cm_callback
 978                          * returned, so we can't refer it anymore
 979                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 980                         rdma_destroy_qp(conn->ibc_cmid);
 981                         conn->ibc_cmid = NULL;
 982
 983                         /* Drop my own and unused rxbuffer refcounts */
 984                         while (i++ <= IBLND_RX_MSGS(conn))
 985                                 kiblnd_conn_decref(conn);
 986
 987                         return NULL;
 988                 }
 989         }
 990
 991         /* Init successful! */
 992         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 993                  state == IBLND_CONN_PASSIVE_WAIT);
 994         conn->ibc_state = state;
 995
 996         /* 1 more conn */
 997         atomic_inc(&peer_ni->ibp_nconns);
 998         atomic_inc(&net->ibn_nconns);
 999         return conn;
1000
1001  failed_2:
1002         kiblnd_destroy_conn(conn);
1003         LIBCFS_FREE(conn, sizeof(*conn));
1004  failed_0:
1005         return NULL;
1006 }
1007
1008 void
1009 kiblnd_destroy_conn(struct kib_conn *conn)
1010 {
1011         struct rdma_cm_id *cmid = conn->ibc_cmid;
1012         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1013
1014         LASSERT (!in_interrupt());
1015         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1016         LASSERT(list_empty(&conn->ibc_early_rxs));
1017         LASSERT(list_empty(&conn->ibc_tx_noops));
1018         LASSERT(list_empty(&conn->ibc_tx_queue));
1019         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1020         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1021         LASSERT(list_empty(&conn->ibc_active_txs));
1022         LASSERT (conn->ibc_noops_posted == 0);
1023         LASSERT (conn->ibc_nsends_posted == 0);
1024
1025         switch (conn->ibc_state) {
1026         default:
1027                 /* conn must be completely disengaged from the network */
1028                 LBUG();
1029
1030         case IBLND_CONN_DISCONNECTED:
1031                 /* connvars should have been freed already */
1032                 LASSERT (conn->ibc_connvars == NULL);
1033                 break;
1034
1035         case IBLND_CONN_INIT:
1036                 break;
1037         }
1038
1039         /* conn->ibc_cmid might be destroyed by CM already */
1040         if (cmid != NULL && cmid->qp != NULL)
1041                 rdma_destroy_qp(cmid);
1042
1043         if (conn->ibc_cq)
1044                 ib_destroy_cq(conn->ibc_cq);
1045
1046         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1047                            LNET_MSG_STATUS_OK);
1048
1049         if (conn->ibc_rx_pages != NULL)
1050                 kiblnd_unmap_rx_descs(conn);
1051
1052         if (conn->ibc_rxs != NULL)
1053                 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1054
1055         if (conn->ibc_connvars != NULL)
1056                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1057
1058         if (conn->ibc_hdev != NULL)
1059                 kiblnd_hdev_decref(conn->ibc_hdev);
1060
1061         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1062         if (conn->ibc_state != IBLND_CONN_INIT) {
1063                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1064
1065                 kiblnd_peer_decref(peer_ni);
1066                 rdma_destroy_id(cmid);
1067                 atomic_dec(&peer_ni->ibp_nconns);
1068                 atomic_dec(&net->ibn_nconns);
1069         }
1070 }
1071
1072 int
1073 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1074 {
1075         struct kib_conn *conn;
1076         struct kib_conn *cnxt;
1077         int count = 0;
1078
1079         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1080                                  ibc_list) {
1081                 CDEBUG(D_NET, "Closing conn -> %s, "
1082                               "version: %x, reason: %d\n",
1083                        libcfs_nid2str(peer_ni->ibp_nid),
1084                        conn->ibc_version, why);
1085
1086                 kiblnd_close_conn_locked(conn, why);
1087                 count++;
1088         }
1089
1090         return count;
1091 }
1092
1093 int
1094 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1095                                 int version, __u64 incarnation)
1096 {
1097         struct kib_conn *conn;
1098         struct kib_conn *cnxt;
1099         int count = 0;
1100
1101         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1102                                  ibc_list) {
1103                 if (conn->ibc_version     == version &&
1104                     conn->ibc_incarnation == incarnation)
1105                         continue;
1106
1107                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1108                               "incarnation:%#llx(%x, %#llx)\n",
1109                        libcfs_nid2str(peer_ni->ibp_nid),
1110                        conn->ibc_version, conn->ibc_incarnation,
1111                        version, incarnation);
1112
1113                 kiblnd_close_conn_locked(conn, -ESTALE);
1114                 count++;
1115         }
1116
1117         return count;
1118 }
1119
1120 static int
1121 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1122 {
1123         struct kib_peer_ni *peer_ni;
1124         struct hlist_node *pnxt;
1125         int lo;
1126         int hi;
1127         int i;
1128         unsigned long flags;
1129         int count = 0;
1130
1131         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1132
1133         if (nid != LNET_NID_ANY) {
1134                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1135                 hi = lo;
1136         } else {
1137                 lo = 0;
1138                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1139         }
1140
1141         for (i = lo; i <= hi; i++) {
1142                 hlist_for_each_entry_safe(peer_ni, pnxt,
1143                                           &kiblnd_data.kib_peers[i], ibp_list) {
1144                         LASSERT(!kiblnd_peer_idle(peer_ni));
1145
1146                         if (peer_ni->ibp_ni != ni)
1147                                 continue;
1148
1149                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1150                                 continue;
1151
1152                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1153                 }
1154         }
1155
1156         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1157
1158         /* wildcards always succeed */
1159         if (nid == LNET_NID_ANY)
1160                 return 0;
1161
1162         return (count == 0) ? -ENOENT : 0;
1163 }
1164
1165 static int
1166 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1167 {
1168         struct libcfs_ioctl_data *data = arg;
1169         int                       rc = -EINVAL;
1170
1171         switch(cmd) {
1172         case IOC_LIBCFS_GET_PEER: {
1173                 lnet_nid_t   nid = 0;
1174                 int          count = 0;
1175
1176                 rc = kiblnd_get_peer_info(ni, data->ioc_nid, data->ioc_count,
1177                                           &nid, &count);
1178                 data->ioc_nid    = nid;
1179                 data->ioc_count  = count;
1180                 break;
1181         }
1182
1183         case IOC_LIBCFS_DEL_PEER: {
1184                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1185                 break;
1186         }
1187         case IOC_LIBCFS_GET_CONN: {
1188                 struct kib_conn *conn;
1189
1190                 rc = 0;
1191                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1192                 if (conn == NULL) {
1193                         rc = -ENOENT;
1194                         break;
1195                 }
1196
1197                 LASSERT(conn->ibc_cmid != NULL);
1198                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1199                 if (conn->ibc_cmid->route.path_rec == NULL)
1200                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1201                 else
1202                         data->ioc_u32[0] =
1203                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1204                 kiblnd_conn_decref(conn);
1205                 break;
1206         }
1207         case IOC_LIBCFS_CLOSE_CONNECTION: {
1208                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1209                 break;
1210         }
1211
1212         default:
1213                 break;
1214         }
1215
1216         return rc;
1217 }
1218
1219 static const struct ln_key_list kiblnd_tunables_keys = {
1220         .lkl_maxattr                    = LNET_NET_O2IBLND_TUNABLES_ATTR_MAX,
1221         .lkl_list                       = {
1222                 [LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS]  = {
1223                         .lkp_value      = "peercredits_hiw",
1224                         .lkp_data_type  = NLA_U32
1225                 },
1226                 [LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND]  = {
1227                         .lkp_value      = "map_on_demand",
1228                         .lkp_data_type  = NLA_FLAG
1229                 },
1230                 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS]  = {
1231                         .lkp_value      = "concurrent_sends",
1232                         .lkp_data_type  = NLA_U32
1233                 },
1234                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE]  = {
1235                         .lkp_value      = "fmr_pool_size",
1236                         .lkp_data_type  = NLA_U32
1237                 },
1238                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER]  = {
1239                         .lkp_value      = "fmr_flush_trigger",
1240                         .lkp_data_type  = NLA_U32
1241                 },
1242                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE]  = {
1243                         .lkp_value      = "fmr_cache",
1244                         .lkp_data_type  = NLA_U32
1245                 },
1246                 [LNET_NET_O2IBLND_TUNABLES_ATTR_NTX]  = {
1247                         .lkp_value      = "ntx",
1248                         .lkp_data_type  = NLA_U16
1249                 },
1250                 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER]  = {
1251                         .lkp_value      = "conns_per_peer",
1252                         .lkp_data_type  = NLA_U16
1253                 },
1254                 [LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT] = {
1255                         .lkp_value      = "timeout",
1256                         .lkp_data_type  = NLA_U32,
1257                 },
1258         },
1259 };
1260
1261 static int
1262 kiblnd_nl_get(int cmd, struct sk_buff *msg, int type, void *data)
1263 {
1264         struct lnet_ioctl_config_o2iblnd_tunables *tuns;
1265         struct lnet_ni *ni = data;
1266
1267         if (!ni || !msg)
1268                 return -EINVAL;
1269
1270         if (cmd != LNET_CMD_NETS || type != LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES)
1271                 return -EOPNOTSUPP;
1272
1273         tuns = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1274         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS,
1275                     tuns->lnd_peercredits_hiw);
1276         if (tuns->lnd_map_on_demand) {
1277                 nla_put_flag(msg,
1278                              LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND);
1279         }
1280         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS,
1281                     tuns->lnd_concurrent_sends);
1282         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE,
1283                     tuns->lnd_fmr_pool_size);
1284         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER,
1285                     tuns->lnd_fmr_flush_trigger);
1286         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE,
1287                     tuns->lnd_fmr_cache);
1288         nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_NTX, tuns->lnd_ntx);
1289         nla_put_u16(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER,
1290                     tuns->lnd_conns_per_peer);
1291         nla_put_u32(msg, LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT,
1292                     kiblnd_timeout());
1293
1294         return 0;
1295 }
1296
1297 static int
1298 kiblnd_nl_set(int cmd, struct nlattr *attr, int type, void *data)
1299 {
1300         struct lnet_lnd_tunables *tunables = data;
1301         int rc = 0;
1302         s64 num;
1303
1304         if (cmd != LNET_CMD_NETS)
1305                 return -EOPNOTSUPP;
1306
1307         if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
1308                 return -EINVAL;
1309
1310         switch (type) {
1311         case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1312                 tunables->lnd_tun_u.lnd_o2ib.lnd_peercredits_hiw = nla_get_s64(attr);
1313                 break;
1314         case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1315                 tunables->lnd_tun_u.lnd_o2ib.lnd_map_on_demand = nla_get_s64(attr);
1316                 break;
1317         case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1318                 tunables->lnd_tun_u.lnd_o2ib.lnd_concurrent_sends = nla_get_s64(attr);
1319                 break;
1320         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1321                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_pool_size = nla_get_s64(attr);
1322                 break;
1323         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1324                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_flush_trigger = nla_get_s64(attr);
1325                 break;
1326         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1327                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_cache = nla_get_s64(attr);
1328                 break;
1329         case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1330                 tunables->lnd_tun_u.lnd_o2ib.lnd_ntx = nla_get_s64(attr);
