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 static int
 428 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
 429                      lnet_nid_t *nidp, int *count)
 430 {
 431         struct kib_peer_ni              *peer_ni;
 432         int                      i;
 433         unsigned long            flags;
 434
 435         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 436
 437         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 438                 LASSERT(!kiblnd_peer_idle(peer_ni));
 439
 440                 if (peer_ni->ibp_ni != ni)
 441                         continue;
 442
 443                 if (index-- > 0)
 444                         continue;
 445
 446                 *nidp = peer_ni->ibp_nid;
 447                 *count = kref_read(&peer_ni->ibp_kref);
 448
 449                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 450                 return 0;
 451         }
 452
 453         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 454         return -ENOENT;
 455 }
 456
 457 static void
 458 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
 459 {
 460         struct kib_conn *cnxt;
 461         struct kib_conn *conn;
 462
 463         if (list_empty(&peer_ni->ibp_conns)) {
 464                 kiblnd_unlink_peer_locked(peer_ni);
 465         } else {
 466                 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 467                                          ibc_list)
 468                         kiblnd_close_conn_locked(conn, 0);
 469                 /* NB closing peer_ni's last conn unlinked it. */
 470         }
 471         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 472          * last ref on it. */
 473 }
 474
 475 static int
 476 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 477 {
 478         LIST_HEAD(zombies);
 479         struct hlist_node *pnxt;
 480         struct kib_peer_ni *peer_ni;
 481         int lo;
 482         int hi;
 483         int i;
 484         unsigned long flags;
 485         int rc = -ENOENT;
 486
 487         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 488
 489         if (nid != LNET_NID_ANY) {
 490                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
 491                 hi = lo;
 492         } else {
 493                 lo = 0;
 494                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
 495         }
 496
 497         for (i = lo; i <= hi; i++) {
 498                 hlist_for_each_entry_safe(peer_ni, pnxt,
 499                                           &kiblnd_data.kib_peers[i], ibp_list) {
 500                         LASSERT(!kiblnd_peer_idle(peer_ni));
 501
 502                         if (peer_ni->ibp_ni != ni)
 503                                 continue;
 504
 505                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 506                                 continue;
 507
 508                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 509                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 510
 511                                 list_splice_init(&peer_ni->ibp_tx_queue,
 512                                                  &zombies);
 513                         }
 514
 515                         kiblnd_del_peer_locked(peer_ni);
 516                         rc = 0;         /* matched something */
 517                 }
 518         }
 519
 520         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 521
 522         kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
 523
 524         return rc;
 525 }
 526
 527 static struct kib_conn *
 528 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 529 {
 530         struct kib_peer_ni *peer_ni;
 531         struct kib_conn *conn;
 532         int i;
 533         unsigned long flags;
 534
 535         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 536
 537         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 538                 LASSERT(!kiblnd_peer_idle(peer_ni));
 539
 540                 if (peer_ni->ibp_ni != ni)
 541                         continue;
 542
 543                 list_for_each_entry(conn, &peer_ni->ibp_conns,
 544                                     ibc_list) {
 545                         if (index-- > 0)
 546                                 continue;
 547
 548                         kiblnd_conn_addref(conn);
 549                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 550                                                flags);
 551                         return conn;
 552                 }
 553         }
 554
 555         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 556         return NULL;
 557 }
 558
 559 static void
 560 kiblnd_debug_rx(struct kib_rx *rx)
 561 {
 562         CDEBUG(D_CONSOLE, "      %p msg_type %x cred %d\n",
 563                rx, rx->rx_msg->ibm_type,
 564                rx->rx_msg->ibm_credits);
 565 }
 566
 567 static void
 568 kiblnd_debug_tx(struct kib_tx *tx)
 569 {
 570         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 571                "cookie %#llx msg %s%s type %x cred %d\n",
 572                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 573                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 574                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 575                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 576                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 577 }
 578
 579 void
 580 kiblnd_debug_conn(struct kib_conn *conn)
 581 {
 582         struct list_head        *tmp;
 583         int                     i;
 584
 585         spin_lock(&conn->ibc_lock);
 586
 587         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 588                atomic_read(&conn->ibc_refcount), conn,
 589                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 590         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 591                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 592                conn->ibc_nsends_posted, conn->ibc_credits,
 593                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 594         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 595
 596         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 597         list_for_each(tmp, &conn->ibc_early_rxs)
 598                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 599
 600         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 601         list_for_each(tmp, &conn->ibc_tx_noops)
 602                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 603
 604         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 605         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 606                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 607
 608         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 609         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 610                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 611
 612         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 613         list_for_each(tmp, &conn->ibc_tx_queue)
 614                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 615
 616         CDEBUG(D_CONSOLE, "   active_txs:\n");
 617         list_for_each(tmp, &conn->ibc_active_txs)
 618                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   rxs:\n");
 621         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 622                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 623
 624         spin_unlock(&conn->ibc_lock);
 625 }
 626
 627 static void
 628 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 629 {
 630         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 631         if (cmid->route.path_rec == NULL)
 632                 return;
 633
 634         if (*kiblnd_tunables.kib_ib_mtu)
 635                 cmid->route.path_rec->mtu =
 636                         ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
 637 }
 638
 639 static int
 640 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 641 {
 642         cpumask_var_t   *mask;
 643         int             vectors;
 644         int             off;
 645         int             i;
 646         lnet_nid_t      ibp_nid;
 647
 648         vectors = conn->ibc_cmid->device->num_comp_vectors;
 649         if (vectors <= 1)
 650                 return 0;
 651
 652         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 653
 654         /* hash NID to CPU id in this partition... when targeting a single peer
 655          * with multiple QPs, to engage more cores in CQ processing to a single
 656          * peer, use ibp_nconns to salt the value the comp_vector value
 657          */
 658         ibp_nid = conn->ibc_peer->ibp_nid +
 659                 atomic_read(&conn->ibc_peer->ibp_nconns);
 660         off = do_div(ibp_nid, cpumask_weight(*mask));
 661         for_each_cpu(i, *mask) {
 662                 if (off-- == 0)
 663                         return i % vectors;
 664         }
 665
 666         LBUG();
 667         return 1;
 668 }
 669
 670 /*
 671  * Get the scheduler bound to this CPT. If the scheduler has no
 672  * threads, which means that the CPT has no CPUs, then grab the
 673  * next scheduler that we can use.
 674  *
 675  * This case would be triggered if a NUMA node is configured with
 676  * no associated CPUs.
 677  */
 678 static struct kib_sched_info *
 679 kiblnd_get_scheduler(int cpt)
 680 {
 681         struct kib_sched_info *sched;
 682         int i;
 683
 684         sched = kiblnd_data.kib_scheds[cpt];
 685
 686         if (sched->ibs_nthreads > 0)
 687                 return sched;
 688
 689         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 690                 if (sched->ibs_nthreads > 0) {
 691                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 692                                         cpt, sched->ibs_cpt);
 693                         return sched;
 694                 }
 695         }
 696
 697         return NULL;
 698 }
 699
 700 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 701 {
 702         /*
 703          * One WR for the LNet message
 704          * And ibc_max_frags for the transfer WRs
 705          */
 706         int ret;
 707         int multiplier = 1 + conn->ibc_max_frags;
 708
 709         /* FastReg needs two extra WRs for map and invalidate */
 710         if (IS_FAST_REG_DEV(conn->ibc_hdev->ibh_dev))
 711                 multiplier += 2;
 712
 713         /* account for a maximum of ibc_queue_depth in-flight transfers */
 714         ret = multiplier * conn->ibc_queue_depth;
 715
 716         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 717                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 718                        "request size %d larger than maximum %d device "
 719                        "can handle\n", conn->ibc_queue_depth, ret,
 720                        conn->ibc_hdev->ibh_max_qp_wr);
 721                 conn->ibc_queue_depth =
 722                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 723         }
 724
 725         /* don't go beyond the maximum the device can handle */
 726         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 727 }
 728
 729 struct kib_conn *
 730 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 731                    int state, int version)
 732 {
 733         /* CAVEAT EMPTOR:
 734          * If the new conn is created successfully it takes over the caller's
 735          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 736          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 737          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 738          * to destroy 'cmid' here since I'm called from the CM which still has
 739          * its ref on 'cmid'). */
 740         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 741         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 742         struct kib_dev *dev;
 743         struct ib_qp_init_attr init_qp_attr = {};
 744         struct kib_sched_info   *sched;
 745 #ifdef HAVE_IB_CQ_INIT_ATTR
 746         struct ib_cq_init_attr  cq_attr = {};
 747 #endif
 748         struct kib_conn *conn;
 749         struct ib_cq            *cq;
 750         unsigned long           flags;
 751         int                     cpt;
 752         int                     rc;
 753         int                     i;
 754
 755         LASSERT(net != NULL);
 756         LASSERT(!in_interrupt());
 757
 758         dev = net->ibn_dev;
 759
 760         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 761         sched = kiblnd_get_scheduler(cpt);
 762
 763         if (sched == NULL) {
 764                 CERROR("no schedulers available. node is unhealthy\n");
 765                 goto failed_0;
 766         }
 767
 768         /*
 769          * The cpt might have changed if we ended up selecting a non cpt
 770          * native scheduler. So use the scheduler's cpt instead.