1331                 break;
1332         case LNET_NET_O2IBLND_TUNABLES_ATTR_LND_TIMEOUT:
1333                 tunables->lnd_tun_u.lnd_o2ib.lnd_timeout = nla_get_s64(attr);
1334                 break;
1335         case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1336                 num = nla_get_s64(attr);
1337                 if (num > -1 || num < 128)
1338                         tunables->lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = num;
1339                 else
1340                         rc = -ERANGE;
1341                 fallthrough;
1342         default:
1343                 break;
1344         }
1345
1346         return rc;
1347 }
1348
1349 static void
1350 kiblnd_free_pages(struct kib_pages *p)
1351 {
1352         int     npages = p->ibp_npages;
1353         int     i;
1354
1355         for (i = 0; i < npages; i++) {
1356                 if (p->ibp_pages[i] != NULL)
1357                         __free_page(p->ibp_pages[i]);
1358         }
1359
1360         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1361 }
1362
1363 int
1364 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1365 {
1366         struct kib_pages *p;
1367         int i;
1368
1369         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1370                          offsetof(struct kib_pages, ibp_pages[npages]));
1371         if (p == NULL) {
1372                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1373                 return -ENOMEM;
1374         }
1375
1376         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1377         p->ibp_npages = npages;
1378
1379         for (i = 0; i < npages; i++) {
1380                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1381                                                      GFP_NOFS);
1382                 if (p->ibp_pages[i] == NULL) {
1383                         CERROR("Can't allocate page %d of %d\n", i, npages);
1384                         kiblnd_free_pages(p);
1385                         return -ENOMEM;
1386                 }
1387         }
1388
1389         *pp = p;
1390         return 0;
1391 }
1392
1393 void
1394 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1395 {
1396         struct kib_rx *rx;
1397         int       i;
1398
1399         LASSERT (conn->ibc_rxs != NULL);
1400         LASSERT (conn->ibc_hdev != NULL);
1401
1402         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1403                 rx = &conn->ibc_rxs[i];
1404
1405                 LASSERT(rx->rx_nob >= 0); /* not posted */
1406
1407                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1408                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1409                                                           rx->rx_msgaddr),
1410                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1411         }
1412
1413         kiblnd_free_pages(conn->ibc_rx_pages);
1414
1415         conn->ibc_rx_pages = NULL;
1416 }
1417
1418 void
1419 kiblnd_map_rx_descs(struct kib_conn *conn)
1420 {
1421         struct kib_rx *rx;
1422         struct page    *pg;
1423         int             pg_off;
1424         int             ipg;
1425         int             i;
1426
1427         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1428                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1429                 rx = &conn->ibc_rxs[i];
1430
1431                 rx->rx_conn = conn;
1432                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1433
1434                 rx->rx_msgaddr =
1435                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1436                                               rx->rx_msg, IBLND_MSG_SIZE,
1437                                               DMA_FROM_DEVICE);
1438                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1439                                                   rx->rx_msgaddr));
1440                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1441
1442                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1443                        i, rx->rx_msg, rx->rx_msgaddr,
1444                        (__u64)(page_to_phys(pg) + pg_off));
1445
1446                 pg_off += IBLND_MSG_SIZE;
1447                 LASSERT(pg_off <= PAGE_SIZE);
1448
1449                 if (pg_off == PAGE_SIZE) {
1450                         pg_off = 0;
1451                         ipg++;
1452                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1453                 }
1454         }
1455 }
1456
1457 static void
1458 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1459 {
1460         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1461         struct kib_tx *tx;
1462         int i;
1463
1464         LASSERT (tpo->tpo_pool.po_allocated == 0);
1465
1466         if (hdev == NULL)
1467                 return;
1468
1469         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1470                 tx = &tpo->tpo_tx_descs[i];
1471                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1472                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1473                                                           tx->tx_msgaddr),
1474                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1475         }
1476
1477         kiblnd_hdev_decref(hdev);
1478         tpo->tpo_hdev = NULL;
1479 }
1480
1481 static struct kib_hca_dev *
1482 kiblnd_current_hdev(struct kib_dev *dev)
1483 {
1484         struct kib_hca_dev *hdev;
1485         unsigned long  flags;
1486         int            i = 0;
1487
1488         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1489         while (dev->ibd_failover) {
1490                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1491                 if (i++ % 50 == 0)
1492                         CDEBUG(D_NET, "%s: Wait for failover\n",
1493                                dev->ibd_ifname);
1494                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1495
1496                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1497         }
1498
1499         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1500         hdev = dev->ibd_hdev;
1501
1502         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1503
1504         return hdev;
1505 }
1506
1507 static void
1508 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1509 {
1510         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1511         struct kib_pool *pool = &tpo->tpo_pool;
1512         struct kib_net      *net   = pool->po_owner->ps_net;
1513         struct kib_dev *dev;
1514         struct page *page;
1515         struct kib_tx *tx;
1516         int             page_offset;
1517         int             ipage;
1518         int             i;
1519
1520         LASSERT (net != NULL);
1521
1522         dev = net->ibn_dev;
1523
1524         /* pre-mapped messages are not bigger than 1 page */
1525         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1526
1527         /* No fancy arithmetic when we do the buffer calculations */
1528         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1529
1530         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1531
1532         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1533                 page = txpgs->ibp_pages[ipage];
1534                 tx = &tpo->tpo_tx_descs[i];
1535
1536                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1537                                                 page_offset);
1538
1539                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1540                                                        tx->tx_msg,
1541                                                        IBLND_MSG_SIZE,
1542                                                        DMA_TO_DEVICE);
1543                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1544                                                   tx->tx_msgaddr));
1545                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1546
1547                 list_add(&tx->tx_list, &pool->po_free_list);
1548
1549                 page_offset += IBLND_MSG_SIZE;
1550                 LASSERT(page_offset <= PAGE_SIZE);
1551
1552                 if (page_offset == PAGE_SIZE) {
1553                         page_offset = 0;
1554                         ipage++;
1555                         LASSERT(ipage <= txpgs->ibp_npages);
1556                 }
1557         }
1558 }
1559
1560 static void
1561 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1562 {
1563         LASSERT(fpo->fpo_map_count == 0);
1564
1565 #ifdef HAVE_OFED_FMR_POOL_API
1566         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1567                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1568         } else
1569 #endif /* HAVE_OFED_FMR_POOL_API */
1570         {
1571                 struct kib_fast_reg_descriptor *frd, *tmp;
1572                 int i = 0;
1573
1574                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1575                                          frd_list) {
1576                         list_del(&frd->frd_list);
1577 #ifndef HAVE_OFED_IB_MAP_MR_SG
1578                         ib_free_fast_reg_page_list(frd->frd_frpl);
1579 #endif
1580                         ib_dereg_mr(frd->frd_mr);
1581                         LIBCFS_FREE(frd, sizeof(*frd));
1582                         i++;
1583                 }
1584                 if (i < fpo->fast_reg.fpo_pool_size)
1585                         CERROR("FastReg pool still has %d regions registered\n",
1586                                 fpo->fast_reg.fpo_pool_size - i);
1587         }
1588
1589         if (fpo->fpo_hdev)
1590                 kiblnd_hdev_decref(fpo->fpo_hdev);
1591
1592         LIBCFS_FREE(fpo, sizeof(*fpo));
1593 }
1594
1595 static void
1596 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1597 {
1598         struct kib_fmr_pool *fpo, *tmp;
1599
1600         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1601                 list_del(&fpo->fpo_list);
1602                 kiblnd_destroy_fmr_pool(fpo);
1603         }
1604 }
1605
1606 static int
1607 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1608                      int ncpts)
1609 {
1610         int size = tunables->lnd_fmr_pool_size / ncpts;
1611
1612         return max(IBLND_FMR_POOL, size);
1613 }
1614
1615 static int
1616 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1617                          int ncpts)
1618 {
1619         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1620
1621         return max(IBLND_FMR_POOL_FLUSH, size);
1622 }
1623
1624 #ifdef HAVE_OFED_FMR_POOL_API
1625 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1626                                  struct kib_fmr_pool *fpo)
1627 {
1628         struct ib_fmr_pool_param param = {
1629                 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1630                 .page_shift        = PAGE_SHIFT,
1631                 .access            = (IB_ACCESS_LOCAL_WRITE |
1632                                       IB_ACCESS_REMOTE_WRITE),
1633                 .pool_size         = fps->fps_pool_size,
1634                 .dirty_watermark   = fps->fps_flush_trigger,
1635                 .flush_function    = NULL,
1636                 .flush_arg         = NULL,
1637                 .cache             = !!fps->fps_cache };
1638         int rc = 0;
1639
1640         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1641                                                    &param);
1642         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1643                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1644                 if (rc != -ENOSYS)
1645                         CERROR("Failed to create FMR pool: %d\n", rc);
1646                 else
1647                         CERROR("FMRs are not supported\n");
1648         }
1649         fpo->fpo_is_fmr = true;
1650
1651         return rc;
1652 }
1653 #endif /* HAVE_OFED_FMR_POOL_API */
1654
1655 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1656                                   struct kib_fmr_pool *fpo,
1657                                   enum kib_dev_caps dev_caps)
1658 {
1659         struct kib_fast_reg_descriptor *frd, *tmp;
1660         int i, rc;
1661
1662 #ifdef HAVE_OFED_FMR_POOL_API
1663         fpo->fpo_is_fmr = false;
1664 #endif
1665
1666         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1667         fpo->fast_reg.fpo_pool_size = 0;
1668         for (i = 0; i < fps->fps_pool_size; i++) {
1669                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1670                                  sizeof(*frd));
1671                 if (!frd) {
1672                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1673                         rc = -ENOMEM;
1674                         goto out;
1675                 }
1676                 frd->frd_mr = NULL;
1677
1678 #ifndef HAVE_OFED_IB_MAP_MR_SG
1679                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1680                                                             IBLND_MAX_RDMA_FRAGS);
1681                 if (IS_ERR(frd->frd_frpl)) {
1682                         rc = PTR_ERR(frd->frd_frpl);
1683                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1684                                 rc);
1685                         frd->frd_frpl = NULL;
1686                         goto out_middle;
1687                 }
1688 #endif
1689
1690 #ifdef HAVE_OFED_IB_ALLOC_FAST_REG_MR
1691                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1692                                                    IBLND_MAX_RDMA_FRAGS);
1693 #else
1694                 /*
1695                  * it is expected to get here if this is an MLX-5 card.
1696                  * MLX-4 cards will always use FMR and MLX-5 cards will
1697                  * always use fast_reg. It turns out that some MLX-5 cards
1698                  * (possibly due to older FW versions) do not natively support
1699                  * gaps. So we will need to track them here.
1700                  */
1701                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1702 #ifdef IB_MR_TYPE_SG_GAPS
1703                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1704                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1705                                                 IB_MR_TYPE_SG_GAPS :
1706                                                 IB_MR_TYPE_MEM_REG,
1707 #else
1708                                                 IB_MR_TYPE_MEM_REG,
1709 #endif
1710                                           IBLND_MAX_RDMA_FRAGS);
1711                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1712                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1713                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1714 #endif
1715                 if (IS_ERR(frd->frd_mr)) {
1716                         rc = PTR_ERR(frd->frd_mr);
1717                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1718                         frd->frd_mr = NULL;
1719                         goto out_middle;
1720                 }
1721
1722                 /* indicate that the local invalidate needs to be generated */
1723                 frd->frd_valid = false;
1724
1725                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1726                 fpo->fast_reg.fpo_pool_size++;
1727         }
1728
1729         return 0;
1730
1731 out_middle:
1732         if (frd->frd_mr)
1733                 ib_dereg_mr(frd->frd_mr);
1734 #ifndef HAVE_OFED_IB_MAP_MR_SG
1735         if (frd->frd_frpl)
1736                 ib_free_fast_reg_page_list(frd->frd_frpl);
1737 #endif
1738         LIBCFS_FREE(frd, sizeof(*frd));
1739
1740 out:
1741         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1742                                  frd_list) {
1743                 list_del(&frd->frd_list);
1744 #ifndef HAVE_OFED_IB_MAP_MR_SG
1745                 ib_free_fast_reg_page_list(frd->frd_frpl);
1746 #endif
1747                 ib_dereg_mr(frd->frd_mr);
1748                 LIBCFS_FREE(frd, sizeof(*frd));
1749         }
1750
1751         return rc;
1752 }
1753
1754 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1755                                   struct kib_fmr_pool **pp_fpo)
1756 {
1757         struct kib_dev *dev = fps->fps_net->ibn_dev;
1758         struct kib_fmr_pool *fpo;
1759         int rc;
1760
1761         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1762         if (!fpo) {
1763                 return -ENOMEM;
1764         }
1765         memset(fpo, 0, sizeof(*fpo));
1766
1767         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1768
1769 #ifdef HAVE_OFED_FMR_POOL_API
1770         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1771                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1772         else
1773 #endif /* HAVE_OFED_FMR_POOL_API */
1774                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1775         if (rc)
1776                 goto out_fpo;
1777
1778         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1779         fpo->fpo_owner = fps;
1780         *pp_fpo = fpo;
1781
1782         return 0;
1783
1784 out_fpo:
1785         kiblnd_hdev_decref(fpo->fpo_hdev);
1786         LIBCFS_FREE(fpo, sizeof(*fpo));
1787         return rc;
1788 }
1789
1790 static void
1791 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1792 {
1793         struct kib_fmr_pool *fpo;
1794
1795         if (fps->fps_net == NULL) /* intialized? */
1796                 return;
1797
1798         spin_lock(&fps->fps_lock);
1799
1800         while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1801                                                struct kib_fmr_pool,
1802                                                fpo_list)) != NULL) {
1803                 fpo->fpo_failed = 1;
1804                 if (fpo->fpo_map_count == 0)
1805                         list_move(&fpo->fpo_list, zombies);
1806                 else
1807                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1808         }
1809
1810         spin_unlock(&fps->fps_lock);
1811 }
1812
1813 static void
1814 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1815 {
1816         if (fps->fps_net != NULL) { /* initialized? */
1817                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1818                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1819         }
1820 }
1821
1822 static int
1823 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1824                         struct kib_net *net,
1825                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1826 {
1827         struct kib_fmr_pool *fpo;
1828         int rc;
1829
1830         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1831
1832         fps->fps_net = net;
1833         fps->fps_cpt = cpt;
1834
1835         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1836         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1837         fps->fps_cache = tunables->lnd_fmr_cache;
1838
1839         spin_lock_init(&fps->fps_lock);
1840         INIT_LIST_HEAD(&fps->fps_pool_list);
1841         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1842
1843         rc = kiblnd_create_fmr_pool(fps, &fpo);
1844         if (rc == 0)
1845                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1846
1847         return rc;
1848 }
1849
1850 static int
1851 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1852 {
1853         if (fpo->fpo_map_count != 0) /* still in use */
1854                 return 0;
1855         if (fpo->fpo_failed)
1856                 return 1;
1857         return now >= fpo->fpo_deadline;
1858 }
1859
1860 #if defined(HAVE_OFED_FMR_POOL_API) || !defined(HAVE_OFED_IB_MAP_MR_SG)
1861 static int
1862 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1863 {
1864         struct kib_hca_dev *hdev;
1865         __u64           *pages = tx->tx_pages;
1866         int             npages;
1867         int             size;
1868         int             i;
1869
1870         hdev = tx->tx_pool->tpo_hdev;
1871
1872         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1873                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1874                         size += hdev->ibh_page_size) {
1875                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1876                                            hdev->ibh_page_mask) + size;
1877                 }
1878         }
1879
1880         return npages;
1881 }
1882 #endif
1883
1884 void
1885 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1886 {
1887         LIST_HEAD(zombies);
1888         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1889         struct kib_fmr_poolset *fps;
1890         time64_t now = ktime_get_seconds();
1891         struct kib_fmr_pool *tmp;
1892
1893         if (!fpo)
1894                 return;
1895
1896         fps = fpo->fpo_owner;
1897
1898 #ifdef HAVE_OFED_FMR_POOL_API
1899         if (fpo->fpo_is_fmr) {
1900                 if (fmr->fmr_pfmr) {
1901                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1902                         fmr->fmr_pfmr = NULL;
1903                 }
1904
1905                 if (status) {
1906                         int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1907                         LASSERT(!rc);
1908                 }
1909         } else
1910 #endif /* HAVE_OFED_FMR_POOL_API */
1911         {
1912                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1913                 if (frd) {
1914                         frd->frd_posted = false;
1915                         fmr->fmr_frd = NULL;
1916                         spin_lock(&fps->fps_lock);
1917                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1918                         spin_unlock(&fps->fps_lock);
1919                 }
1920         }
1921         fmr->fmr_pool = NULL;
1922
1923         spin_lock(&fps->fps_lock);
1924         fpo->fpo_map_count--;   /* decref the pool */
1925
1926         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1927                 /* the first pool is persistent */
1928                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1929                         continue;
1930
1931                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1932                         list_move(&fpo->fpo_list, &zombies);
1933                         fps->fps_version++;
1934                 }
1935         }
1936         spin_unlock(&fps->fps_lock);
1937
1938         if (!list_empty(&zombies))
1939                 kiblnd_destroy_fmr_pool_list(&zombies);
1940 }
1941
1942 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1943                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1944                         struct kib_fmr *fmr)
1945 {
1946         struct kib_fmr_pool *fpo;
1947         __u64 version;
1948         bool is_rx = (rd != tx->tx_rd);
1949 #ifdef HAVE_OFED_FMR_POOL_API
1950         __u64 *pages = tx->tx_pages;
1951         bool tx_pages_mapped = false;
1952         int npages = 0;
1953 #endif
1954         int rc;
1955
1956 again:
1957         spin_lock(&fps->fps_lock);
1958         version = fps->fps_version;
1959         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1960                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1961                 fpo->fpo_map_count++;
1962
1963 #ifdef HAVE_OFED_FMR_POOL_API
1964                 fmr->fmr_pfmr = NULL;
1965                 if (fpo->fpo_is_fmr) {
1966                         struct ib_pool_fmr *pfmr;
1967
1968                         spin_unlock(&fps->fps_lock);
1969
1970                         if (!tx_pages_mapped) {
1971                                 npages = kiblnd_map_tx_pages(tx, rd);
1972                                 tx_pages_mapped = true;
1973                         }
1974
1975                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1976                                                 pages, npages, iov);
1977                         if (likely(!IS_ERR(pfmr))) {
1978                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1979                                         : pfmr->fmr->lkey;
1980                                 fmr->fmr_frd  = NULL;
1981                                 fmr->fmr_pfmr = pfmr;
1982                                 fmr->fmr_pool = fpo;
1983                                 return 0;
1984                         }
1985                         rc = PTR_ERR(pfmr);
1986                 } else
1987 #endif /* HAVE_OFED_FMR_POOL_API */
1988                 {
1989                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1990                                 struct kib_fast_reg_descriptor *frd;
1991 #ifdef HAVE_OFED_IB_MAP_MR_SG
1992                                 struct ib_reg_wr *wr;
1993                                 int n;
1994 #else
1995                                 struct ib_rdma_wr *wr;
1996                                 struct ib_fast_reg_page_list *frpl;
1997 #endif
1998                                 struct ib_mr *mr;
1999
2000                                 frd = list_first_entry(
2001                                         &fpo->fast_reg.fpo_pool_list,
2002                                         struct kib_fast_reg_descriptor,
2003                                         frd_list);
2004                                 list_del(&frd->frd_list);
2005                                 spin_unlock(&fps->fps_lock);
2006
2007 #ifndef HAVE_OFED_IB_MAP_MR_SG
2008                                 frpl = frd->frd_frpl;
2009 #endif
2010                                 mr   = frd->frd_mr;
2011
2012                                 if (!frd->frd_valid) {
2013                                         struct ib_rdma_wr *inv_wr;
2014                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
2015
2016                                         frd->frd_valid = true;
2017                                         inv_wr = &frd->frd_inv_wr;
2018                                         memset(inv_wr, 0, sizeof(*inv_wr));
2019
2020                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
2021                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
2022                                         inv_wr->wr.ex.invalidate_rkey = key;
2023
2024                                         /* Bump the key */
2025                                         key = ib_inc_rkey(key);
2026                                         ib_update_fast_reg_key(mr, key);
2027                                 }
2028
2029 #ifdef HAVE_OFED_IB_MAP_MR_SG
2030 #ifdef HAVE_OFED_IB_MAP_MR_SG_5ARGS
2031                                 n = ib_map_mr_sg(mr, tx->tx_frags,
2032                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
2033 #else
2034                                 n = ib_map_mr_sg(mr, tx->tx_frags,
2035                                                  rd->rd_nfrags, PAGE_SIZE);
2036 #endif /* HAVE_OFED_IB_MAP_MR_SG_5ARGS */
2037                                 if (unlikely(n != rd->rd_nfrags)) {
2038                                         CERROR("Failed to map mr %d/%d elements\n",
2039                                                n, rd->rd_nfrags);
2040                                         return n < 0 ? n : -EINVAL;
2041                                 }
2042
2043                                 wr = &frd->frd_fastreg_wr;
2044                                 memset(wr, 0, sizeof(*wr));
2045
2046                                 wr->wr.opcode = IB_WR_REG_MR;
2047                                 wr->wr.wr_id  = IBLND_WID_MR;
2048                                 wr->wr.num_sge = 0;
2049                                 wr->wr.send_flags = 0;
2050                                 wr->mr = mr;
2051                                 wr->key = is_rx ? mr->rkey : mr->lkey;
2052                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
2053                                               IB_ACCESS_REMOTE_WRITE);
2054 #else /* HAVE_OFED_IB_MAP_MR_SG */
2055                                 if (!tx_pages_mapped) {
2056                                         npages = kiblnd_map_tx_pages(tx, rd);
2057                                         tx_pages_mapped = true;
2058                                 }
2059
2060                                 LASSERT(npages <= frpl->max_page_list_len);
2061                                 memcpy(frpl->page_list, pages,
2062                                        sizeof(*pages) * npages);
2063
2064                                 /* Prepare FastReg WR */
2065                                 wr = &frd->frd_fastreg_wr;
2066                                 memset(wr, 0, sizeof(*wr));