 771          */
 772         cpt = sched->ibs_cpt;
 773
 774         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 775         if (conn == NULL) {
 776                 CERROR("Can't allocate connection for %s\n",
 777                        libcfs_nid2str(peer_ni->ibp_nid));
 778                 goto failed_0;
 779         }
 780
 781         conn->ibc_state = IBLND_CONN_INIT;
 782         conn->ibc_version = version;
 783         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 784         cmid->context = conn;                   /* for future CM callbacks */
 785         conn->ibc_cmid = cmid;
 786         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 787         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 788         conn->ibc_rxs = NULL;
 789         conn->ibc_rx_pages = NULL;
 790
 791         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 792         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 793         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 794         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 795         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 796         INIT_LIST_HEAD(&conn->ibc_active_txs);
 797         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 798         spin_lock_init(&conn->ibc_lock);
 799
 800         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 801                          sizeof(*conn->ibc_connvars));
 802         if (conn->ibc_connvars == NULL) {
 803                 CERROR("Can't allocate in-progress connection state\n");
 804                 goto failed_2;
 805         }
 806
 807         write_lock_irqsave(glock, flags);
 808         if (dev->ibd_failover) {
 809                 write_unlock_irqrestore(glock, flags);
 810                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 811                 goto failed_2;
 812         }
 813
 814         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 815                 /* wakeup failover thread and teardown connection */
 816                 if (kiblnd_dev_can_failover(dev)) {
 817                         list_add_tail(&dev->ibd_fail_list,
 818                                       &kiblnd_data.kib_failed_devs);
 819                         wake_up(&kiblnd_data.kib_failover_waitq);
 820                 }
 821
 822                 write_unlock_irqrestore(glock, flags);
 823                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 824                        cmid->device->name, dev->ibd_ifname);
 825                 goto failed_2;
 826         }
 827
 828         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 829         conn->ibc_hdev = dev->ibd_hdev;
 830
 831         kiblnd_setup_mtu_locked(cmid);
 832
 833         write_unlock_irqrestore(glock, flags);
 834
 835 #ifdef HAVE_IB_CQ_INIT_ATTR
 836         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 837         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 838         cq = ib_create_cq(cmid->device,
 839                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 840                           &cq_attr);
 841 #else
 842         cq = ib_create_cq(cmid->device,
 843                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 844                           IBLND_CQ_ENTRIES(conn),
 845                           kiblnd_get_completion_vector(conn, cpt));
 846 #endif
 847         if (IS_ERR(cq)) {
 848                 /*
 849                  * on MLX-5 (possibly MLX-4 as well) this error could be
 850                  * hit if the concurrent_sends and/or peer_tx_credits is set
 851                  * too high. Or due to an MLX-5 bug which tries to
 852                  * allocate 256kb via kmalloc for WR cookie array
 853                  */
 854                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 855                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 856                 goto failed_2;
 857         }
 858
 859         conn->ibc_cq = cq;
 860
 861         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 862         if (rc != 0) {
 863                 CERROR("Can't request completion notification: %d\n", rc);
 864                 goto failed_2;
 865         }
 866
 867         init_qp_attr.event_handler = kiblnd_qp_event;
 868         init_qp_attr.qp_context = conn;
 869         init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 870         init_qp_attr.cap.max_recv_sge = 1;
 871         init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 872         init_qp_attr.qp_type = IB_QPT_RC;
 873         init_qp_attr.send_cq = cq;
 874         init_qp_attr.recv_cq = cq;
 875
 876         if (peer_ni->ibp_queue_depth_mod &&
 877             peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
 878                 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
 879                 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
 880                        peer_ni->ibp_queue_depth_mod,
 881                        peer_ni->ibp_queue_depth);
 882         }
 883
 884         do {
 885                 /* kiblnd_send_wrs() can change the connection's queue depth if
 886                  * the maximum work requests for the device is maxed out
 887                  */
 888                 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
 889                 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
 890                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
 891                                     &init_qp_attr);
 892                 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
 893                         break;
 894                 conn->ibc_queue_depth--;
 895         } while (rc);
 896
 897         if (rc) {
 898                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 899                        "send_sge: %d, recv_sge: %d\n",
 900                        rc, init_qp_attr.cap.max_send_wr,
 901                        init_qp_attr.cap.max_recv_wr,
 902                        init_qp_attr.cap.max_send_sge,
 903                        init_qp_attr.cap.max_recv_sge);
 904                 goto failed_2;
 905         }
 906
 907         conn->ibc_sched = sched;
 908
 909         if (!peer_ni->ibp_queue_depth_mod &&
 910             conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
 911                 CWARN("peer %s - queue depth reduced from %u to %u"
 912                       "  to allow for qp creation\n",
 913                       libcfs_nid2str(peer_ni->ibp_nid),
 914                       peer_ni->ibp_queue_depth,
 915                       conn->ibc_queue_depth);
 916                 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
 917         }
 918
 919         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 920                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 921         if (conn->ibc_rxs == NULL) {
 922                 CERROR("Cannot allocate RX buffers\n");
 923                 goto failed_2;
 924         }
 925
 926         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 927                                 IBLND_RX_MSG_PAGES(conn));
 928         if (rc != 0)
 929                 goto failed_2;
 930
 931         kiblnd_map_rx_descs(conn);
 932
 933         /* 1 ref for caller and each rxmsg */
 934         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 935         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 936
 937         /* post receives */
 938         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 939                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 940                 if (rc != 0) {
 941                         CERROR("Can't post rxmsg: %d\n", rc);
 942
 943                         /* Make posted receives complete */
 944                         kiblnd_abort_receives(conn);
 945
 946                         /* correct # of posted buffers
 947                          * NB locking needed now I'm racing with completion */
 948                         spin_lock_irqsave(&sched->ibs_lock, flags);
 949                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 950                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 951
 952                         /* cmid will be destroyed by CM(ofed) after cm_callback
 953                          * returned, so we can't refer it anymore
 954                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 955                         rdma_destroy_qp(conn->ibc_cmid);
 956                         conn->ibc_cmid = NULL;
 957
 958                         /* Drop my own and unused rxbuffer refcounts */
 959                         while (i++ <= IBLND_RX_MSGS(conn))
 960                                 kiblnd_conn_decref(conn);
 961
 962                         return NULL;
 963                 }
 964         }
 965
 966         /* Init successful! */
 967         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 968                  state == IBLND_CONN_PASSIVE_WAIT);
 969         conn->ibc_state = state;
 970
 971         /* 1 more conn */
 972         atomic_inc(&peer_ni->ibp_nconns);
 973         atomic_inc(&net->ibn_nconns);
 974         return conn;
 975
 976  failed_2:
 977         kiblnd_destroy_conn(conn);
 978         LIBCFS_FREE(conn, sizeof(*conn));
 979  failed_0:
 980         return NULL;
 981 }
 982
 983 void
 984 kiblnd_destroy_conn(struct kib_conn *conn)
 985 {
 986         struct rdma_cm_id *cmid = conn->ibc_cmid;
 987         struct kib_peer_ni *peer_ni = conn->ibc_peer;
 988
 989         LASSERT (!in_interrupt());
 990         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 991         LASSERT(list_empty(&conn->ibc_early_rxs));
 992         LASSERT(list_empty(&conn->ibc_tx_noops));
 993         LASSERT(list_empty(&conn->ibc_tx_queue));
 994         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 995         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 996         LASSERT(list_empty(&conn->ibc_active_txs));
 997         LASSERT (conn->ibc_noops_posted == 0);
 998         LASSERT (conn->ibc_nsends_posted == 0);
 999
1000         switch (conn->ibc_state) {
1001         default:
1002                 /* conn must be completely disengaged from the network */
1003                 LBUG();
1004
1005         case IBLND_CONN_DISCONNECTED:
1006                 /* connvars should have been freed already */
1007                 LASSERT (conn->ibc_connvars == NULL);
1008                 break;
1009
1010         case IBLND_CONN_INIT:
1011                 break;
1012         }
1013
1014         /* conn->ibc_cmid might be destroyed by CM already */
1015         if (cmid != NULL && cmid->qp != NULL)
1016                 rdma_destroy_qp(cmid);
1017
1018         if (conn->ibc_cq)
1019                 ib_destroy_cq(conn->ibc_cq);
1020
1021         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1022                            LNET_MSG_STATUS_OK);
1023
1024         if (conn->ibc_rx_pages != NULL)
1025                 kiblnd_unmap_rx_descs(conn);
1026
1027         if (conn->ibc_rxs != NULL)
1028                 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1029
1030         if (conn->ibc_connvars != NULL)
1031                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1032
1033         if (conn->ibc_hdev != NULL)
1034                 kiblnd_hdev_decref(conn->ibc_hdev);
1035
1036         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1037         if (conn->ibc_state != IBLND_CONN_INIT) {
1038                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1039
1040                 kiblnd_peer_decref(peer_ni);
1041                 rdma_destroy_id(cmid);
1042                 atomic_dec(&peer_ni->ibp_nconns);
1043                 atomic_dec(&net->ibn_nconns);
1044         }
1045 }
1046
1047 int
1048 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1049 {
1050         struct kib_conn *conn;
1051         struct kib_conn *cnxt;
1052         int count = 0;
1053
1054         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1055                                  ibc_list) {
1056                 CDEBUG(D_NET, "Closing conn -> %s, "
1057                               "version: %x, reason: %d\n",
1058                        libcfs_nid2str(peer_ni->ibp_nid),
1059                        conn->ibc_version, why);
1060
1061                 kiblnd_close_conn_locked(conn, why);
1062                 count++;
1063         }
1064
1065         return count;
1066 }
1067
1068 int
1069 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1070                                 int version, __u64 incarnation)
1071 {
1072         struct kib_conn *conn;
1073         struct kib_conn *cnxt;
1074         int count = 0;
1075
1076         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1077                                  ibc_list) {
1078                 if (conn->ibc_version     == version &&
1079                     conn->ibc_incarnation == incarnation)
1080                         continue;
1081
1082                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1083                               "incarnation:%#llx(%x, %#llx)\n",
1084                        libcfs_nid2str(peer_ni->ibp_nid),
1085                        conn->ibc_version, conn->ibc_incarnation,
1086                        version, incarnation);
1087
1088                 kiblnd_close_conn_locked(conn, -ESTALE);
1089                 count++;
1090         }
1091
1092         return count;
1093 }
1094
1095 static int
1096 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1097 {
1098         struct kib_peer_ni *peer_ni;
1099         struct hlist_node *pnxt;
1100         int lo;
1101         int hi;
1102         int i;
1103         unsigned long flags;
1104         int count = 0;
1105
1106         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1107
1108         if (nid != LNET_NID_ANY) {
1109                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1110                 hi = lo;
1111         } else {
1112                 lo = 0;
1113                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1114         }
1115
1116         for (i = lo; i <= hi; i++) {
1117                 hlist_for_each_entry_safe(peer_ni, pnxt,
1118                                           &kiblnd_data.kib_peers[i], ibp_list) {
1119                         LASSERT(!kiblnd_peer_idle(peer_ni));
1120
1121                         if (peer_ni->ibp_ni != ni)
1122                                 continue;
1123
1124                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1125                                 continue;
1126
1127                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1128                 }
1129         }
1130
1131         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1132
1133         /* wildcards always succeed */
1134         if (nid == LNET_NID_ANY)
1135                 return 0;
1136
1137         return (count == 0) ? -ENOENT : 0;
1138 }
1139
1140 static int
1141 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1142 {
1143         struct libcfs_ioctl_data *data = arg;
1144         int                       rc = -EINVAL;
1145
1146         switch(cmd) {
1147         case IOC_LIBCFS_GET_PEER: {
1148                 lnet_nid_t   nid = 0;
1149                 int          count = 0;
1150
1151                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1152                                           &nid, &count);
1153                 data->ioc_nid    = nid;
1154                 data->ioc_count  = count;
1155                 break;
1156         }
1157
1158         case IOC_LIBCFS_DEL_PEER: {
1159                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1160                 break;
1161         }
1162         case IOC_LIBCFS_GET_CONN: {
1163                 struct kib_conn *conn;
1164
1165                 rc = 0;
1166                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1167                 if (conn == NULL) {