2067
2068                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
2069                                 wr->wr.wr_id  = IBLND_WID_MR;
2070
2071                                 wr->wr.wr.fast_reg.iova_start = iov;
2072                                 wr->wr.wr.fast_reg.page_list  = frpl;
2073                                 wr->wr.wr.fast_reg.page_list_len = npages;
2074                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2075                                 wr->wr.wr.fast_reg.length = nob;
2076                                 wr->wr.wr.fast_reg.rkey =
2077                                         is_rx ? mr->rkey : mr->lkey;
2078                                 wr->wr.wr.fast_reg.access_flags =
2079                                         (IB_ACCESS_LOCAL_WRITE |
2080                                          IB_ACCESS_REMOTE_WRITE);
2081 #endif /* HAVE_OFED_IB_MAP_MR_SG */
2082
2083                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
2084                                 fmr->fmr_frd  = frd;
2085                                 fmr->fmr_pool = fpo;
2086                                 frd->frd_posted = false;
2087                                 return 0;
2088                         }
2089                         spin_unlock(&fps->fps_lock);
2090                         rc = -EAGAIN;
2091                 }
2092
2093                 spin_lock(&fps->fps_lock);
2094                 fpo->fpo_map_count--;
2095                 if (rc != -EAGAIN) {
2096                         spin_unlock(&fps->fps_lock);
2097                         return rc;
2098                 }
2099
2100                 /* EAGAIN and ... */
2101                 if (version != fps->fps_version) {
2102                         spin_unlock(&fps->fps_lock);
2103                         goto again;
2104                 }
2105         }
2106
2107         if (fps->fps_increasing) {
2108                 spin_unlock(&fps->fps_lock);
2109                 CDEBUG(D_NET, "Another thread is allocating new "
2110                        "FMR pool, waiting for her to complete\n");
2111                 wait_var_event(fps, !fps->fps_increasing);
2112                 goto again;
2113
2114         }
2115
2116         if (ktime_get_seconds() < fps->fps_next_retry) {
2117                 /* someone failed recently */
2118                 spin_unlock(&fps->fps_lock);
2119                 return -EAGAIN;
2120         }
2121
2122         fps->fps_increasing = 1;
2123         spin_unlock(&fps->fps_lock);
2124
2125         CDEBUG(D_NET, "Allocate new FMR pool\n");
2126         rc = kiblnd_create_fmr_pool(fps, &fpo);
2127         spin_lock(&fps->fps_lock);
2128         fps->fps_increasing = 0;
2129         wake_up_var(fps);
2130         if (rc == 0) {
2131                 fps->fps_version++;
2132                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2133         } else {
2134                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2135         }
2136         spin_unlock(&fps->fps_lock);
2137
2138         goto again;
2139 }
2140
2141 static void
2142 kiblnd_fini_pool(struct kib_pool *pool)
2143 {
2144         LASSERT(list_empty(&pool->po_free_list));
2145         LASSERT(pool->po_allocated == 0);
2146
2147         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2148 }
2149
2150 static void
2151 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2152 {
2153         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2154
2155         memset(pool, 0, sizeof(struct kib_pool));
2156         INIT_LIST_HEAD(&pool->po_free_list);
2157         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2158         pool->po_owner = ps;
2159         pool->po_size = size;
2160 }
2161
2162 static void
2163 kiblnd_destroy_pool_list(struct list_head *head)
2164 {
2165         struct kib_pool *pool;
2166
2167         while ((pool = list_first_entry_or_null(head,
2168                                                 struct kib_pool,
2169                                                 po_list)) != NULL) {
2170                 list_del(&pool->po_list);
2171
2172                 LASSERT(pool->po_owner != NULL);
2173                 pool->po_owner->ps_pool_destroy(pool);
2174         }
2175 }
2176
2177 static void
2178 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2179 {
2180         struct kib_pool *po;
2181
2182         if (ps->ps_net == NULL) /* intialized? */
2183                 return;
2184
2185         spin_lock(&ps->ps_lock);
2186         while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2187                                               struct kib_pool,
2188                                               po_list)) != NULL) {
2189                 po->po_failed = 1;
2190                 if (po->po_allocated == 0)
2191                         list_move(&po->po_list, zombies);
2192                 else
2193                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2194         }
2195         spin_unlock(&ps->ps_lock);
2196 }
2197
2198 static void
2199 kiblnd_fini_poolset(struct kib_poolset *ps)
2200 {
2201         if (ps->ps_net != NULL) { /* initialized? */
2202                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2203                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2204         }
2205 }
2206
2207 static int
2208 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2209                     struct kib_net *net, char *name, int size,
2210                     kib_ps_pool_create_t po_create,
2211                     kib_ps_pool_destroy_t po_destroy,
2212                     kib_ps_node_init_t nd_init,
2213                     kib_ps_node_fini_t nd_fini)
2214 {
2215         struct kib_pool *pool;
2216         int rc;
2217
2218         memset(ps, 0, sizeof(struct kib_poolset));
2219
2220         ps->ps_cpt          = cpt;
2221         ps->ps_net          = net;
2222         ps->ps_pool_create  = po_create;
2223         ps->ps_pool_destroy = po_destroy;
2224         ps->ps_node_init    = nd_init;
2225         ps->ps_node_fini    = nd_fini;
2226         ps->ps_pool_size    = size;
2227         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2228             >= sizeof(ps->ps_name))
2229                 return -E2BIG;
2230         spin_lock_init(&ps->ps_lock);
2231         INIT_LIST_HEAD(&ps->ps_pool_list);
2232         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2233
2234         rc = ps->ps_pool_create(ps, size, &pool);
2235         if (rc == 0)
2236                 list_add(&pool->po_list, &ps->ps_pool_list);
2237         else
2238                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2239
2240         return rc;
2241 }
2242
2243 static int
2244 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2245 {
2246         if (pool->po_allocated != 0) /* still in use */
2247                 return 0;
2248         if (pool->po_failed)
2249                 return 1;
2250         return now >= pool->po_deadline;
2251 }
2252
2253 void
2254 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2255 {
2256         LIST_HEAD(zombies);
2257         struct kib_poolset *ps = pool->po_owner;
2258         struct kib_pool *tmp;
2259         time64_t now = ktime_get_seconds();
2260
2261         spin_lock(&ps->ps_lock);
2262
2263         if (ps->ps_node_fini != NULL)
2264                 ps->ps_node_fini(pool, node);
2265
2266         LASSERT(pool->po_allocated > 0);
2267         list_add(node, &pool->po_free_list);
2268         pool->po_allocated--;
2269
2270         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2271                 /* the first pool is persistent */
2272                 if (ps->ps_pool_list.next == &pool->po_list)
2273                         continue;
2274
2275                 if (kiblnd_pool_is_idle(pool, now))
2276                         list_move(&pool->po_list, &zombies);
2277         }
2278         spin_unlock(&ps->ps_lock);
2279
2280         if (!list_empty(&zombies))
2281                 kiblnd_destroy_pool_list(&zombies);
2282 }
2283
2284 struct list_head *
2285 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2286 {
2287         struct list_head        *node;
2288         struct kib_pool *pool;
2289         int                     rc;
2290         unsigned int            interval = 1;
2291         ktime_t time_before;
2292         unsigned int trips = 0;
2293
2294 again:
2295         spin_lock(&ps->ps_lock);
2296         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2297                 if (list_empty(&pool->po_free_list))
2298                         continue;
2299
2300                 pool->po_allocated++;
2301                 pool->po_deadline = ktime_get_seconds() +
2302                                     IBLND_POOL_DEADLINE;
2303                 node = pool->po_free_list.next;
2304                 list_del(node);
2305
2306                 if (ps->ps_node_init != NULL) {
2307                         /* still hold the lock */
2308                         ps->ps_node_init(pool, node);
2309                 }
2310                 spin_unlock(&ps->ps_lock);
2311                 return node;
2312         }
2313
2314         /* no available tx pool and ... */
2315         if (ps->ps_increasing) {
2316                 /* another thread is allocating a new pool */
2317                 spin_unlock(&ps->ps_lock);
2318                 trips++;
2319                 CDEBUG(D_NET,
2320                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2321                        ps->ps_name, interval, trips);
2322
2323                 schedule_timeout_interruptible(interval);
2324                 if (interval < cfs_time_seconds(1))
2325                         interval *= 2;
2326
2327                 goto again;
2328         }
2329
2330         if (ktime_get_seconds() < ps->ps_next_retry) {
2331                 /* someone failed recently */
2332                 spin_unlock(&ps->ps_lock);
2333                 return NULL;
2334         }
2335
2336         ps->ps_increasing = 1;
2337         spin_unlock(&ps->ps_lock);
2338
2339         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2340         time_before = ktime_get();
2341         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2342         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2343                ktime_ms_delta(ktime_get(), time_before));
2344
2345         spin_lock(&ps->ps_lock);
2346         ps->ps_increasing = 0;
2347         if (rc == 0) {
2348                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2349         } else {
2350                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2351                 CERROR("Can't allocate new %s pool because out of memory\n",
2352                        ps->ps_name);
2353         }
2354         spin_unlock(&ps->ps_lock);
2355
2356         goto again;
2357 }
2358
2359 static void
2360 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2361 {
2362         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2363                                                tpo_pool);
2364         int i;
2365
2366         LASSERT (pool->po_allocated == 0);
2367
2368         if (tpo->tpo_tx_pages != NULL) {
2369                 kiblnd_unmap_tx_pool(tpo);
2370                 kiblnd_free_pages(tpo->tpo_tx_pages);
2371         }
2372
2373         if (tpo->tpo_tx_descs == NULL)
2374                 goto out;
2375
2376         for (i = 0; i < pool->po_size; i++) {
2377                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2378                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2379
2380                 list_del(&tx->tx_list);
2381                 if (tx->tx_pages != NULL)
2382                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2383                 if (tx->tx_frags != NULL)
2384                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2385                                            IBLND_MAX_RDMA_FRAGS);
2386                 if (tx->tx_wrq != NULL)
2387                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2388                                            IBLND_MAX_RDMA_FRAGS);
2389                 if (tx->tx_sge != NULL) {
2390                         /* +1 is for the lnet header/message itself */
2391                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2392                                            (IBLND_MAX_RDMA_FRAGS *
2393                                            wrq_sge + 1));
2394                 }
2395                 if (tx->tx_rd != NULL)
2396                         LIBCFS_FREE(tx->tx_rd,
2397                                     offsetof(struct kib_rdma_desc,
2398                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2399         }
2400
2401         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2402 out:
2403         kiblnd_fini_pool(pool);
2404         CFS_FREE_PTR(tpo);
2405 }
2406
2407 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2408 {
2409         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2410         int ntx;
2411
2412         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2413         ntx = tunables->lnd_ntx / ncpts;
2414
2415         return max(IBLND_TX_POOL, ntx);
2416 }
2417
2418 static int
2419 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2420 {
2421         int            i;
2422         int            npg;
2423         struct kib_pool *pool;
2424         struct kib_tx_pool *tpo;
2425
2426         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2427         if (tpo == NULL) {
2428                 CERROR("Failed to allocate TX pool\n");
2429                 return -ENOMEM;
2430         }
2431
2432         pool = &tpo->tpo_pool;
2433         kiblnd_init_pool(ps, pool, size);
2434         tpo->tpo_tx_descs = NULL;
2435         tpo->tpo_tx_pages = NULL;
2436
2437         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2438         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2439                 CERROR("Can't allocate tx pages: %d\n", npg);
2440                 CFS_FREE_PTR(tpo);
2441                 return -ENOMEM;
2442         }
2443
2444         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2445                          size * sizeof(struct kib_tx));
2446         if (tpo->tpo_tx_descs == NULL) {
2447                 CERROR("Can't allocate %d tx descriptors\n", size);
2448                 ps->ps_pool_destroy(pool);
2449                 return -ENOMEM;
2450         }
2451
2452         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2453
2454         for (i = 0; i < size; i++) {
2455                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2456                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2457