1168                         rc = -ENOENT;
1169                         break;
1170                 }
1171
1172                 LASSERT(conn->ibc_cmid != NULL);
1173                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1174                 if (conn->ibc_cmid->route.path_rec == NULL)
1175                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1176                 else
1177                         data->ioc_u32[0] =
1178                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1179                 kiblnd_conn_decref(conn);
1180                 break;
1181         }
1182         case IOC_LIBCFS_CLOSE_CONNECTION: {
1183                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1184                 break;
1185         }
1186
1187         default:
1188                 break;
1189         }
1190
1191         return rc;
1192 }
1193
1194 static void
1195 kiblnd_free_pages(struct kib_pages *p)
1196 {
1197         int     npages = p->ibp_npages;
1198         int     i;
1199
1200         for (i = 0; i < npages; i++) {
1201                 if (p->ibp_pages[i] != NULL)
1202                         __free_page(p->ibp_pages[i]);
1203         }
1204
1205         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1206 }
1207
1208 int
1209 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1210 {
1211         struct kib_pages *p;
1212         int i;
1213
1214         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1215                          offsetof(struct kib_pages, ibp_pages[npages]));
1216         if (p == NULL) {
1217                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1218                 return -ENOMEM;
1219         }
1220
1221         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1222         p->ibp_npages = npages;
1223
1224         for (i = 0; i < npages; i++) {
1225                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1226                                                      GFP_NOFS);
1227                 if (p->ibp_pages[i] == NULL) {
1228                         CERROR("Can't allocate page %d of %d\n", i, npages);
1229                         kiblnd_free_pages(p);
1230                         return -ENOMEM;
1231                 }
1232         }
1233
1234         *pp = p;
1235         return 0;
1236 }
1237
1238 void
1239 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1240 {
1241         struct kib_rx *rx;
1242         int       i;
1243
1244         LASSERT (conn->ibc_rxs != NULL);
1245         LASSERT (conn->ibc_hdev != NULL);
1246
1247         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1248                 rx = &conn->ibc_rxs[i];
1249
1250                 LASSERT(rx->rx_nob >= 0); /* not posted */
1251
1252                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1253                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1254                                                           rx->rx_msgaddr),
1255                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1256         }
1257
1258         kiblnd_free_pages(conn->ibc_rx_pages);
1259
1260         conn->ibc_rx_pages = NULL;
1261 }
1262
1263 void
1264 kiblnd_map_rx_descs(struct kib_conn *conn)
1265 {
1266         struct kib_rx *rx;
1267         struct page    *pg;
1268         int             pg_off;
1269         int             ipg;
1270         int             i;
1271
1272         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1273                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1274                 rx = &conn->ibc_rxs[i];
1275
1276                 rx->rx_conn = conn;
1277                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1278
1279                 rx->rx_msgaddr =
1280                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1281                                               rx->rx_msg, IBLND_MSG_SIZE,
1282                                               DMA_FROM_DEVICE);
1283                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1284                                                   rx->rx_msgaddr));
1285                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1286
1287                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1288                        i, rx->rx_msg, rx->rx_msgaddr,
1289                        (__u64)(page_to_phys(pg) + pg_off));
1290
1291                 pg_off += IBLND_MSG_SIZE;
1292                 LASSERT(pg_off <= PAGE_SIZE);
1293
1294                 if (pg_off == PAGE_SIZE) {
1295                         pg_off = 0;
1296                         ipg++;
1297                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1298                 }
1299         }
1300 }
1301
1302 static void
1303 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1304 {
1305         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1306         struct kib_tx *tx;
1307         int i;
1308
1309         LASSERT (tpo->tpo_pool.po_allocated == 0);
1310
1311         if (hdev == NULL)
1312                 return;
1313
1314         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1315                 tx = &tpo->tpo_tx_descs[i];
1316                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1317                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1318                                                           tx->tx_msgaddr),
1319                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1320         }
1321
1322         kiblnd_hdev_decref(hdev);
1323         tpo->tpo_hdev = NULL;
1324 }
1325
1326 static struct kib_hca_dev *
1327 kiblnd_current_hdev(struct kib_dev *dev)
1328 {
1329         struct kib_hca_dev *hdev;
1330         unsigned long  flags;
1331         int            i = 0;
1332
1333         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1334         while (dev->ibd_failover) {
1335                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1336                 if (i++ % 50 == 0)
1337                         CDEBUG(D_NET, "%s: Wait for failover\n",
1338                                dev->ibd_ifname);
1339                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1340
1341                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1342         }
1343
1344         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1345         hdev = dev->ibd_hdev;
1346
1347         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1348
1349         return hdev;
1350 }
1351
1352 static void
1353 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1354 {
1355         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1356         struct kib_pool *pool = &tpo->tpo_pool;
1357         struct kib_net      *net   = pool->po_owner->ps_net;
1358         struct kib_dev *dev;
1359         struct page *page;
1360         struct kib_tx *tx;
1361         int             page_offset;
1362         int             ipage;
1363         int             i;
1364
1365         LASSERT (net != NULL);
1366
1367         dev = net->ibn_dev;
1368
1369         /* pre-mapped messages are not bigger than 1 page */
1370         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1371
1372         /* No fancy arithmetic when we do the buffer calculations */
1373         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1374
1375         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1376
1377         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1378                 page = txpgs->ibp_pages[ipage];
1379                 tx = &tpo->tpo_tx_descs[i];
1380
1381                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1382                                                 page_offset);
1383
1384                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1385                                                        tx->tx_msg,
1386                                                        IBLND_MSG_SIZE,
1387                                                        DMA_TO_DEVICE);
1388                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1389                                                   tx->tx_msgaddr));
1390                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1391
1392                 list_add(&tx->tx_list, &pool->po_free_list);
1393
1394                 page_offset += IBLND_MSG_SIZE;
1395                 LASSERT(page_offset <= PAGE_SIZE);
1396
1397                 if (page_offset == PAGE_SIZE) {
1398                         page_offset = 0;
1399                         ipage++;
1400                         LASSERT(ipage <= txpgs->ibp_npages);
1401                 }
1402         }
1403 }
1404
1405 static void
1406 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1407 {
1408         LASSERT(fpo->fpo_map_count == 0);
1409
1410 #ifdef HAVE_FMR_POOL_API
1411         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1412                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1413         } else
1414 #endif /* HAVE_FMR_POOL_API */
1415         {
1416                 struct kib_fast_reg_descriptor *frd, *tmp;
1417                 int i = 0;
1418
1419                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1420                                          frd_list) {
1421                         list_del(&frd->frd_list);
1422 #ifndef HAVE_IB_MAP_MR_SG
1423                         ib_free_fast_reg_page_list(frd->frd_frpl);
1424 #endif
1425                         ib_dereg_mr(frd->frd_mr);
1426                         LIBCFS_FREE(frd, sizeof(*frd));
1427                         i++;
1428                 }
1429                 if (i < fpo->fast_reg.fpo_pool_size)
1430                         CERROR("FastReg pool still has %d regions registered\n",
1431                                 fpo->fast_reg.fpo_pool_size - i);
1432         }
1433
1434         if (fpo->fpo_hdev)
1435                 kiblnd_hdev_decref(fpo->fpo_hdev);
1436
1437         LIBCFS_FREE(fpo, sizeof(*fpo));
1438 }
1439
1440 static void
1441 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1442 {
1443         struct kib_fmr_pool *fpo, *tmp;
1444
1445         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1446                 list_del(&fpo->fpo_list);
1447                 kiblnd_destroy_fmr_pool(fpo);
1448         }
1449 }
1450
1451 static int
1452 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1453                      int ncpts)
1454 {
1455         int size = tunables->lnd_fmr_pool_size / ncpts;
1456
1457         return max(IBLND_FMR_POOL, size);
1458 }
1459
1460 static int
1461 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1462                          int ncpts)
1463 {
1464         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1465
1466         return max(IBLND_FMR_POOL_FLUSH, size);
1467 }
1468
1469 #ifdef HAVE_FMR_POOL_API
1470 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1471                                  struct kib_fmr_pool *fpo)
1472 {
1473         struct ib_fmr_pool_param param = {
1474                 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1475                 .page_shift        = PAGE_SHIFT,
1476                 .access            = (IB_ACCESS_LOCAL_WRITE |
1477                                       IB_ACCESS_REMOTE_WRITE),
1478                 .pool_size         = fps->fps_pool_size,
1479                 .dirty_watermark   = fps->fps_flush_trigger,
1480                 .flush_function    = NULL,
1481                 .flush_arg         = NULL,
1482                 .cache             = !!fps->fps_cache };
1483         int rc = 0;
1484
1485         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1486                                                    &param);
1487         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1488                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1489                 if (rc != -ENOSYS)
1490                         CERROR("Failed to create FMR pool: %d\n", rc);
1491                 else
1492                         CERROR("FMRs are not supported\n");
1493         }
1494         fpo->fpo_is_fmr = true;
1495
1496         return rc;
1497 }
1498 #endif /* HAVE_FMR_POOL_API */
1499
1500 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1501                                   struct kib_fmr_pool *fpo,
1502                                   enum kib_dev_caps dev_caps)
1503 {
1504         struct kib_fast_reg_descriptor *frd, *tmp;
1505         int i, rc;
1506
1507 #ifdef HAVE_FMR_POOL_API
1508         fpo->fpo_is_fmr = false;
1509 #endif
1510
1511         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1512         fpo->fast_reg.fpo_pool_size = 0;
1513         for (i = 0; i < fps->fps_pool_size; i++) {
1514                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1515                                  sizeof(*frd));
1516                 if (!frd) {
1517                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1518                         rc = -ENOMEM;
1519                         goto out;
1520                 }
1521                 frd->frd_mr = NULL;
1522
1523 #ifndef HAVE_IB_MAP_MR_SG
1524                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1525                                                             IBLND_MAX_RDMA_FRAGS);
1526                 if (IS_ERR(frd->frd_frpl)) {
1527                         rc = PTR_ERR(frd->frd_frpl);
1528                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1529                                 rc);
1530                         frd->frd_frpl = NULL;
1531                         goto out_middle;
1532                 }
1533 #endif
1534
1535 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1536                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1537                                                    IBLND_MAX_RDMA_FRAGS);
1538 #else
1539                 /*
1540                  * it is expected to get here if this is an MLX-5 card.
1541                  * MLX-4 cards will always use FMR and MLX-5 cards will
1542                  * always use fast_reg. It turns out that some MLX-5 cards
1543                  * (possibly due to older FW versions) do not natively support
1544                  * gaps. So we will need to track them here.
1545                  */
1546                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1547 #ifdef IB_MR_TYPE_SG_GAPS
1548                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1549                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1550                                                 IB_MR_TYPE_SG_GAPS :
1551                                                 IB_MR_TYPE_MEM_REG,
1552 #else
1553                                                 IB_MR_TYPE_MEM_REG,
1554 #endif
1555                                           IBLND_MAX_RDMA_FRAGS);
1556                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1557                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1558                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1559 #endif
1560                 if (IS_ERR(frd->frd_mr)) {
1561                         rc = PTR_ERR(frd->frd_mr);
1562                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1563                         frd->frd_mr = NULL;
1564                         goto out_middle;
1565                 }
1566
1567                 /* There appears to be a bug in MLX5 code where you must
1568                  * invalidate the rkey of a new FastReg pool before first
1569                  * using it. Thus, I am marking the FRD invalid here. */
1570                 frd->frd_valid = false;
1571
1572                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1573                 fpo->fast_reg.fpo_pool_size++;
1574         }
1575
1576         return 0;
1577
1578 out_middle:
1579         if (frd->frd_mr)
1580                 ib_dereg_mr(frd->frd_mr);
1581 #ifndef HAVE_IB_MAP_MR_SG