2458                 tx->tx_pool = tpo;
2459                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2460                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2461                                          lnet_cpt_table(), ps->ps_cpt,
2462                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2463                         if (tx->tx_pages == NULL)
2464                                 break;
2465                 }
2466
2467                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2468                                  IBLND_MAX_RDMA_FRAGS *
2469                                  sizeof(*tx->tx_frags));
2470                 if (tx->tx_frags == NULL)
2471                         break;
2472
2473                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2474
2475                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2476                                  IBLND_MAX_RDMA_FRAGS *
2477                                  sizeof(*tx->tx_wrq));
2478                 if (tx->tx_wrq == NULL)
2479                         break;
2480
2481                 /* +1 is for the lnet header/message itself */
2482                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2483                                  (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2484                                  sizeof(*tx->tx_sge));
2485                 if (tx->tx_sge == NULL)
2486                         break;
2487
2488                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2489                                  offsetof(struct kib_rdma_desc,
2490                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2491                 if (tx->tx_rd == NULL)
2492                         break;
2493         }
2494
2495         if (i == size) {
2496                 kiblnd_map_tx_pool(tpo);
2497                 *pp_po = pool;
2498                 return 0;
2499         }
2500
2501         ps->ps_pool_destroy(pool);
2502         return -ENOMEM;
2503 }
2504
2505 static void
2506 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2507 {
2508         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2509                                                   struct kib_tx_poolset,
2510                                                   tps_poolset);
2511         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2512
2513         tx->tx_cookie = tps->tps_next_tx_cookie++;
2514 }
2515
2516 static void
2517 kiblnd_net_fini_pools(struct kib_net *net)
2518 {
2519         int     i;
2520
2521         cfs_cpt_for_each(i, lnet_cpt_table()) {
2522                 struct kib_tx_poolset *tps;
2523                 struct kib_fmr_poolset *fps;
2524
2525                 if (net->ibn_tx_ps != NULL) {
2526                         tps = net->ibn_tx_ps[i];
2527                         kiblnd_fini_poolset(&tps->tps_poolset);
2528                 }
2529
2530                 if (net->ibn_fmr_ps != NULL) {
2531                         fps = net->ibn_fmr_ps[i];
2532                         kiblnd_fini_fmr_poolset(fps);
2533                 }
2534         }
2535
2536         if (net->ibn_tx_ps != NULL) {
2537                 cfs_percpt_free(net->ibn_tx_ps);
2538                 net->ibn_tx_ps = NULL;
2539         }
2540
2541         if (net->ibn_fmr_ps != NULL) {
2542                 cfs_percpt_free(net->ibn_fmr_ps);
2543                 net->ibn_fmr_ps = NULL;
2544         }
2545 }
2546
2547 static int
2548 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2549                       int ncpts)
2550 {
2551         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2552 #ifdef HAVE_OFED_IB_GET_DMA_MR
2553         unsigned long   flags;
2554 #endif
2555         int             cpt;
2556         int             rc;
2557         int             i;
2558
2559         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2560
2561 #ifdef HAVE_OFED_IB_GET_DMA_MR
2562         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2563         /*
2564          * if lnd_map_on_demand is zero then we have effectively disabled
2565          * FMR or FastReg and we're using global memory regions
2566          * exclusively.
2567          */
2568         if (!tunables->lnd_map_on_demand) {
2569                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2570                                            flags);
2571                 goto create_tx_pool;
2572         }
2573
2574         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2575 #endif
2576
2577         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2578                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2579                        tunables->lnd_fmr_pool_size,
2580                        tunables->lnd_ntx / 4);
2581                 rc = -EINVAL;
2582                 goto failed;
2583         }
2584
2585         /* TX pool must be created later than FMR, see LU-2268
2586          * for details */
2587         LASSERT(net->ibn_tx_ps == NULL);
2588
2589         /* premapping can fail if ibd_nmr > 1, so we always create
2590          * FMR pool and map-on-demand if premapping failed */
2591
2592         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2593                                            sizeof(struct kib_fmr_poolset));
2594         if (net->ibn_fmr_ps == NULL) {
2595                 CERROR("Failed to allocate FMR pool array\n");
2596                 rc = -ENOMEM;
2597                 goto failed;
2598         }
2599
2600         for (i = 0; i < ncpts; i++) {
2601                 cpt = (cpts == NULL) ? i : cpts[i];
2602                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2603                                              net, tunables);
2604                 if (rc != 0) {
2605                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2606                                cpt, rc);
2607                         goto failed;
2608                 }
2609         }
2610
2611         if (i > 0)
2612                 LASSERT(i == ncpts);
2613
2614 #ifdef HAVE_OFED_IB_GET_DMA_MR
2615  create_tx_pool:
2616 #endif
2617         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2618                                           sizeof(struct kib_tx_poolset));
2619         if (net->ibn_tx_ps == NULL) {
2620                 CERROR("Failed to allocate tx pool array\n");
2621                 rc = -ENOMEM;
2622                 goto failed;
2623         }
2624
2625         for (i = 0; i < ncpts; i++) {
2626                 cpt = (cpts == NULL) ? i : cpts[i];
2627                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2628                                          cpt, net, "TX",
2629                                          kiblnd_tx_pool_size(ni, ncpts),
2630                                          kiblnd_create_tx_pool,
2631                                          kiblnd_destroy_tx_pool,
2632                                          kiblnd_tx_init, NULL);
2633                 if (rc != 0) {
2634                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2635                                cpt, rc);
2636                         goto failed;
2637                 }
2638         }
2639
2640         return 0;
2641  failed:
2642         kiblnd_net_fini_pools(net);
2643         LASSERT(rc != 0);
2644         return rc;
2645 }
2646
2647 static int
2648 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2649 {
2650         struct ib_port_attr *port_attr;
2651         int rc;
2652         unsigned long flags;
2653         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2654
2655         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2656         if (port_attr == NULL) {
2657                 CDEBUG(D_NETERROR, "Out of memory\n");
2658                 return -ENOMEM;
2659         }
2660
2661         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2662
2663         write_lock_irqsave(g_lock, flags);
2664
2665         if (rc == 0)
2666                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2667                                  ? IBLND_DEV_PORT_ACTIVE
2668                                  : IBLND_DEV_PORT_DOWN;
2669
2670         write_unlock_irqrestore(g_lock, flags);
2671         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2672
2673         if (rc != 0) {
2674                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2675                 return rc;
2676         }
2677         return 0;
2678 }
2679
2680 static inline void
2681 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2682 {
2683         struct kib_net  *net;
2684         __u32 ni_state_before;
2685         bool update_ping_buf = false;
2686         struct lnet_ni *ni = NULL;
2687
2688         /* for health check */
2689         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2690                 ni = net->ibn_ni;
2691                 if (val)
2692                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2693                                         libcfs_nidstr(&ni->ni_nid));
2694                 ni_state_before = lnet_set_link_fatal_state(ni, val);
2695
2696                 if (!update_ping_buf &&
2697                     (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2698                     (val != ni_state_before) &&
2699                     (net->ibn_init == IBLND_INIT_ALL))
2700                         update_ping_buf = true;
2701         }
2702
2703         if (update_ping_buf)
2704                 lnet_mark_ping_buffer_for_update();
2705 }
2706
2707 static void
2708 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2709 {
2710         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2711         struct kib_hca_dev  *hdev;
2712         unsigned long flags;
2713
2714         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2715
2716         write_lock_irqsave(g_lock, flags);
2717
2718         switch (event->event) {
2719         case IB_EVENT_DEVICE_FATAL:
2720                 CDEBUG(D_NET, "IB device fatal\n");
2721                 hdev->ibh_state = IBLND_DEV_FATAL;
2722                 kiblnd_set_ni_fatal_on(hdev, 1);
2723                 break;
2724         case IB_EVENT_PORT_ACTIVE:
2725                 CDEBUG(D_NET, "IB port active\n");
2726                 if (event->element.port_num == hdev->ibh_port) {
2727                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2728                         kiblnd_set_ni_fatal_on(hdev, 0);
2729                 }
2730                 break;
2731         case IB_EVENT_PORT_ERR:
2732                 CDEBUG(D_NET, "IB port err\n");
2733                 if (event->element.port_num == hdev->ibh_port) {
2734                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2735                         kiblnd_set_ni_fatal_on(hdev, 1);
2736                 }
2737                 break;
2738         default:
2739                 break;
2740         }
2741         write_unlock_irqrestore(g_lock, flags);
2742 }
2743
2744 static int
2745 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2746 {
2747         struct ib_device_attr *dev_attr;
2748         int rc = 0;
2749         int rc2 = 0;
2750
2751         /* It's safe to assume a HCA can handle a page size
2752          * matching that of the native system */
2753         hdev->ibh_page_shift = PAGE_SHIFT;
2754         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2755         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2756
2757 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2758         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2759         if (dev_attr == NULL) {
2760                 CERROR("Out of memory\n");
2761                 return -ENOMEM;
2762         }
2763
2764         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2765         if (rc != 0) {
2766                 CERROR("Failed to query IB device: %d\n", rc);
2767                 goto out_clean_attr;
2768         }
2769 #else
2770         dev_attr = &hdev->ibh_ibdev->attrs;
2771 #endif
2772
2773         hdev->ibh_mr_size = dev_attr->max_mr_size;
2774         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2775
2776         /* Setup device Memory Registration capabilities */
2777 #ifdef HAVE_OFED_FMR_POOL_API
2778 #ifdef HAVE_OFED_IB_DEVICE_OPS
2779         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2780             hdev->ibh_ibdev->ops.dealloc_fmr &&
2781             hdev->ibh_ibdev->ops.map_phys_fmr &&
2782             hdev->ibh_ibdev->ops.unmap_fmr) {
2783 #else
2784         if (hdev->ibh_ibdev->alloc_fmr &&
2785             hdev->ibh_ibdev->dealloc_fmr &&
2786             hdev->ibh_ibdev->map_phys_fmr &&
2787             hdev->ibh_ibdev->unmap_fmr) {
2788 #endif
2789                 LCONSOLE_INFO("Using FMR for registration\n");
2790                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2791         } else
2792 #endif /* HAVE_OFED_FMR_POOL_API */
2793         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2794                 LCONSOLE_INFO("Using FastReg for registration\n");
2795                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2796 #ifndef HAVE_OFED_IB_ALLOC_FAST_REG_MR
2797 #ifdef IB_DEVICE_SG_GAPS_REG
2798                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2799                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2800 #endif
2801 #endif
2802         } else {
2803                 rc = -ENOSYS;
2804         }
2805
2806         rc2 = kiblnd_port_get_attr(hdev);
2807         if (rc2 != 0)
2808                 return rc2;
2809
2810         if (rc != 0)
2811                 rc = -EINVAL;
2812
2813 #ifndef HAVE_OFED_IB_DEVICE_ATTRS
2814 out_clean_attr:
2815         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2816 #endif
2817
2818         if (rc == -ENOSYS)
2819                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2820                        "register memory: %d\n", rc);
2821         else if (rc == -EINVAL)
2822                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2823         return rc;
2824 }
2825
2826 #ifdef HAVE_OFED_IB_GET_DMA_MR
2827 static void
2828 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2829 {
2830         if (hdev->ibh_mrs == NULL)
2831                 return;
2832
2833         ib_dereg_mr(hdev->ibh_mrs);
2834
2835         hdev->ibh_mrs = NULL;
2836 }
2837 #endif
2838
2839 void
2840 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2841 {
2842         if (hdev->ibh_event_handler.device != NULL)
2843                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2844
2845 #ifdef HAVE_OFED_IB_GET_DMA_MR
2846         kiblnd_hdev_cleanup_mrs(hdev);
2847 #endif
2848
2849         if (hdev->ibh_pd != NULL)
2850                 ib_dealloc_pd(hdev->ibh_pd);
2851
2852         if (hdev->ibh_cmid != NULL)
2853                 rdma_destroy_id(hdev->ibh_cmid);
2854
2855         LIBCFS_FREE(hdev, sizeof(*hdev));
2856 }
2857
2858 #ifdef HAVE_OFED_IB_GET_DMA_MR
2859 static int
2860 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2861 {
2862         struct ib_mr *mr;
2863         int           acflags = IB_ACCESS_LOCAL_WRITE |
2864                                 IB_ACCESS_REMOTE_WRITE;
2865
2866         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2867         if (IS_ERR(mr)) {
2868                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2869                 kiblnd_hdev_cleanup_mrs(hdev);
2870                 return PTR_ERR(mr);
2871         }
2872
2873         hdev->ibh_mrs = mr;
2874
2875         return 0;
2876 }
2877 #endif
2878
2879 static int
2880 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2881 {       /* DUMMY */
2882         return 0;
2883 }
2884
2885 static int
2886 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2887 {
2888         struct rdma_cm_id  *cmid;
2889         struct sockaddr_in  srcaddr;
2890         struct sockaddr_in  dstaddr;
2891         int                 rc;
2892
2893         if (dev->ibd_hdev == NULL || /* initializing */
2894             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2895             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2896                 return 1;
2897
2898         /* XXX: it's UGLY, but I don't have better way to find
2899          * ib-bonding HCA failover because:
2900          *
2901          * a. no reliable CM event for HCA failover...
2902          * b. no OFED API to get ib_device for current net_device...
2903          *
2904          * We have only two choices at this point:
2905          *
2906          * a. rdma_bind_addr(), it will conflict with listener cmid
2907          * b. rdma_resolve_addr() to zero addr */
2908         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2909                                      RDMA_PS_TCP, IB_QPT_RC);
2910         if (IS_ERR(cmid)) {
2911                 rc = PTR_ERR(cmid);
2912                 CERROR("Failed to create cmid for failover: %d\n", rc);
2913                 return rc;
2914         }
2915
2916         memset(&srcaddr, 0, sizeof(srcaddr));
2917         srcaddr.sin_family      = AF_INET;
2918         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2919
2920         memset(&dstaddr, 0, sizeof(dstaddr));
2921         dstaddr.sin_family = AF_INET;
2922         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2923                                (struct sockaddr *)&dstaddr, 1);
2924         if (rc != 0 || cmid->device == NULL) {
2925                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2926                        dev->ibd_ifname, &dev->ibd_ifip,
2927                        cmid->device, rc);
2928                 rdma_destroy_id(cmid);
2929                 return rc;
2930         }
2931
2932         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2933         rdma_destroy_id(cmid);
2934         return rc;
2935 }
2936
2937 int
2938 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2939 {
2940         LIST_HEAD(zombie_tpo);
2941         LIST_HEAD(zombie_ppo);
2942         LIST_HEAD(zombie_fpo);
2943         struct rdma_cm_id  *cmid  = NULL;
2944         struct kib_hca_dev *hdev  = NULL;
2945         struct kib_hca_dev *old;
2946         struct ib_pd       *pd;
2947         struct kib_net *net;
2948         struct sockaddr_in  addr;
2949         struct net_device *netdev;
2950         unsigned long       flags;
2951         int                 rc = 0;
2952         int                 i;
2953         bool                set_fatal = true;
2954
2955         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2956                 dev->ibd_can_failover ||
2957                 dev->ibd_hdev == NULL);
2958
2959         rc = kiblnd_dev_need_failover(dev, ns);
2960         if (rc <= 0)
2961                 goto out;
2962
2963         if (dev->ibd_hdev != NULL &&
2964             dev->ibd_hdev->ibh_cmid != NULL) {
2965                 /* XXX it's not good to close old listener at here,
2966                  * because we can fail to create new listener.
2967                  * But we have to close it now, otherwise rdma_bind_addr
2968                  * will return EADDRINUSE... How crap! */
2969                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2970
2971                 cmid = dev->ibd_hdev->ibh_cmid;
2972                 /* make next schedule of kiblnd_dev_need_failover()
2973                  * return 1 for me */
2974                 dev->ibd_hdev->ibh_cmid  = NULL;
2975                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2976
2977                 rdma_destroy_id(cmid);
2978         }
2979
2980         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2981                                      IB_QPT_RC);
2982         if (IS_ERR(cmid)) {
2983                 rc = PTR_ERR(cmid);
2984                 CERROR("Failed to create cmid for failover: %d\n", rc);
2985                 goto out;
2986         }
2987
2988         memset(&addr, 0, sizeof(addr));
2989         addr.sin_family      = AF_INET;
2990         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2991         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2992
2993         /* Bind to failover device or port */
2994         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2995         if (rc != 0 || cmid->device == NULL) {
2996                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2997                        dev->ibd_ifname, &dev->ibd_ifip,
2998                        cmid->device, rc);
2999                 if (!rc && !cmid->device)
3000                         set_fatal = false;
3001                 rdma_destroy_id(cmid);
3002                 goto out;
3003         }
3004
3005         LIBCFS_ALLOC(hdev, sizeof(*hdev));
3006         if (hdev == NULL) {
3007                 CERROR("Failed to allocate kib_hca_dev\n");
3008                 rdma_destroy_id(cmid);
3009                 rc = -ENOMEM;
3010                 goto out;
3011         }
3012
3013         atomic_set(&hdev->ibh_ref, 1);
3014         hdev->ibh_dev   = dev;
3015         hdev->ibh_cmid  = cmid;
3016         hdev->ibh_ibdev = cmid->device;
3017         hdev->ibh_port  = cmid->port_num;
3018
3019 #ifdef HAVE_OFED_IB_ALLOC_PD_2ARGS
3020         pd = ib_alloc_pd(cmid->device, 0);
3021 #else
3022         pd = ib_alloc_pd(cmid->device);
3023 #endif
3024         if (IS_ERR(pd)) {
3025                 rc = PTR_ERR(pd);
3026                 CERROR("Can't allocate PD: %d\n", rc);
3027                 goto out;
3028         }
3029
3030         hdev->ibh_pd = pd;
3031
3032         rc = rdma_listen(cmid, 0);
3033         if (rc != 0) {
3034                 CERROR("Can't start new listener: %d\n", rc);
3035                 goto out;
3036         }
3037
3038         rc = kiblnd_hdev_get_attr(hdev);
3039         if (rc != 0) {
3040                 CERROR("Can't get device attributes: %d\n", rc);
3041                 goto out;
3042         }
3043
3044 #ifdef HAVE_OFED_IB_GET_DMA_MR
3045         rc = kiblnd_hdev_setup_mrs(hdev);
3046         if (rc != 0) {
3047                 CERROR("Can't setup device: %d\n", rc);
3048                 goto out;
3049         }
3050 #endif
3051
3052         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
3053                                 hdev->ibh_ibdev, kiblnd_event_handler);
3054         ib_register_event_handler(&hdev->ibh_event_handler);
3055
3056         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3057
3058         old = dev->ibd_hdev;
3059         dev->ibd_hdev = hdev;   /* take over the refcount */
3060         hdev = old;
3061
3062         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
3063                 cfs_cpt_for_each(i, lnet_cpt_table()) {
3064                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
3065                                             &zombie_tpo);
3066
3067                         if (net->ibn_fmr_ps != NULL)
3068                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
3069                                                         &zombie_fpo);
3070                 }
3071         }
3072
3073         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3074  out:
3075         if (!list_empty(&zombie_tpo))
3076                 kiblnd_destroy_pool_list(&zombie_tpo);
3077         if (!list_empty(&zombie_ppo))
3078                 kiblnd_destroy_pool_list(&zombie_ppo);
3079         if (!list_empty(&zombie_fpo))
3080                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
3081         if (hdev != NULL)
3082                 kiblnd_hdev_decref(hdev);
3083
3084         if (rc != 0) {
3085                 dev->ibd_failed_failover++;
3086         } else {
3087                 dev->ibd_failed_failover = 0;
3088
3089                 if (set_fatal) {
3090                         rcu_read_lock();
3091                         netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
3092                         if (netdev && (lnet_get_link_status(netdev) == 1))
3093                                 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
3094                         rcu_read_unlock();
3095                 }
3096         }
3097
3098         return rc;
3099 }
3100
3101 void
3102 kiblnd_destroy_dev(struct kib_dev *dev)
3103 {
3104         LASSERT(dev->ibd_nnets == 0);
3105         LASSERT(list_empty(&dev->ibd_nets));
3106
3107         list_del(&dev->ibd_fail_list);
3108         list_del(&dev->ibd_list);
3109
3110         if (dev->ibd_hdev != NULL)
3111                 kiblnd_hdev_decref(dev->ibd_hdev);
3112
3113         LIBCFS_FREE(dev, sizeof(*dev));
3114 }
3115
3116 static struct kib_dev *
3117 kiblnd_dev_search(char *ifname)
3118 {
3119         struct kib_dev *alias = NULL;
3120         struct kib_dev *dev;
3121         char *colon;
3122         char *colon2;
3123
3124         colon = strchr(ifname, ':');
3125         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3126                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3127                         return dev;
3128
3129                 if (alias != NULL)
3130                         continue;
3131
3132                 colon2 = strchr(dev->ibd_ifname, ':');
3133                 if (colon != NULL)
3134                         *colon = 0;
3135                 if (colon2 != NULL)
3136                         *colon2 = 0;
3137
3138                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3139                         alias = dev;
3140
3141                 if (colon != NULL)
3142                         *colon = ':';
3143                 if (colon2 != NULL)
3144                         *colon2 = ':';
3145         }
3146         return alias;
3147 }
3148
3149 static int
3150 kiblnd_handle_link_state_change(struct net_device *dev,
3151                                 unsigned char operstate)
3152 {
3153         struct lnet_ni *ni = NULL;
3154         struct kib_dev *event_kibdev;
3155         struct kib_net *net;
3156         struct kib_net *cnxt;
3157         bool link_down = !(operstate == IF_OPER_UP);
3158         struct in_device *in_dev;
3159         bool found_ip = false;
3160         __u32 ni_state_before;
3161         bool update_ping_buf = false;
3162         int state;
3163         DECLARE_CONST_IN_IFADDR(ifa);
3164
3165         event_kibdev = kiblnd_dev_search(dev->name);
3166
3167         if (!event_kibdev)
3168                 goto out;
3169
3170         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3171                 found_ip = false;
3172                 ni = net->ibn_ni;
3173
3174                 in_dev = __in_dev_get_rtnl(dev);
3175                 if (!in_dev) {
3176                         CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3177                                dev->name);
3178                         ni_state_before = lnet_set_link_fatal_state(ni, 1);
3179                         goto ni_done;
3180                 }
3181                 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3182                         if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3183                                 found_ip = true;
3184                 }
3185                 endfor_ifa(in_dev);
3186
3187                 if (!found_ip) {
3188                         CDEBUG(D_NET, "Interface %s has no matching ip\n",
3189                                dev->name);
3190                         ni_state_before = lnet_set_link_fatal_state(ni, 1);
3191                         goto ni_done;
3192                 }
3193
3194                 if (link_down) {
3195                         ni_state_before = lnet_set_link_fatal_state(ni, 1);
3196                 } else {
3197                         state = (lnet_get_link_status(dev) == 0);
3198                         ni_state_before = lnet_set_link_fatal_state(ni,
3199                                                                     state);
3200                 }
3201 ni_done:
3202                 if (!update_ping_buf &&
3203                     (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3204                     (atomic_read(&ni->ni_fatal_error_on) != ni_state_before) &&
3205                     (net->ibn_init == IBLND_INIT_ALL))
3206                         update_ping_buf = true;
3207         }
3208
3209         if (update_ping_buf)
3210                 lnet_mark_ping_buffer_for_update();
3211 out:
3212         return 0;
3213 }
3214
3215 static int
3216 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3217 {
3218         struct kib_dev *event_kibdev;
3219         struct kib_net *net;
3220         struct kib_net *cnxt;
3221         struct net_device *event_netdev = ifa->ifa_dev->dev;
3222         __u32 ni_state_before;
3223         bool update_ping_buf = false;
3224         struct lnet_ni *ni = NULL;
3225         bool link_down;
3226
3227         event_kibdev = kiblnd_dev_search(event_netdev->name);
3228
3229         if (!event_kibdev)
3230                 goto out;
3231
3232         if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3233                 goto out;
3234
3235         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3236                                  ibn_list) {
3237                 ni = net->ibn_ni;
3238                 link_down = (event == NETDEV_DOWN);
3239                 ni_state_before = lnet_set_link_fatal_state(ni, link_down);
3240                 if (!update_ping_buf &&
3241                     (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