1582         if (frd->frd_frpl)
1583                 ib_free_fast_reg_page_list(frd->frd_frpl);
1584 #endif
1585         LIBCFS_FREE(frd, sizeof(*frd));
1586
1587 out:
1588         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1589                                  frd_list) {
1590                 list_del(&frd->frd_list);
1591 #ifndef HAVE_IB_MAP_MR_SG
1592                 ib_free_fast_reg_page_list(frd->frd_frpl);
1593 #endif
1594                 ib_dereg_mr(frd->frd_mr);
1595                 LIBCFS_FREE(frd, sizeof(*frd));
1596         }
1597
1598         return rc;
1599 }
1600
1601 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1602                                   struct kib_fmr_pool **pp_fpo)
1603 {
1604         struct kib_dev *dev = fps->fps_net->ibn_dev;
1605         struct kib_fmr_pool *fpo;
1606         int rc;
1607
1608         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1609         if (!fpo) {
1610                 return -ENOMEM;
1611         }
1612         memset(fpo, 0, sizeof(*fpo));
1613
1614         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1615
1616 #ifdef HAVE_FMR_POOL_API
1617         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1618                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1619         else
1620 #endif /* HAVE_FMR_POOL_API */
1621                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1622         if (rc)
1623                 goto out_fpo;
1624
1625         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1626         fpo->fpo_owner = fps;
1627         *pp_fpo = fpo;
1628
1629         return 0;
1630
1631 out_fpo:
1632         kiblnd_hdev_decref(fpo->fpo_hdev);
1633         LIBCFS_FREE(fpo, sizeof(*fpo));
1634         return rc;
1635 }
1636
1637 static void
1638 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1639 {
1640         struct kib_fmr_pool *fpo;
1641
1642         if (fps->fps_net == NULL) /* intialized? */
1643                 return;
1644
1645         spin_lock(&fps->fps_lock);
1646
1647         while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1648                                                struct kib_fmr_pool,
1649                                                fpo_list)) != NULL) {
1650                 fpo->fpo_failed = 1;
1651                 if (fpo->fpo_map_count == 0)
1652                         list_move(&fpo->fpo_list, zombies);
1653                 else
1654                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1655         }
1656
1657         spin_unlock(&fps->fps_lock);
1658 }
1659
1660 static void
1661 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1662 {
1663         if (fps->fps_net != NULL) { /* initialized? */
1664                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1665                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1666         }
1667 }
1668
1669 static int
1670 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1671                         struct kib_net *net,
1672                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1673 {
1674         struct kib_fmr_pool *fpo;
1675         int rc;
1676
1677         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1678
1679         fps->fps_net = net;
1680         fps->fps_cpt = cpt;
1681
1682         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1683         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1684         fps->fps_cache = tunables->lnd_fmr_cache;
1685
1686         spin_lock_init(&fps->fps_lock);
1687         INIT_LIST_HEAD(&fps->fps_pool_list);
1688         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1689
1690         rc = kiblnd_create_fmr_pool(fps, &fpo);
1691         if (rc == 0)
1692                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1693
1694         return rc;
1695 }
1696
1697 static int
1698 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1699 {
1700         if (fpo->fpo_map_count != 0) /* still in use */
1701                 return 0;
1702         if (fpo->fpo_failed)
1703                 return 1;
1704         return now >= fpo->fpo_deadline;
1705 }
1706
1707 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1708 static int
1709 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1710 {
1711         struct kib_hca_dev *hdev;
1712         __u64           *pages = tx->tx_pages;
1713         int             npages;
1714         int             size;
1715         int             i;
1716
1717         hdev = tx->tx_pool->tpo_hdev;
1718
1719         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1720                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1721                         size += hdev->ibh_page_size) {
1722                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1723                                            hdev->ibh_page_mask) + size;
1724                 }
1725         }
1726
1727         return npages;
1728 }
1729 #endif
1730
1731 void
1732 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1733 {
1734         LIST_HEAD(zombies);
1735         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1736         struct kib_fmr_poolset *fps;
1737         time64_t now = ktime_get_seconds();
1738         struct kib_fmr_pool *tmp;
1739
1740         if (!fpo)
1741                 return;
1742
1743         fps = fpo->fpo_owner;
1744
1745 #ifdef HAVE_FMR_POOL_API
1746         if (fpo->fpo_is_fmr) {
1747                 if (fmr->fmr_pfmr) {
1748                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1749                         fmr->fmr_pfmr = NULL;
1750                 }
1751
1752                 if (status) {
1753                         int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1754                         LASSERT(!rc);
1755                 }
1756         } else
1757 #endif /* HAVE_FMR_POOL_API */
1758         {
1759                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1760
1761                 if (frd) {
1762                         frd->frd_valid = false;
1763                         frd->frd_posted = false;
1764                         fmr->fmr_frd = NULL;
1765                         spin_lock(&fps->fps_lock);
1766                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1767                         spin_unlock(&fps->fps_lock);
1768                 }
1769         }
1770         fmr->fmr_pool = NULL;
1771
1772         spin_lock(&fps->fps_lock);
1773         fpo->fpo_map_count--;   /* decref the pool */
1774
1775         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1776                 /* the first pool is persistent */
1777                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1778                         continue;
1779
1780                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1781                         list_move(&fpo->fpo_list, &zombies);
1782                         fps->fps_version++;
1783                 }
1784         }
1785         spin_unlock(&fps->fps_lock);
1786
1787         if (!list_empty(&zombies))
1788                 kiblnd_destroy_fmr_pool_list(&zombies);
1789 }
1790
1791 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1792                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1793                         struct kib_fmr *fmr)
1794 {
1795         struct kib_fmr_pool *fpo;
1796         __u64 version;
1797         bool is_rx = (rd != tx->tx_rd);
1798 #ifdef HAVE_FMR_POOL_API
1799         __u64 *pages = tx->tx_pages;
1800         bool tx_pages_mapped = false;
1801         int npages = 0;
1802 #endif
1803         int rc;
1804
1805 again:
1806         spin_lock(&fps->fps_lock);
1807         version = fps->fps_version;
1808         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1809                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1810                 fpo->fpo_map_count++;
1811
1812 #ifdef HAVE_FMR_POOL_API
1813                 fmr->fmr_pfmr = NULL;
1814                 if (fpo->fpo_is_fmr) {
1815                         struct ib_pool_fmr *pfmr;
1816
1817                         spin_unlock(&fps->fps_lock);
1818
1819                         if (!tx_pages_mapped) {
1820                                 npages = kiblnd_map_tx_pages(tx, rd);
1821                                 tx_pages_mapped = true;
1822                         }
1823
1824                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1825                                                 pages, npages, iov);
1826                         if (likely(!IS_ERR(pfmr))) {
1827                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1828                                         : pfmr->fmr->lkey;
1829                                 fmr->fmr_frd  = NULL;
1830                                 fmr->fmr_pfmr = pfmr;
1831                                 fmr->fmr_pool = fpo;
1832                                 return 0;
1833                         }
1834                         rc = PTR_ERR(pfmr);
1835                 } else
1836 #endif /* HAVE_FMR_POOL_API */
1837                 {
1838                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1839                                 struct kib_fast_reg_descriptor *frd;
1840 #ifdef HAVE_IB_MAP_MR_SG
1841                                 struct ib_reg_wr *wr;
1842                                 int n;
1843 #else
1844                                 struct ib_rdma_wr *wr;
1845                                 struct ib_fast_reg_page_list *frpl;
1846 #endif
1847                                 struct ib_mr *mr;
1848
1849                                 frd = list_first_entry(
1850                                         &fpo->fast_reg.fpo_pool_list,
1851                                         struct kib_fast_reg_descriptor,
1852                                         frd_list);
1853                                 list_del(&frd->frd_list);
1854                                 spin_unlock(&fps->fps_lock);
1855
1856 #ifndef HAVE_IB_MAP_MR_SG
1857                                 frpl = frd->frd_frpl;
1858 #endif
1859                                 mr   = frd->frd_mr;
1860
1861                                 if (!frd->frd_valid) {
1862                                         struct ib_rdma_wr *inv_wr;
1863                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1864
1865                                         inv_wr = &frd->frd_inv_wr;
1866                                         memset(inv_wr, 0, sizeof(*inv_wr));
1867
1868                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1869                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1870                                         inv_wr->wr.ex.invalidate_rkey = key;
1871
1872                                         /* Bump the key */
1873                                         key = ib_inc_rkey(key);
1874                                         ib_update_fast_reg_key(mr, key);
1875                                 }
1876
1877 #ifdef HAVE_IB_MAP_MR_SG
1878 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1879                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1880                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
1881 #else
1882                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1883                                                  rd->rd_nfrags, PAGE_SIZE);
1884 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1885                                 if (unlikely(n != rd->rd_nfrags)) {
1886                                         CERROR("Failed to map mr %d/%d elements\n",
1887                                                n, rd->rd_nfrags);
1888                                         return n < 0 ? n : -EINVAL;
1889                                 }
1890
1891                                 wr = &frd->frd_fastreg_wr;
1892                                 memset(wr, 0, sizeof(*wr));
1893
1894                                 wr->wr.opcode = IB_WR_REG_MR;
1895                                 wr->wr.wr_id  = IBLND_WID_MR;
1896                                 wr->wr.num_sge = 0;
1897                                 wr->wr.send_flags = 0;
1898                                 wr->mr = mr;
1899                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1900                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1901                                               IB_ACCESS_REMOTE_WRITE);
1902 #else /* HAVE_IB_MAP_MR_SG */
1903                                 if (!tx_pages_mapped) {
1904                                         npages = kiblnd_map_tx_pages(tx, rd);
1905                                         tx_pages_mapped = true;
1906                                 }
1907
1908                                 LASSERT(npages <= frpl->max_page_list_len);
1909                                 memcpy(frpl->page_list, pages,
1910                                        sizeof(*pages) * npages);
1911
1912                                 /* Prepare FastReg WR */
1913                                 wr = &frd->frd_fastreg_wr;
1914                                 memset(wr, 0, sizeof(*wr));
1915
1916                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1917                                 wr->wr.wr_id  = IBLND_WID_MR;
1918
1919                                 wr->wr.wr.fast_reg.iova_start = iov;
1920                                 wr->wr.wr.fast_reg.page_list  = frpl;
1921                                 wr->wr.wr.fast_reg.page_list_len = npages;
1922                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1923                                 wr->wr.wr.fast_reg.length = nob;
1924                                 wr->wr.wr.fast_reg.rkey =
1925                                         is_rx ? mr->rkey : mr->lkey;
1926                                 wr->wr.wr.fast_reg.access_flags =
1927                                         (IB_ACCESS_LOCAL_WRITE |
1928                                          IB_ACCESS_REMOTE_WRITE);
1929 #endif /* HAVE_IB_MAP_MR_SG */
1930
1931                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1932                                 fmr->fmr_frd  = frd;
1933                                 fmr->fmr_pool = fpo;
1934                                 frd->frd_posted = false;
1935                                 return 0;
1936                         }
1937                         spin_unlock(&fps->fps_lock);
1938                         rc = -EAGAIN;
1939                 }
1940
1941                 spin_lock(&fps->fps_lock);
1942                 fpo->fpo_map_count--;
1943                 if (rc != -EAGAIN) {
1944                         spin_unlock(&fps->fps_lock);
1945                         return rc;
1946                 }
1947
1948                 /* EAGAIN and ... */
1949                 if (version != fps->fps_version) {
1950                         spin_unlock(&fps->fps_lock);
1951                         goto again;
1952                 }
1953         }
1954
1955         if (fps->fps_increasing) {
1956                 spin_unlock(&fps->fps_lock);
1957                 CDEBUG(D_NET, "Another thread is allocating new "
1958                        "FMR pool, waiting for her to complete\n");
1959                 wait_var_event(fps, !fps->fps_increasing);
1960                 goto again;
1961
1962         }
1963
1964         if (ktime_get_seconds() < fps->fps_next_retry) {
1965                 /* someone failed recently */
1966                 spin_unlock(&fps->fps_lock);
1967                 return -EAGAIN;
1968         }
1969
1970         fps->fps_increasing = 1;
1971         spin_unlock(&fps->fps_lock);
1972
1973         CDEBUG(D_NET, "Allocate new FMR pool\n");
1974         rc = kiblnd_create_fmr_pool(fps, &fpo);
1975         spin_lock(&fps->fps_lock);
1976         fps->fps_increasing = 0;
1977         wake_up_var(fps);
1978         if (rc == 0) {
1979                 fps->fps_version++;
1980                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1981         } else {
1982                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1983         }
1984         spin_unlock(&fps->fps_lock);
1985
1986         goto again;
1987 }
1988
1989 static void
1990 kiblnd_fini_pool(struct kib_pool *pool)
1991 {
1992         LASSERT(list_empty(&pool->po_free_list));
1993         LASSERT(pool->po_allocated == 0);
1994
1995         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1996 }
1997
1998 static void
1999 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2000 {
2001         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2002
2003         memset(pool, 0, sizeof(struct kib_pool));
2004         INIT_LIST_HEAD(&pool->po_free_list);
2005         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2006         pool->po_owner = ps;
2007         pool->po_size = size;
2008 }
2009
2010 static void
2011 kiblnd_destroy_pool_list(struct list_head *head)
2012 {
2013         struct kib_pool *pool;
2014
2015         while ((pool = list_first_entry_or_null(head,
2016                                                 struct kib_pool,
2017                                                 po_list)) != NULL) {
2018                 list_del(&pool->po_list);
2019
2020                 LASSERT(pool->po_owner != NULL);
2021                 pool->po_owner->ps_pool_destroy(pool);
2022         }
2023 }
2024
2025 static void
2026 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2027 {
2028         struct kib_pool *po;
2029
2030         if (ps->ps_net == NULL) /* intialized? */
2031                 return;
2032
2033         spin_lock(&ps->ps_lock);
2034         while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2035                                               struct kib_pool,
2036                                               po_list)) != NULL) {
2037                 po->po_failed = 1;
2038                 if (po->po_allocated == 0)
2039                         list_move(&po->po_list, zombies);
2040                 else
2041                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2042         }
2043         spin_unlock(&ps->ps_lock);
2044 }
2045
2046 static void
2047 kiblnd_fini_poolset(struct kib_poolset *ps)
2048 {
2049         if (ps->ps_net != NULL) { /* initialized? */
2050                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2051                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2052         }
2053 }
2054
2055 static int
2056 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2057                     struct kib_net *net, char *name, int size,
2058                     kib_ps_pool_create_t po_create,
2059                     kib_ps_pool_destroy_t po_destroy,
2060                     kib_ps_node_init_t nd_init,
2061                     kib_ps_node_fini_t nd_fini)
2062 {
2063         struct kib_pool *pool;
2064         int rc;
2065
2066         memset(ps, 0, sizeof(struct kib_poolset));
2067
2068         ps->ps_cpt          = cpt;
2069         ps->ps_net          = net;
2070         ps->ps_pool_create  = po_create;
2071         ps->ps_pool_destroy = po_destroy;
2072         ps->ps_node_init    = nd_init;
2073         ps->ps_node_fini    = nd_fini;
2074         ps->ps_pool_size    = size;
2075         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2076             >= sizeof(ps->ps_name))
2077                 return -E2BIG;
2078         spin_lock_init(&ps->ps_lock);
2079         INIT_LIST_HEAD(&ps->ps_pool_list);
2080         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2081
2082         rc = ps->ps_pool_create(ps, size, &pool);
2083         if (rc == 0)
2084                 list_add(&pool->po_list, &ps->ps_pool_list);
2085         else
2086                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2087
2088         return rc;
2089 }
2090
2091 static int
2092 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2093 {
2094         if (pool->po_allocated != 0) /* still in use */
2095                 return 0;
2096         if (pool->po_failed)
2097                 return 1;
2098         return now >= pool->po_deadline;
2099 }
2100
2101 void
2102 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2103 {
2104         LIST_HEAD(zombies);
2105         struct kib_poolset *ps = pool->po_owner;
2106         struct kib_pool *tmp;
2107         time64_t now = ktime_get_seconds();
2108
2109         spin_lock(&ps->ps_lock);
2110
2111         if (ps->ps_node_fini != NULL)
2112                 ps->ps_node_fini(pool, node);
2113
2114         LASSERT(pool->po_allocated > 0);
2115         list_add(node, &pool->po_free_list);
2116         pool->po_allocated--;
2117
2118         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2119                 /* the first pool is persistent */
2120                 if (ps->ps_pool_list.next == &pool->po_list)
2121                         continue;
2122
2123                 if (kiblnd_pool_is_idle(pool, now))
2124                         list_move(&pool->po_list, &zombies);
2125         }
2126         spin_unlock(&ps->ps_lock);
2127
2128         if (!list_empty(&zombies))
2129                 kiblnd_destroy_pool_list(&zombies);
2130 }
2131
2132 struct list_head *
2133 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2134 {
2135         struct list_head        *node;
2136         struct kib_pool *pool;
2137         int                     rc;
2138         unsigned int            interval = 1;
2139         ktime_t time_before;
2140         unsigned int trips = 0;
2141
2142 again:
2143         spin_lock(&ps->ps_lock);
2144         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2145                 if (list_empty(&pool->po_free_list))
2146                         continue;
2147
2148                 pool->po_allocated++;
2149                 pool->po_deadline = ktime_get_seconds() +
2150                                     IBLND_POOL_DEADLINE;
2151                 node = pool->po_free_list.next;
2152                 list_del(node);
2153
2154                 if (ps->ps_node_init != NULL) {
2155                         /* still hold the lock */
2156                         ps->ps_node_init(pool, node);
2157                 }
2158                 spin_unlock(&ps->ps_lock);
2159                 return node;
2160         }
2161
2162         /* no available tx pool and ... */
2163         if (ps->ps_increasing) {
2164                 /* another thread is allocating a new pool */
2165                 spin_unlock(&ps->ps_lock);
2166                 trips++;
2167                 CDEBUG(D_NET,
2168                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2169                        ps->ps_name, interval, trips);
2170
2171                 schedule_timeout_interruptible(interval);
2172                 if (interval < cfs_time_seconds(1))
2173                         interval *= 2;
2174
2175                 goto again;
2176         }
2177
2178         if (ktime_get_seconds() < ps->ps_next_retry) {
2179                 /* someone failed recently */
2180                 spin_unlock(&ps->ps_lock);
2181                 return NULL;
2182         }
2183
2184         ps->ps_increasing = 1;
2185         spin_unlock(&ps->ps_lock);
2186
2187         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2188         time_before = ktime_get();
2189         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2190         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2191                ktime_ms_delta(ktime_get(), time_before));
2192
2193         spin_lock(&ps->ps_lock);
2194         ps->ps_increasing = 0;
2195         if (rc == 0) {
2196                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2197         } else {
2198                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2199                 CERROR("Can't allocate new %s pool because out of memory\n",
2200                        ps->ps_name);
2201         }
2202         spin_unlock(&ps->ps_lock);
2203
2204         goto again;
2205 }
2206
2207 static void
2208 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2209 {
2210         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2211                                                tpo_pool);
2212         int i;
2213
2214         LASSERT (pool->po_allocated == 0);
2215
2216         if (tpo->tpo_tx_pages != NULL) {
2217                 kiblnd_unmap_tx_pool(tpo);
2218                 kiblnd_free_pages(tpo->tpo_tx_pages);
2219         }
2220
2221         if (tpo->tpo_tx_descs == NULL)
2222                 goto out;
2223
2224         for (i = 0; i < pool->po_size; i++) {
2225                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2226                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2227
2228                 list_del(&tx->tx_list);
2229                 if (tx->tx_pages != NULL)
2230                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2231                 if (tx->tx_frags != NULL)
2232                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2233                                            IBLND_MAX_RDMA_FRAGS);
2234                 if (tx->tx_wrq != NULL)
2235                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2236                                            IBLND_MAX_RDMA_FRAGS);
2237                 if (tx->tx_sge != NULL) {
2238                         /* +1 is for the lnet header/message itself */
2239                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2240                                            (IBLND_MAX_RDMA_FRAGS *
2241                                            wrq_sge + 1));
2242                 }
2243                 if (tx->tx_rd != NULL)
2244                         LIBCFS_FREE(tx->tx_rd,
2245                                     offsetof(struct kib_rdma_desc,
2246                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2247         }
2248
2249         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2250 out:
2251         kiblnd_fini_pool(pool);
2252         CFS_FREE_PTR(tpo);
2253 }
2254
2255 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2256 {
2257         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2258         int ntx;
2259
2260         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2261         ntx = tunables->lnd_ntx / ncpts;
2262
2263         return max(IBLND_TX_POOL, ntx);
2264 }
2265
2266 static int
2267 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2268 {
2269         int            i;
2270         int            npg;
2271         struct kib_pool *pool;
2272         struct kib_tx_pool *tpo;
2273
2274         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2275         if (tpo == NULL) {
2276                 CERROR("Failed to allocate TX pool\n");
2277                 return -ENOMEM;
2278         }
2279
2280         pool = &tpo->tpo_pool;
2281         kiblnd_init_pool(ps, pool, size);
2282         tpo->tpo_tx_descs = NULL;
2283         tpo->tpo_tx_pages = NULL;
2284
2285         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2286         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2287                 CERROR("Can't allocate tx pages: %d\n", npg);
2288                 CFS_FREE_PTR(tpo);
2289                 return -ENOMEM;
2290         }
2291
2292         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2293                          size * sizeof(struct kib_tx));
2294         if (tpo->tpo_tx_descs == NULL) {
2295                 CERROR("Can't allocate %d tx descriptors\n", size);
2296                 ps->ps_pool_destroy(pool);
2297                 return -ENOMEM;
2298         }
2299
2300         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2301
2302         for (i = 0; i < size; i++) {
2303                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2304                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2305
2306                 tx->tx_pool = tpo;
2307                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2308                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2309                                          lnet_cpt_table(), ps->ps_cpt,
2310                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2311                         if (tx->tx_pages == NULL)
2312                                 break;
2313                 }
2314
2315                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2316                                  IBLND_MAX_RDMA_FRAGS *
2317                                  sizeof(*tx->tx_frags));
2318                 if (tx->tx_frags == NULL)
2319                         break;
2320
2321                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2322
2323                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2324                                  IBLND_MAX_RDMA_FRAGS *
2325                                  sizeof(*tx->tx_wrq));
2326                 if (tx->tx_wrq == NULL)
2327                         break;
2328
2329                 /* +1 is for the lnet header/message itself */
2330                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2331                                  (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2332                                  sizeof(*tx->tx_sge));
2333                 if (tx->tx_sge == NULL)
2334                         break;
2335
2336                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2337                                  offsetof(struct kib_rdma_desc,
2338                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2339                 if (tx->tx_rd == NULL)
2340                         break;
2341         }
2342
2343         if (i == size) {
2344                 kiblnd_map_tx_pool(tpo);
2345                 *pp_po = pool;
2346                 return 0;
2347         }
2348
2349         ps->ps_pool_destroy(pool);
2350         return -ENOMEM;
2351 }
2352
2353 static void
2354 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2355 {
2356         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2357                                                   struct kib_tx_poolset,
2358                                                   tps_poolset);
2359         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2360
2361         tx->tx_cookie = tps->tps_next_tx_cookie++;
2362 }
2363
2364 static void
2365 kiblnd_net_fini_pools(struct kib_net *net)
2366 {
2367         int     i;
2368
2369         cfs_cpt_for_each(i, lnet_cpt_table()) {
2370                 struct kib_tx_poolset *tps;
2371                 struct kib_fmr_poolset *fps;
2372
2373                 if (net->ibn_tx_ps != NULL) {
2374                         tps = net->ibn_tx_ps[i];
2375                         kiblnd_fini_poolset(&tps->tps_poolset);
2376                 }
2377
2378                 if (net->ibn_fmr_ps != NULL) {
2379                         fps = net->ibn_fmr_ps[i];
2380                         kiblnd_fini_fmr_poolset(fps);
2381                 }
2382         }
2383
2384         if (net->ibn_tx_ps != NULL) {
2385                 cfs_percpt_free(net->ibn_tx_ps);
2386                 net->ibn_tx_ps = NULL;
2387         }
2388
2389         if (net->ibn_fmr_ps != NULL) {
2390                 cfs_percpt_free(net->ibn_fmr_ps);
2391                 net->ibn_fmr_ps = NULL;
2392         }
2393 }
2394
2395 static int
2396 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2397                       int ncpts)
2398 {
2399         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2400 #ifdef HAVE_IB_GET_DMA_MR
2401         unsigned long   flags;
2402 #endif
2403         int             cpt;
2404         int             rc;
2405         int             i;
2406
2407         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2408
2409 #ifdef HAVE_IB_GET_DMA_MR
2410         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2411         /*
2412          * if lnd_map_on_demand is zero then we have effectively disabled
2413          * FMR or FastReg and we're using global memory regions
2414          * exclusively.