3242                     ((event == NETDEV_DOWN) != ni_state_before) &&
3243                     (net->ibn_init == IBLND_INIT_ALL))
3244                         update_ping_buf = true;
3245         }
3246
3247         if (update_ping_buf)
3248                 lnet_mark_ping_buffer_for_update();
3249 out:
3250         return 0;
3251 }
3252
3253
3254 /************************************
3255  * Net device notifier event handler
3256  ************************************/
3257 static int kiblnd_device_event(struct notifier_block *unused,
3258                                  unsigned long event, void *ptr)
3259 {
3260         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3261         unsigned char operstate;
3262
3263         operstate = dev->operstate;
3264
3265         CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3266                event, dev->name, dev->ifindex, operstate);
3267
3268         switch (event) {
3269         case NETDEV_UP:
3270         case NETDEV_DOWN:
3271         case NETDEV_CHANGE:
3272                 kiblnd_handle_link_state_change(dev, operstate);
3273                 break;
3274         }
3275
3276         return NOTIFY_OK;
3277 }
3278
3279 /************************************
3280  * Inetaddr notifier event handler
3281  ************************************/
3282 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3283                                  unsigned long event, void *ptr)
3284 {
3285         struct in_ifaddr *ifa = ptr;
3286
3287         CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3288                event, &ifa->ifa_address, &ifa->ifa_mask);
3289
3290         switch (event) {
3291         case NETDEV_UP:
3292         case NETDEV_DOWN:
3293         case NETDEV_CHANGE:
3294                 kiblnd_handle_inetaddr_change(ifa, event);
3295                 break;
3296
3297         }
3298         return NOTIFY_OK;
3299 }
3300
3301 static struct notifier_block kiblnd_dev_notifier_block = {
3302         .notifier_call = kiblnd_device_event,
3303 };
3304
3305 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3306         .notifier_call = kiblnd_inetaddr_event,
3307 };
3308
3309 static void
3310 kiblnd_base_shutdown(void)
3311 {
3312         struct kib_sched_info *sched;
3313         struct kib_peer_ni *peer_ni;
3314         int i;
3315
3316         LASSERT(list_empty(&kiblnd_data.kib_devs));
3317
3318         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3319                libcfs_kmem_read());
3320
3321         if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3322                 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3323                 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3324         }
3325
3326         switch (kiblnd_data.kib_init) {
3327         default:
3328                 LBUG();
3329
3330         case IBLND_INIT_ALL:
3331         case IBLND_INIT_DATA:
3332                 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3333                         LASSERT(0);
3334                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3335                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3336                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3337                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3338
3339                 /* flag threads to terminate; wake and wait for them to die */
3340                 kiblnd_data.kib_shutdown = 1;
3341
3342                 /* NB: we really want to stop scheduler threads net by net
3343                  * instead of the whole module, this should be improved
3344                  * with dynamic configuration LNet.
3345                  */
3346                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3347                         wake_up_all(&sched->ibs_waitq);
3348
3349                 wake_up(&kiblnd_data.kib_connd_waitq);
3350                 wake_up(&kiblnd_data.kib_failover_waitq);
3351
3352                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3353                                        !atomic_read(&kiblnd_data.kib_nthreads),
3354                                        "Waiting for %d threads to terminate\n",
3355                                        atomic_read(&kiblnd_data.kib_nthreads));
3356                 fallthrough;
3357
3358         case IBLND_INIT_NOTHING:
3359                 break;
3360         }
3361
3362         if (kiblnd_data.kib_scheds != NULL)
3363                 cfs_percpt_free(kiblnd_data.kib_scheds);
3364
3365         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3366                libcfs_kmem_read());
3367
3368         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3369         module_put(THIS_MODULE);
3370 }
3371
3372 static void
3373 kiblnd_shutdown(struct lnet_ni *ni)
3374 {
3375         struct kib_net *net = ni->ni_data;
3376         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3377         unsigned long     flags;
3378
3379         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3380
3381         if (net == NULL)
3382                 goto out;
3383
3384         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3385                libcfs_kmem_read());
3386
3387         write_lock_irqsave(g_lock, flags);
3388         net->ibn_shutdown = 1;
3389         write_unlock_irqrestore(g_lock, flags);
3390
3391         switch (net->ibn_init) {
3392         default:
3393                 LBUG();
3394
3395         case IBLND_INIT_ALL:
3396                 /* nuke all existing peers within this net */
3397                 kiblnd_del_peer(ni, LNET_NID_ANY);
3398
3399                 /* Wait for all peer_ni state to clean up */
3400                 wait_var_event_warning(&net->ibn_npeers,
3401                                        atomic_read(&net->ibn_npeers) == 0,
3402                                        "%s: waiting for %d peers to disconnect\n",
3403                                        libcfs_nidstr(&ni->ni_nid),
3404                                        atomic_read(&net->ibn_npeers));
3405
3406                 kiblnd_net_fini_pools(net);
3407
3408                 write_lock_irqsave(g_lock, flags);
3409                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3410                 net->ibn_dev->ibd_nnets--;
3411                 list_del(&net->ibn_list);
3412                 write_unlock_irqrestore(g_lock, flags);
3413
3414                 wake_up_all(&kiblnd_data.kib_connd_waitq);
3415                 wait_var_event_warning(&net->ibn_nconns,
3416                                        atomic_read(&net->ibn_nconns) == 0,
3417                                        "%s: waiting for %d conns to clean\n",
3418                                        libcfs_nidstr(&ni->ni_nid),
3419                                        atomic_read(&net->ibn_nconns));
3420                 fallthrough;
3421
3422         case IBLND_INIT_NOTHING:
3423                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3424
3425                 if (net->ibn_dev != NULL &&
3426                     net->ibn_dev->ibd_nnets == 0)
3427                         kiblnd_destroy_dev(net->ibn_dev);
3428
3429                 break;
3430         }
3431
3432         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3433                libcfs_kmem_read());
3434
3435         net->ibn_init = IBLND_INIT_NOTHING;
3436         ni->ni_data = NULL;
3437
3438         LIBCFS_FREE(net, sizeof(*net));
3439
3440 out:
3441         if (list_empty(&kiblnd_data.kib_devs))
3442                 kiblnd_base_shutdown();
3443 }
3444
3445 static int
3446 kiblnd_base_startup(struct net *ns)
3447 {
3448         struct kib_sched_info *sched;
3449         int rc;
3450         int i;
3451
3452         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3453
3454         if (!try_module_get(THIS_MODULE))
3455                 goto failed;
3456
3457         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3458
3459         rwlock_init(&kiblnd_data.kib_global_lock);
3460
3461         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3462         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3463
3464         hash_init(kiblnd_data.kib_peers);
3465
3466         spin_lock_init(&kiblnd_data.kib_connd_lock);
3467         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3468         INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3469         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3470         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3471         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3472
3473         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3474         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3475
3476         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3477                                                   sizeof(*sched));
3478         if (kiblnd_data.kib_scheds == NULL)
3479                 goto failed;
3480
3481         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3482                 int     nthrs;
3483
3484                 spin_lock_init(&sched->ibs_lock);
3485                 INIT_LIST_HEAD(&sched->ibs_conns);
3486                 init_waitqueue_head(&sched->ibs_waitq);
3487
3488                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3489                 if (*kiblnd_tunables.kib_nscheds > 0) {
3490                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3491                 } else {
3492                         /* max to half of CPUs, another half is reserved for
3493                          * upper layer modules */
3494                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3495                 }
3496
3497                 sched->ibs_nthreads_max = nthrs;
3498                 sched->ibs_cpt = i;
3499         }
3500
3501         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3502
3503         /* lists/ptrs/locks initialised */
3504         kiblnd_data.kib_init = IBLND_INIT_DATA;
3505         /*****************************************************/
3506
3507         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3508         if (rc != 0) {
3509                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3510                 goto failed;
3511         }
3512
3513         if (*kiblnd_tunables.kib_dev_failover != 0)
3514                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3515                                          "kiblnd_failover");
3516
3517         if (rc != 0) {
3518                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3519                 goto failed;
3520         }
3521
3522         register_netdevice_notifier(&kiblnd_dev_notifier_block);
3523         register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3524
3525         /* flag everything initialised */
3526         kiblnd_data.kib_init = IBLND_INIT_ALL;
3527         /*****************************************************/
3528
3529         return 0;
3530
3531  failed:
3532         kiblnd_base_shutdown();
3533         return -ENETDOWN;
3534 }
3535
3536 static int
3537 kiblnd_start_schedulers(struct kib_sched_info *sched)
3538 {
3539         int     rc = 0;
3540         int     nthrs;
3541         int     i;
3542
3543         if (sched->ibs_nthreads == 0) {
3544                 if (*kiblnd_tunables.kib_nscheds > 0) {
3545                         nthrs = sched->ibs_nthreads_max;
3546                 } else {
3547                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3548                                                sched->ibs_cpt);
3549                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3550                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3551                 }
3552         } else {
3553                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3554                 /* increase one thread if there is new interface */
3555                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3556         }
3557
3558         for (i = 0; i < nthrs; i++) {
3559                 long    id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3560
3561                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3562                                          "kiblnd_sd_%02ld_%02ld",
3563                                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3564                 if (rc == 0)
3565                         continue;
3566
3567                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3568                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3569                 break;
3570         }
3571
3572         sched->ibs_nthreads += i;
3573         return rc;
3574 }
3575
3576 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3577                                     int ncpts)
3578 {
3579         int     cpt;
3580         int     rc;
3581         int     i;
3582
3583         for (i = 0; i < ncpts; i++) {
3584                 struct kib_sched_info *sched;
3585
3586                 cpt = (cpts == NULL) ? i : cpts[i];
3587                 sched = kiblnd_data.kib_scheds[cpt];
3588
3589                 if (!newdev && sched->ibs_nthreads > 0)
3590                         continue;
3591
3592                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3593                 if (rc != 0) {
3594                         CERROR("Failed to start scheduler threads for %s\n",
3595                                dev->ibd_ifname);
3596                         return rc;
3597                 }
3598         }
3599         return 0;
3600 }
3601
3602 static int
3603 kiblnd_startup(struct lnet_ni *ni)
3604 {
3605         char *ifname = NULL;
3606         struct lnet_inetdev *ifaces = NULL;
3607         struct kib_dev *ibdev = NULL;
3608         struct kib_net *net = NULL;
3609         unsigned long flags;
3610         int rc;
3611         int i;
3612         bool newdev;
3613         struct net_device *netdev;
3614
3615         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3616
3617         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3618                 rc = kiblnd_base_startup(ni->ni_net_ns);
3619                 if (rc != 0)
3620                         return rc;
3621         }
3622
3623         LIBCFS_ALLOC(net, sizeof(*net));
3624         ni->ni_data = net;
3625         if (net == NULL) {
3626                 rc = -ENOMEM;
3627                 goto failed;
3628         }
3629
3630         net->ibn_ni = ni;
3631         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3632
3633         kiblnd_tunables_setup(ni);
3634
3635         /* Multi-Rail wants each secondary
3636          * IP to be treated as an unique 'struct ni' interface.
3637          */
3638         if (ni->ni_interface != NULL) {