2415          */
2416         if (!tunables->lnd_map_on_demand) {
2417                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2418                                            flags);
2419                 goto create_tx_pool;
2420         }
2421
2422         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2423 #endif
2424
2425         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2426                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2427                        tunables->lnd_fmr_pool_size,
2428                        tunables->lnd_ntx / 4);
2429                 rc = -EINVAL;
2430                 goto failed;
2431         }
2432
2433         /* TX pool must be created later than FMR, see LU-2268
2434          * for details */
2435         LASSERT(net->ibn_tx_ps == NULL);
2436
2437         /* premapping can fail if ibd_nmr > 1, so we always create
2438          * FMR pool and map-on-demand if premapping failed */
2439
2440         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2441                                            sizeof(struct kib_fmr_poolset));
2442         if (net->ibn_fmr_ps == NULL) {
2443                 CERROR("Failed to allocate FMR pool array\n");
2444                 rc = -ENOMEM;
2445                 goto failed;
2446         }
2447
2448         for (i = 0; i < ncpts; i++) {
2449                 cpt = (cpts == NULL) ? i : cpts[i];
2450                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2451                                              net, tunables);
2452                 if (rc != 0) {
2453                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2454                                cpt, rc);
2455                         goto failed;
2456                 }
2457         }
2458
2459         if (i > 0)
2460                 LASSERT(i == ncpts);
2461
2462 #ifdef HAVE_IB_GET_DMA_MR
2463  create_tx_pool:
2464 #endif
2465         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2466                                           sizeof(struct kib_tx_poolset));
2467         if (net->ibn_tx_ps == NULL) {
2468                 CERROR("Failed to allocate tx pool array\n");
2469                 rc = -ENOMEM;
2470                 goto failed;
2471         }
2472
2473         for (i = 0; i < ncpts; i++) {
2474                 cpt = (cpts == NULL) ? i : cpts[i];
2475                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2476                                          cpt, net, "TX",
2477                                          kiblnd_tx_pool_size(ni, ncpts),
2478                                          kiblnd_create_tx_pool,
2479                                          kiblnd_destroy_tx_pool,
2480                                          kiblnd_tx_init, NULL);
2481                 if (rc != 0) {
2482                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2483                                cpt, rc);
2484                         goto failed;
2485                 }
2486         }
2487
2488         return 0;
2489  failed:
2490         kiblnd_net_fini_pools(net);
2491         LASSERT(rc != 0);
2492         return rc;
2493 }
2494
2495 static int
2496 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2497 {
2498         struct ib_port_attr *port_attr;
2499         int rc;
2500         unsigned long flags;
2501         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2502
2503         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2504         if (port_attr == NULL) {
2505                 CDEBUG(D_NETERROR, "Out of memory\n");
2506                 return -ENOMEM;
2507         }
2508
2509         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2510
2511         write_lock_irqsave(g_lock, flags);
2512
2513         if (rc == 0)
2514                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2515                                  ? IBLND_DEV_PORT_ACTIVE
2516                                  : IBLND_DEV_PORT_DOWN;
2517
2518         write_unlock_irqrestore(g_lock, flags);
2519         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2520
2521         if (rc != 0) {
2522                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2523                 return rc;
2524         }
2525         return 0;
2526 }
2527
2528 static inline void
2529 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2530 {
2531         struct kib_net  *net;
2532
2533         /* for health check */
2534         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2535                 if (val)
2536                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2537                                         libcfs_nidstr(&net->ibn_ni->ni_nid));
2538                 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2539         }
2540 }
2541
2542 void
2543 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2544 {
2545         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2546         struct kib_hca_dev  *hdev;
2547         unsigned long flags;
2548
2549         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2550
2551         write_lock_irqsave(g_lock, flags);
2552
2553         switch (event->event) {
2554         case IB_EVENT_DEVICE_FATAL:
2555                 CDEBUG(D_NET, "IB device fatal\n");
2556                 hdev->ibh_state = IBLND_DEV_FATAL;
2557                 kiblnd_set_ni_fatal_on(hdev, 1);
2558                 break;
2559         case IB_EVENT_PORT_ACTIVE:
2560                 CDEBUG(D_NET, "IB port active\n");
2561                 if (event->element.port_num == hdev->ibh_port) {
2562                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2563                         kiblnd_set_ni_fatal_on(hdev, 0);
2564                 }
2565                 break;
2566         case IB_EVENT_PORT_ERR:
2567                 CDEBUG(D_NET, "IB port err\n");
2568                 if (event->element.port_num == hdev->ibh_port) {
2569                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2570                         kiblnd_set_ni_fatal_on(hdev, 1);
2571                 }
2572                 break;
2573         default:
2574                 break;
2575         }
2576         write_unlock_irqrestore(g_lock, flags);
2577 }
2578
2579 static int
2580 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2581 {
2582         struct ib_device_attr *dev_attr;
2583         int rc = 0;
2584         int rc2 = 0;
2585
2586         /* It's safe to assume a HCA can handle a page size
2587          * matching that of the native system */
2588         hdev->ibh_page_shift = PAGE_SHIFT;
2589         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2590         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2591
2592 #ifndef HAVE_IB_DEVICE_ATTRS
2593         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2594         if (dev_attr == NULL) {
2595                 CERROR("Out of memory\n");
2596                 return -ENOMEM;
2597         }
2598
2599         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2600         if (rc != 0) {
2601                 CERROR("Failed to query IB device: %d\n", rc);
2602                 goto out_clean_attr;
2603         }
2604 #else
2605         dev_attr = &hdev->ibh_ibdev->attrs;
2606 #endif
2607
2608         hdev->ibh_mr_size = dev_attr->max_mr_size;
2609         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2610
2611         /* Setup device Memory Registration capabilities */
2612 #ifdef HAVE_FMR_POOL_API
2613 #ifdef HAVE_IB_DEVICE_OPS
2614         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2615             hdev->ibh_ibdev->ops.dealloc_fmr &&
2616             hdev->ibh_ibdev->ops.map_phys_fmr &&
2617             hdev->ibh_ibdev->ops.unmap_fmr) {
2618 #else
2619         if (hdev->ibh_ibdev->alloc_fmr &&
2620             hdev->ibh_ibdev->dealloc_fmr &&
2621             hdev->ibh_ibdev->map_phys_fmr &&
2622             hdev->ibh_ibdev->unmap_fmr) {
2623 #endif
2624                 LCONSOLE_INFO("Using FMR for registration\n");
2625                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2626         } else
2627 #endif /* HAVE_FMR_POOL_API */
2628         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2629                 LCONSOLE_INFO("Using FastReg for registration\n");
2630                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2631 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2632 #ifdef IB_DEVICE_SG_GAPS_REG
2633                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2634                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2635 #endif
2636 #endif
2637         } else {
2638                 rc = -ENOSYS;
2639         }
2640
2641         rc2 = kiblnd_port_get_attr(hdev);
2642         if (rc2 != 0)
2643                 return rc2;
2644
2645         if (rc != 0)
2646                 rc = -EINVAL;
2647
2648 #ifndef HAVE_IB_DEVICE_ATTRS
2649 out_clean_attr:
2650         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2651 #endif
2652
2653         if (rc == -ENOSYS)
2654                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2655                        "register memory: %d\n", rc);
2656         else if (rc == -EINVAL)
2657                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2658         return rc;
2659 }
2660
2661 #ifdef HAVE_IB_GET_DMA_MR
2662 static void
2663 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2664 {
2665         if (hdev->ibh_mrs == NULL)
2666                 return;
2667
2668         ib_dereg_mr(hdev->ibh_mrs);
2669
2670         hdev->ibh_mrs = NULL;
2671 }
2672 #endif
2673
2674 void
2675 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2676 {
2677         if (hdev->ibh_event_handler.device != NULL)
2678                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2679
2680 #ifdef HAVE_IB_GET_DMA_MR
2681         kiblnd_hdev_cleanup_mrs(hdev);
2682 #endif
2683
2684         if (hdev->ibh_pd != NULL)
2685                 ib_dealloc_pd(hdev->ibh_pd);
2686
2687         if (hdev->ibh_cmid != NULL)
2688                 rdma_destroy_id(hdev->ibh_cmid);
2689
2690         LIBCFS_FREE(hdev, sizeof(*hdev));
2691 }
2692
2693 #ifdef HAVE_IB_GET_DMA_MR
2694 static int
2695 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2696 {
2697         struct ib_mr *mr;
2698         int           acflags = IB_ACCESS_LOCAL_WRITE |
2699                                 IB_ACCESS_REMOTE_WRITE;
2700
2701         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2702         if (IS_ERR(mr)) {
2703                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2704                 kiblnd_hdev_cleanup_mrs(hdev);
2705                 return PTR_ERR(mr);
2706         }
2707
2708         hdev->ibh_mrs = mr;
2709
2710         return 0;
2711 }
2712 #endif
2713
2714 static int
2715 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2716 {       /* DUMMY */
2717         return 0;
2718 }
2719
2720 static int kiblnd_get_link_status(struct net_device *dev)
2721 {
2722         int ret = -1;
2723
2724         LASSERT(dev);
2725
2726         if (!netif_running(dev))
2727                 ret = 0;
2728         /* Some devices may not be providing link settings */
2729         else if (dev->ethtool_ops->get_link)
2730                 ret = dev->ethtool_ops->get_link(dev);
2731
2732         return ret;
2733 }
2734
2735 static int
2736 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2737 {
2738         struct rdma_cm_id  *cmid;
2739         struct sockaddr_in  srcaddr;
2740         struct sockaddr_in  dstaddr;
2741         int                 rc;
2742
2743         if (dev->ibd_hdev == NULL || /* initializing */
2744             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2745             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2746                 return 1;
2747
2748         /* XXX: it's UGLY, but I don't have better way to find
2749          * ib-bonding HCA failover because:
2750          *
2751          * a. no reliable CM event for HCA failover...
2752          * b. no OFED API to get ib_device for current net_device...
2753          *
2754          * We have only two choices at this point:
2755          *
2756          * a. rdma_bind_addr(), it will conflict with listener cmid
2757          * b. rdma_resolve_addr() to zero addr */
2758         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2759                                      RDMA_PS_TCP, IB_QPT_RC);
2760         if (IS_ERR(cmid)) {
2761                 rc = PTR_ERR(cmid);
2762                 CERROR("Failed to create cmid for failover: %d\n", rc);
2763                 return rc;
2764         }
2765
2766         memset(&srcaddr, 0, sizeof(srcaddr));
2767         srcaddr.sin_family      = AF_INET;
2768         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2769
2770         memset(&dstaddr, 0, sizeof(dstaddr));
2771         dstaddr.sin_family = AF_INET;
2772         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2773                                (struct sockaddr *)&dstaddr, 1);
2774         if (rc != 0 || cmid->device == NULL) {
2775                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2776                        dev->ibd_ifname, &dev->ibd_ifip,
2777                        cmid->device, rc);
2778                 rdma_destroy_id(cmid);
2779                 return rc;
2780         }
2781
2782         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2783         rdma_destroy_id(cmid);
2784         return rc;
2785 }
2786
2787 int
2788 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2789 {
2790         LIST_HEAD(zombie_tpo);
2791         LIST_HEAD(zombie_ppo);
2792         LIST_HEAD(zombie_fpo);
2793         struct rdma_cm_id  *cmid  = NULL;
2794         struct kib_hca_dev *hdev  = NULL;
2795         struct kib_hca_dev *old;
2796         struct ib_pd       *pd;
2797         struct kib_net *net;
2798         struct sockaddr_in  addr;
2799         struct net_device *netdev;
2800         unsigned long       flags;
2801         int                 rc = 0;
2802         int                 i;
2803         bool                set_fatal = true;
2804
2805         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2806                 dev->ibd_can_failover ||
2807                 dev->ibd_hdev == NULL);
2808
2809         rc = kiblnd_dev_need_failover(dev, ns);
2810         if (rc <= 0)
2811                 goto out;
2812
2813         if (dev->ibd_hdev != NULL &&
2814             dev->ibd_hdev->ibh_cmid != NULL) {
2815                 /* XXX it's not good to close old listener at here,
2816                  * because we can fail to create new listener.
2817                  * But we have to close it now, otherwise rdma_bind_addr
2818                  * will return EADDRINUSE... How crap! */
2819                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2820
2821                 cmid = dev->ibd_hdev->ibh_cmid;
2822                 /* make next schedule of kiblnd_dev_need_failover()
2823                  * return 1 for me */
2824                 dev->ibd_hdev->ibh_cmid  = NULL;
2825                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2826
2827                 rdma_destroy_id(cmid);
2828         }
2829
2830         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2831                                      IB_QPT_RC);
2832         if (IS_ERR(cmid)) {
2833                 rc = PTR_ERR(cmid);
2834                 CERROR("Failed to create cmid for failover: %d\n", rc);
2835                 goto out;
2836         }
2837
2838         memset(&addr, 0, sizeof(addr));
2839         addr.sin_family      = AF_INET;
2840         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2841         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2842
2843         /* Bind to failover device or port */
2844         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2845         if (rc != 0 || cmid->device == NULL) {
2846                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2847                        dev->ibd_ifname, &dev->ibd_ifip,
2848                        cmid->device, rc);
2849                 if (!rc && !cmid->device)
2850                         set_fatal = false;
2851                 rdma_destroy_id(cmid);
2852                 goto out;
2853         }
2854
2855         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2856         if (hdev == NULL) {
2857                 CERROR("Failed to allocate kib_hca_dev\n");
2858                 rdma_destroy_id(cmid);
2859                 rc = -ENOMEM;
2860                 goto out;
2861         }
2862
2863         atomic_set(&hdev->ibh_ref, 1);
2864         hdev->ibh_dev   = dev;
2865         hdev->ibh_cmid  = cmid;
2866         hdev->ibh_ibdev = cmid->device;
2867         hdev->ibh_port  = cmid->port_num;
2868
2869 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2870         pd = ib_alloc_pd(cmid->device, 0);
2871 #else
2872         pd = ib_alloc_pd(cmid->device);
2873 #endif
2874         if (IS_ERR(pd)) {
2875                 rc = PTR_ERR(pd);
2876                 CERROR("Can't allocate PD: %d\n", rc);
2877                 goto out;
2878         }
2879
2880         hdev->ibh_pd = pd;
2881
2882         rc = rdma_listen(cmid, 0);
2883         if (rc != 0) {
2884                 CERROR("Can't start new listener: %d\n", rc);
2885                 goto out;
2886         }
2887
2888         rc = kiblnd_hdev_get_attr(hdev);
2889         if (rc != 0) {
2890                 CERROR("Can't get device attributes: %d\n", rc);
2891                 goto out;
2892         }
2893
2894 #ifdef HAVE_IB_GET_DMA_MR
2895         rc = kiblnd_hdev_setup_mrs(hdev);
2896         if (rc != 0) {
2897                 CERROR("Can't setup device: %d\n", rc);
2898                 goto out;
2899         }
2900 #endif
2901
2902         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2903                                 hdev->ibh_ibdev, kiblnd_event_handler);
2904         ib_register_event_handler(&hdev->ibh_event_handler);
2905
2906         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2907
2908         old = dev->ibd_hdev;
2909         dev->ibd_hdev = hdev;   /* take over the refcount */
2910         hdev = old;
2911
2912         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2913                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2914                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2915                                             &zombie_tpo);
2916
2917                         if (net->ibn_fmr_ps != NULL)
2918                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2919                                                         &zombie_fpo);
2920                 }
2921         }
2922
2923         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2924  out:
2925         if (!list_empty(&zombie_tpo))
2926                 kiblnd_destroy_pool_list(&zombie_tpo);
2927         if (!list_empty(&zombie_ppo))
2928                 kiblnd_destroy_pool_list(&zombie_ppo);
2929         if (!list_empty(&zombie_fpo))
2930                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2931         if (hdev != NULL)
2932                 kiblnd_hdev_decref(hdev);
2933
2934         if (rc != 0) {
2935                 dev->ibd_failed_failover++;
2936         } else {
2937                 dev->ibd_failed_failover = 0;
2938
2939                 if (set_fatal) {
2940                         rcu_read_lock();
2941                         netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
2942                         if (netdev && (kiblnd_get_link_status(netdev) == 1))
2943                                 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
2944                         rcu_read_unlock();
2945                 }
2946         }
2947
2948         return rc;
2949 }
2950
2951 void
2952 kiblnd_destroy_dev(struct kib_dev *dev)
2953 {
2954         LASSERT(dev->ibd_nnets == 0);
2955         LASSERT(list_empty(&dev->ibd_nets));
2956
2957         list_del(&dev->ibd_fail_list);
2958         list_del(&dev->ibd_list);
2959
2960         if (dev->ibd_hdev != NULL)
2961                 kiblnd_hdev_decref(dev->ibd_hdev);
2962
2963         LIBCFS_FREE(dev, sizeof(*dev));
2964 }
2965
2966 static struct kib_dev *
2967 kiblnd_dev_search(char *ifname)
2968 {
2969         struct kib_dev *alias = NULL;
2970         struct kib_dev *dev;
2971         char *colon;
2972         char *colon2;
2973
2974         colon = strchr(ifname, ':');
2975         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2976                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2977                         return dev;
2978
2979                 if (alias != NULL)
2980                         continue;
2981
2982                 colon2 = strchr(dev->ibd_ifname, ':');
2983                 if (colon != NULL)
2984                         *colon = 0;
2985                 if (colon2 != NULL)
2986                         *colon2 = 0;
2987
2988                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2989                         alias = dev;
2990
2991                 if (colon != NULL)
2992                         *colon = ':';
2993                 if (colon2 != NULL)
2994                         *colon2 = ':';
2995         }
2996         return alias;
2997 }
2998
2999 static int
3000 kiblnd_handle_link_state_change(struct net_device *dev,
3001                                 unsigned char operstate)
3002 {
3003         struct lnet_ni *ni = NULL;
3004         struct kib_dev *event_kibdev;
3005         struct kib_net *net;
3006         struct kib_net *cnxt;
3007         bool link_down = !(operstate == IF_OPER_UP);
3008         struct in_device *in_dev;
3009         bool found_ip = false;
3010         DECLARE_CONST_IN_IFADDR(ifa);
3011
3012         event_kibdev = kiblnd_dev_search(dev->name);
3013
3014         if (!event_kibdev)
3015                 goto out;
3016
3017         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3018                 found_ip = false;
3019
3020                 ni = net->ibn_ni;
3021
3022                 in_dev = __in_dev_get_rtnl(dev);
3023                 if (!in_dev) {
3024                         CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3025                                dev->name);
3026                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3027                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3028                         atomic_set(&ni->ni_fatal_error_on, 1);
3029                         continue;
3030                 }
3031                 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3032                         if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3033                                 found_ip = true;
3034                 }
3035                 endfor_ifa(in_dev);
3036
3037                 if (!found_ip) {
3038                         CDEBUG(D_NET, "Interface %s has no matching ip\n",
3039                                dev->name);
3040                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3041                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3042                         atomic_set(&ni->ni_fatal_error_on, 1);
3043                         continue;
3044                 }
3045
3046                 if (link_down) {
3047                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3048                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3049                         atomic_set(&ni->ni_fatal_error_on, link_down);
3050                 } else {
3051                         CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3052                                libcfs_nidstr(&net->ibn_ni->ni_nid),
3053                                (kiblnd_get_link_status(dev) == 0));
3054                         atomic_set(&ni->ni_fatal_error_on,
3055                                    (kiblnd_get_link_status(dev) == 0));
3056                 }
3057         }
3058 out:
3059         return 0;
3060 }
3061
3062 static int
3063 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3064 {
3065         struct kib_dev *event_kibdev;
3066         struct kib_net *net;
3067         struct kib_net *cnxt;
3068         struct net_device *event_netdev = ifa->ifa_dev->dev;
3069
3070         event_kibdev = kiblnd_dev_search(event_netdev->name);
3071
3072         if (!event_kibdev)
3073                 goto out;
3074
3075         if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3076                 goto out;
3077
3078         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3079                                  ibn_list) {
3080                 CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3081                        libcfs_nidstr(&net->ibn_ni->ni_nid),
3082                        (event == NETDEV_DOWN));
3083                 atomic_set(&net->ibn_ni->ni_fatal_error_on,
3084                            (event == NETDEV_DOWN));
3085         }
3086 out:
3087         return 0;
3088 }
3089
3090
3091 /************************************
3092  * Net device notifier event handler
3093  ************************************/
3094 static int kiblnd_device_event(struct notifier_block *unused,
3095                                  unsigned long event, void *ptr)
3096 {
3097         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3098         unsigned char operstate;
3099
3100         operstate = dev->operstate;
3101
3102         CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3103                event, dev->name, dev->ifindex, operstate);
3104
3105         switch (event) {
3106         case NETDEV_UP:
3107         case NETDEV_DOWN:
3108         case NETDEV_CHANGE:
3109                 kiblnd_handle_link_state_change(dev, operstate);
3110                 break;
3111         }
3112
3113         return NOTIFY_OK;
3114 }
3115
3116 /************************************
3117  * Inetaddr notifier event handler
3118  ************************************/
3119 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3120                                  unsigned long event, void *ptr)
3121 {
3122         struct in_ifaddr *ifa = ptr;
3123
3124         CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3125                event, &ifa->ifa_address, &ifa->ifa_mask);
3126
3127         switch (event) {
3128         case NETDEV_UP:
3129         case NETDEV_DOWN:
3130         case NETDEV_CHANGE:
3131                 kiblnd_handle_inetaddr_change(ifa, event);
3132                 break;
3133
3134         }
3135         return NOTIFY_OK;
3136 }
3137
3138 static struct notifier_block kiblnd_dev_notifier_block = {
3139         .notifier_call = kiblnd_device_event,
3140 };
3141
3142 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3143         .notifier_call = kiblnd_inetaddr_event,
3144 };
3145
3146 static void
3147 kiblnd_base_shutdown(void)
3148 {
3149         struct kib_sched_info *sched;
3150         struct kib_peer_ni *peer_ni;
3151         int i;
3152
3153         LASSERT(list_empty(&kiblnd_data.kib_devs));
3154
3155         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3156                libcfs_kmem_read());
3157
3158         if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3159                 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3160                 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3161         }
3162
3163         switch (kiblnd_data.kib_init) {
3164         default:
3165                 LBUG();
3166
3167         case IBLND_INIT_ALL:
3168         case IBLND_INIT_DATA:
3169                 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3170                         LASSERT(0);
3171                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3172                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3173                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3174                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3175
3176                 /* flag threads to terminate; wake and wait for them to die */
3177                 kiblnd_data.kib_shutdown = 1;
3178
3179                 /* NB: we really want to stop scheduler threads net by net
3180                  * instead of the whole module, this should be improved
3181                  * with dynamic configuration LNet.