3639                 /* Use the IPoIB interface specified in 'networks=' */
3640                 ifname = ni->ni_interface;
3641         } else {
3642                 ifname = *kiblnd_tunables.kib_default_ipif;
3643                 rc = libcfs_strnid(&ni->ni_nid, ifname);
3644                 if (rc < 0 || ni->ni_nid.nid_type != O2IBLND)
3645                         memset(&ni->ni_nid, 0, sizeof(ni->ni_nid));
3646         }
3647
3648         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3649                 CERROR("IPoIB interface name too long: %s\n", ifname);
3650                 rc = -E2BIG;
3651                 goto failed;
3652         }
3653
3654         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3655         if (rc < 0)
3656                 goto failed;
3657
3658         i = lnet_inet_select(ni, ifaces, rc);
3659         if (i < 0)
3660                 goto failed;
3661
3662         if (nid_addr_is_set(&ni->ni_nid)) {
3663                 strscpy(ifname, ifaces[i].li_name, sizeof(ifname));
3664         } else if (strcmp(ifname, ifaces[i].li_name) != 0) {
3665                 CERROR("ko2iblnd: No matching interfaces\n");
3666                 rc = -ENOENT;
3667                 goto failed;
3668         }
3669
3670         ibdev = kiblnd_dev_search(ifname);
3671         newdev = ibdev == NULL;
3672         /* hmm...create kib_dev even for alias */
3673         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3674                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3675                 if (!ibdev) {
3676                         rc = -ENOMEM;
3677                         goto failed;
3678                 }
3679
3680                 ibdev->ibd_ifip = ntohl(ifaces[i].li_ipaddr);
3681                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3682                         sizeof(ibdev->ibd_ifname));
3683                 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3684
3685                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3686                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3687                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3688
3689                 /* initialize the device */
3690                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3691                 if (rc) {
3692                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3693                                rc);
3694                         goto failed;
3695                 }
3696
3697                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3698         }
3699
3700         net->ibn_dev = ibdev;
3701         ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3702         if (!ni->ni_interface) {
3703                 rc = lnet_ni_add_interface(ni, ifaces[i].li_name);
3704                 if (rc < 0)
3705                         CWARN("ko2iblnd failed to allocate ni_interface\n");
3706         }
3707         ni->ni_dev_cpt = ifaces[i].li_cpt;
3708
3709         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3710         if (rc != 0)
3711                 goto failed;
3712
3713         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3714         if (rc != 0) {
3715                 CERROR("Failed to initialize NI pools: %d\n", rc);
3716                 goto failed;
3717         }
3718
3719         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3720         ibdev->ibd_nnets++;
3721         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3722         /* for health check */
3723         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3724                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3725
3726         rcu_read_lock();
3727         netdev = dev_get_by_name_rcu(ni->ni_net_ns, net->ibn_dev->ibd_ifname);
3728         if (netdev &&
3729             ((netdev->reg_state == NETREG_UNREGISTERING) ||
3730              (netdev->operstate != IF_OPER_UP) ||
3731             (lnet_get_link_status(netdev) == 0))) {
3732                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3733         }
3734         rcu_read_unlock();
3735
3736         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3737
3738         net->ibn_init = IBLND_INIT_ALL;
3739
3740         return 0;
3741
3742 failed:
3743         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3744                 kiblnd_destroy_dev(ibdev);
3745
3746         kfree(ifaces);
3747         kiblnd_shutdown(ni);
3748
3749         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3750                ifname ? ifname : "", rc);
3751
3752         return -ENETDOWN;
3753 }
3754
3755 static const struct lnet_lnd the_o2iblnd = {
3756         .lnd_type       = O2IBLND,
3757         .lnd_startup    = kiblnd_startup,
3758         .lnd_shutdown   = kiblnd_shutdown,
3759         .lnd_ctl        = kiblnd_ctl,
3760         .lnd_send       = kiblnd_send,
3761         .lnd_recv       = kiblnd_recv,
3762         .lnd_get_dev_prio = kiblnd_get_dev_prio,
3763         .lnd_nl_get     = kiblnd_nl_get,
3764         .lnd_nl_set     = kiblnd_nl_set,
3765         .lnd_keys       = &kiblnd_tunables_keys,
3766 };
3767
3768 static void ko2inlnd_assert_wire_constants(void)
3769 {
3770         BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3771         BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3772         BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3773         BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3774
3775         BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3776         BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3777         BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3778         BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3779         BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3780         BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3781         BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3782         BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3783         BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3784         BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3785
3786         BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3787         BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3788         BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3789         BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3790         BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3791         BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3792         BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3793         BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3794
3795         BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3796         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3797         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3798         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3799         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3800         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3801         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3802
3803         BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3804         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3805         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3806         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3807         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3808
3809         BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3810         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3811         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3812         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3813         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3814
3815         BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3816         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3817         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3818         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3819         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3820         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3821         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3822
3823         BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3824         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3825         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3826         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3827         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3828
3829         BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3830         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3831         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3832         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3833         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3834         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3835         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3836
3837         BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3838         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3839         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3840         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3841         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3842         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3843         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3844
3845         BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3846         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3847         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3848         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3849         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3850
3851         /* Checks for struct kib_msg */
3852         //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3853         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3854         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3855         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3856         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3857         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3858         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3859         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3860         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3861         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3862         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3863         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3864         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3865         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3866         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3867         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3868         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3869         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3870         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3871         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3872         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3873
3874         /* Connparams */
3875         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3876         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3877         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3878         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3879         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3880         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3881
3882         /* Immediate message */
3883         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3884         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3885         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3886         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3887
3888         /* PUT req message */
3889         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3890         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3891         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3892         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3893
3894         /* Put ACK */
3895         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3896         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3897         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3898         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3899         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3900         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3901
3902         /* GET message */
3903         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3904         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3905         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3906         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3907         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3908         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3909
3910         /* Completion message */
3911         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3912         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3913         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3914         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3915
3916         /* Sanity checks */
3917         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3918         BUILD_BUG_ON(offsetof(struct kib_msg,
3919                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3920                      IBLND_MSG_SIZE);
3921         BUILD_BUG_ON(offsetof(struct kib_msg,
3922                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3923                      IBLND_MSG_SIZE);
3924 }
3925
3926 static void __exit ko2iblnd_exit(void)
3927 {
3928         lnet_unregister_lnd(&the_o2iblnd);
3929 }
3930
3931 static int __init ko2iblnd_init(void)
3932 {
3933         int rc;
3934
3935         ko2inlnd_assert_wire_constants();
3936
3937         rc = kiblnd_tunables_init();
3938         if (rc != 0)
3939                 return rc;
3940
3941         rc = libcfs_setup();
3942         if (rc)
3943                 return rc;
3944
3945         lnet_register_lnd(&the_o2iblnd);
3946
3947         return 0;
3948 }
3949
3950 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3951 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3952 MODULE_VERSION("2.8.0");
3953 MODULE_LICENSE("GPL");
3954
3955 module_init(ko2iblnd_init);
3956 module_exit(ko2iblnd_exit);