3182                  */
3183                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3184                         wake_up_all(&sched->ibs_waitq);
3185
3186                 wake_up(&kiblnd_data.kib_connd_waitq);
3187                 wake_up(&kiblnd_data.kib_failover_waitq);
3188
3189                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3190                                        !atomic_read(&kiblnd_data.kib_nthreads),
3191                                        "Waiting for %d threads to terminate\n",
3192                                        atomic_read(&kiblnd_data.kib_nthreads));
3193                 fallthrough;
3194
3195         case IBLND_INIT_NOTHING:
3196                 break;
3197         }
3198
3199         if (kiblnd_data.kib_scheds != NULL)
3200                 cfs_percpt_free(kiblnd_data.kib_scheds);
3201
3202         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3203                libcfs_kmem_read());
3204
3205         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3206         module_put(THIS_MODULE);
3207 }
3208
3209 static void
3210 kiblnd_shutdown(struct lnet_ni *ni)
3211 {
3212         struct kib_net *net = ni->ni_data;
3213         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3214         unsigned long     flags;
3215
3216         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3217
3218         if (net == NULL)
3219                 goto out;
3220
3221         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3222                libcfs_kmem_read());
3223
3224         write_lock_irqsave(g_lock, flags);
3225         net->ibn_shutdown = 1;
3226         write_unlock_irqrestore(g_lock, flags);
3227
3228         switch (net->ibn_init) {
3229         default:
3230                 LBUG();
3231
3232         case IBLND_INIT_ALL:
3233                 /* nuke all existing peers within this net */
3234                 kiblnd_del_peer(ni, LNET_NID_ANY);
3235
3236                 /* Wait for all peer_ni state to clean up */
3237                 wait_var_event_warning(&net->ibn_npeers,
3238                                        atomic_read(&net->ibn_npeers) == 0,
3239                                        "%s: waiting for %d peers to disconnect\n",
3240                                        libcfs_nidstr(&ni->ni_nid),
3241                                        atomic_read(&net->ibn_npeers));
3242
3243                 kiblnd_net_fini_pools(net);
3244
3245                 write_lock_irqsave(g_lock, flags);
3246                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3247                 net->ibn_dev->ibd_nnets--;
3248                 list_del(&net->ibn_list);
3249                 write_unlock_irqrestore(g_lock, flags);
3250
3251                 wake_up_all(&kiblnd_data.kib_connd_waitq);
3252                 wait_var_event_warning(&net->ibn_nconns,
3253                                        atomic_read(&net->ibn_nconns) == 0,
3254                                        "%s: waiting for %d conns to clean\n",
3255                                        libcfs_nidstr(&ni->ni_nid),
3256                                        atomic_read(&net->ibn_nconns));
3257                 fallthrough;
3258
3259         case IBLND_INIT_NOTHING:
3260                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3261
3262                 if (net->ibn_dev != NULL &&
3263                     net->ibn_dev->ibd_nnets == 0)
3264                         kiblnd_destroy_dev(net->ibn_dev);
3265
3266                 break;
3267         }
3268
3269         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3270                libcfs_kmem_read());
3271
3272         net->ibn_init = IBLND_INIT_NOTHING;
3273         ni->ni_data = NULL;
3274
3275         LIBCFS_FREE(net, sizeof(*net));
3276
3277 out:
3278         if (list_empty(&kiblnd_data.kib_devs))
3279                 kiblnd_base_shutdown();
3280 }
3281
3282 static int
3283 kiblnd_base_startup(struct net *ns)
3284 {
3285         struct kib_sched_info *sched;
3286         int rc;
3287         int i;
3288
3289         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3290
3291         if (!try_module_get(THIS_MODULE))
3292                 goto failed;
3293
3294         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3295
3296         rwlock_init(&kiblnd_data.kib_global_lock);
3297
3298         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3299         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3300
3301         hash_init(kiblnd_data.kib_peers);
3302
3303         spin_lock_init(&kiblnd_data.kib_connd_lock);
3304         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3305         INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3306         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3307         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3308         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3309
3310         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3311         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3312
3313         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3314                                                   sizeof(*sched));
3315         if (kiblnd_data.kib_scheds == NULL)
3316                 goto failed;
3317
3318         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3319                 int     nthrs;
3320
3321                 spin_lock_init(&sched->ibs_lock);
3322                 INIT_LIST_HEAD(&sched->ibs_conns);
3323                 init_waitqueue_head(&sched->ibs_waitq);
3324
3325                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3326                 if (*kiblnd_tunables.kib_nscheds > 0) {
3327                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3328                 } else {
3329                         /* max to half of CPUs, another half is reserved for
3330                          * upper layer modules */
3331                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3332                 }
3333
3334                 sched->ibs_nthreads_max = nthrs;
3335                 sched->ibs_cpt = i;
3336         }
3337
3338         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3339
3340         /* lists/ptrs/locks initialised */
3341         kiblnd_data.kib_init = IBLND_INIT_DATA;
3342         /*****************************************************/
3343
3344         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3345         if (rc != 0) {
3346                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3347                 goto failed;
3348         }
3349
3350         if (*kiblnd_tunables.kib_dev_failover != 0)
3351                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3352                                          "kiblnd_failover");
3353
3354         if (rc != 0) {
3355                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3356                 goto failed;
3357         }
3358
3359         register_netdevice_notifier(&kiblnd_dev_notifier_block);
3360         register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3361
3362         /* flag everything initialised */
3363         kiblnd_data.kib_init = IBLND_INIT_ALL;
3364         /*****************************************************/
3365
3366         return 0;
3367
3368  failed:
3369         kiblnd_base_shutdown();
3370         return -ENETDOWN;
3371 }
3372
3373 static int
3374 kiblnd_start_schedulers(struct kib_sched_info *sched)
3375 {
3376         int     rc = 0;
3377         int     nthrs;
3378         int     i;
3379
3380         if (sched->ibs_nthreads == 0) {
3381                 if (*kiblnd_tunables.kib_nscheds > 0) {
3382                         nthrs = sched->ibs_nthreads_max;
3383                 } else {
3384                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3385                                                sched->ibs_cpt);
3386                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3387                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3388                 }
3389         } else {
3390                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3391                 /* increase one thread if there is new interface */
3392                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3393         }
3394
3395         for (i = 0; i < nthrs; i++) {
3396                 long    id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3397
3398                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3399                                          "kiblnd_sd_%02ld_%02ld",
3400                                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3401                 if (rc == 0)
3402                         continue;
3403
3404                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3405                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3406                 break;
3407         }
3408
3409         sched->ibs_nthreads += i;
3410         return rc;
3411 }
3412
3413 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3414                                     int ncpts)
3415 {
3416         int     cpt;
3417         int     rc;
3418         int     i;
3419
3420         for (i = 0; i < ncpts; i++) {
3421                 struct kib_sched_info *sched;
3422
3423                 cpt = (cpts == NULL) ? i : cpts[i];
3424                 sched = kiblnd_data.kib_scheds[cpt];
3425
3426                 if (!newdev && sched->ibs_nthreads > 0)
3427                         continue;
3428
3429                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3430                 if (rc != 0) {
3431                         CERROR("Failed to start scheduler threads for %s\n",
3432                                dev->ibd_ifname);
3433                         return rc;
3434                 }
3435         }
3436         return 0;
3437 }
3438
3439 static int
3440 kiblnd_startup(struct lnet_ni *ni)
3441 {
3442         char *ifname = NULL;
3443         struct lnet_inetdev *ifaces = NULL;
3444         struct kib_dev *ibdev = NULL;
3445         struct kib_net *net = NULL;
3446         unsigned long flags;
3447         int rc;
3448         int i;
3449         bool newdev;
3450
3451         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3452
3453         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3454                 rc = kiblnd_base_startup(ni->ni_net_ns);
3455                 if (rc != 0)
3456                         return rc;
3457         }
3458
3459         LIBCFS_ALLOC(net, sizeof(*net));
3460         ni->ni_data = net;
3461         if (net == NULL) {
3462                 rc = -ENOMEM;
3463                 goto failed;
3464         }
3465
3466         net->ibn_ni = ni;
3467         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3468
3469         kiblnd_tunables_setup(ni);
3470
3471         /*
3472          * Multi-Rail wants each secondary
3473          * IP to be treated as an unique 'struct ni' interface.
3474          */
3475         if (ni->ni_interface != NULL) {
3476                 /* Use the IPoIB interface specified in 'networks=' */
3477                 ifname = ni->ni_interface;
3478         } else {
3479                 ifname = *kiblnd_tunables.kib_default_ipif;
3480         }
3481
3482         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3483                 CERROR("IPoIB interface name too long: %s\n", ifname);
3484                 rc = -E2BIG;
3485                 goto failed;
3486         }
3487
3488         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3489         if (rc < 0)
3490                 goto failed;
3491
3492         for (i = 0; i < rc; i++) {
3493                 if (strcmp(ifname, ifaces[i].li_name) == 0)
3494                         break;
3495         }
3496
3497         if (i == rc) {
3498                 CERROR("ko2iblnd: No matching interfaces\n");
3499                 rc = -ENOENT;
3500                 goto failed;
3501         }
3502
3503         ibdev = kiblnd_dev_search(ifname);
3504         newdev = ibdev == NULL;
3505         /* hmm...create kib_dev even for alias */
3506         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3507                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3508                 if (!ibdev) {
3509                         rc = -ENOMEM;
3510                         goto failed;
3511                 }
3512
3513                 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3514                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3515                         sizeof(ibdev->ibd_ifname));
3516                 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3517
3518                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3519                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3520                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3521
3522                 /* initialize the device */
3523                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3524                 if (rc) {
3525                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3526                                rc);
3527                         goto failed;
3528                 }
3529
3530                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3531         }
3532
3533         net->ibn_dev = ibdev;
3534         ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3535
3536         ni->ni_dev_cpt = ifaces[i].li_cpt;
3537
3538         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3539         if (rc != 0)
3540                 goto failed;
3541
3542         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3543         if (rc != 0) {
3544                 CERROR("Failed to initialize NI pools: %d\n", rc);
3545                 goto failed;
3546         }
3547
3548         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3549         ibdev->ibd_nnets++;
3550         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3551         /* for health check */
3552         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3553                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3554         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3555
3556         net->ibn_init = IBLND_INIT_ALL;
3557
3558         return 0;
3559
3560 failed:
3561         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3562                 kiblnd_destroy_dev(ibdev);
3563
3564         kfree(ifaces);
3565         kiblnd_shutdown(ni);
3566
3567         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3568                ifname ? ifname : "", rc);
3569
3570         return -ENETDOWN;
3571 }
3572
3573 static const struct lnet_lnd the_o2iblnd = {
3574         .lnd_type       = O2IBLND,
3575         .lnd_startup    = kiblnd_startup,
3576         .lnd_shutdown   = kiblnd_shutdown,
3577         .lnd_ctl        = kiblnd_ctl,
3578         .lnd_send       = kiblnd_send,
3579         .lnd_recv       = kiblnd_recv,
3580         .lnd_get_dev_prio = kiblnd_get_dev_prio,
3581 };
3582
3583 static void ko2inlnd_assert_wire_constants(void)
3584 {
3585         BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3586         BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3587         BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3588         BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3589
3590         BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3591         BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3592         BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3593         BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3594         BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3595         BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3596         BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3597         BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3598         BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3599         BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3600
3601         BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3602         BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3603         BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3604         BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3605         BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3606         BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3607         BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3608         BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3609
3610         BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3611         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3612         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3613         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3614         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3615         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3616         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3617
3618         BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3619         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3620         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3621         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3622         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3623
3624         BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3625         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3626         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3627         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3628         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3629
3630         BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3631         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3632         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3633         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3634         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3635         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3636         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3637
3638         BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3639         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3640         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3641         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3642         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3643
3644         BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3645         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3646         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3647         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3648         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3649         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3650         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3651
3652         BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3653         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3654         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3655         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3656         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3657         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3658         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3659
3660         BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3661         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3662         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3663         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3664         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3665
3666         /* Checks for struct kib_msg */
3667         //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3668         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3669         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3670         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3671         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3672         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3673         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3674         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3675         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3676         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3677         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3678         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3679         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3680         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3681         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3682         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3683         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3684         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3685         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3686         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3687         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3688
3689         /* Connparams */
3690         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3691         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3692         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3693         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3694         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3695         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3696
3697         /* Immediate message */
3698         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3699         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3700         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3701         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3702
3703         /* PUT req message */
3704         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3705         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3706         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3707         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3708
3709         /* Put ACK */
3710         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3711         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3712         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3713         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3714         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3715         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3716
3717         /* GET message */
3718         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3719         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3720         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3721         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3722         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3723         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3724
3725         /* Completion message */
3726         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3727         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3728         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3729         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3730
3731         /* Sanity checks */
3732         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3733         BUILD_BUG_ON(offsetof(struct kib_msg,
3734                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3735                      IBLND_MSG_SIZE);
3736         BUILD_BUG_ON(offsetof(struct kib_msg,
3737                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3738                      IBLND_MSG_SIZE);
3739 }
3740
3741 static void __exit ko2iblnd_exit(void)
3742 {
3743         lnet_unregister_lnd(&the_o2iblnd);
3744 }
3745
3746 static int __init ko2iblnd_init(void)
3747 {
3748         int rc;
3749
3750         ko2inlnd_assert_wire_constants();
3751
3752         rc = kiblnd_tunables_init();
3753         if (rc != 0)
3754                 return rc;
3755
3756         lnet_register_lnd(&the_o2iblnd);
3757
3758         return 0;
3759 }
3760
3761 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3762 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3763 MODULE_VERSION("2.8.0");
3764 MODULE_LICENSE("GPL");
3765
3766 module_init(ko2iblnd_init);
3767 module_exit(ko2iblnd_exit);