lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/klnds/o2iblnd/o2iblnd.c
  32  *
  33  * Author: Eric Barton <eric@bartonsoftware.com>
  34  */
  35
  36 #include <asm/page.h>
  37 #include <linux/ethtool.h>
  38 #include <linux/inetdevice.h>
  39
  40 #include "o2iblnd.h"
  41
  42 static const struct lnet_lnd the_o2iblnd;
  43
  44 struct kib_data kiblnd_data;
  45
  46 static __u32
  47 kiblnd_cksum (void *ptr, int nob)
  48 {
  49         char  *c  = ptr;
  50         __u32  sum = 0;
  51
  52         while (nob-- > 0)
  53                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  54
  55         /* ensure I don't return 0 (== no checksum) */
  56         return (sum == 0) ? 1 : sum;
  57 }
  58
  59 static char *
  60 kiblnd_msgtype2str(int type)
  61 {
  62         switch (type) {
  63         case IBLND_MSG_CONNREQ:
  64                 return "CONNREQ";
  65
  66         case IBLND_MSG_CONNACK:
  67                 return "CONNACK";
  68
  69         case IBLND_MSG_NOOP:
  70                 return "NOOP";
  71
  72         case IBLND_MSG_IMMEDIATE:
  73                 return "IMMEDIATE";
  74
  75         case IBLND_MSG_PUT_REQ:
  76                 return "PUT_REQ";
  77
  78         case IBLND_MSG_PUT_NAK:
  79                 return "PUT_NAK";
  80
  81         case IBLND_MSG_PUT_ACK:
  82                 return "PUT_ACK";
  83
  84         case IBLND_MSG_PUT_DONE:
  85                 return "PUT_DONE";
  86
  87         case IBLND_MSG_GET_REQ:
  88                 return "GET_REQ";
  89
  90         case IBLND_MSG_GET_DONE:
  91                 return "GET_DONE";
  92
  93         default:
  94                 return "???";
  95         }
  96 }
  97
  98 static int
  99 kiblnd_msgtype2size(int type)
 100 {
 101         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 102
 103         switch (type) {
 104         case IBLND_MSG_CONNREQ:
 105         case IBLND_MSG_CONNACK:
 106                 return hdr_size + sizeof(struct kib_connparams);
 107
 108         case IBLND_MSG_NOOP:
 109                 return hdr_size;
 110
 111         case IBLND_MSG_IMMEDIATE:
 112                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 113
 114         case IBLND_MSG_PUT_REQ:
 115                 return hdr_size + sizeof(struct kib_putreq_msg);
 116
 117         case IBLND_MSG_PUT_ACK:
 118                 return hdr_size + sizeof(struct kib_putack_msg);
 119
 120         case IBLND_MSG_GET_REQ:
 121                 return hdr_size + sizeof(struct kib_get_msg);
 122
 123         case IBLND_MSG_PUT_NAK:
 124         case IBLND_MSG_PUT_DONE:
 125         case IBLND_MSG_GET_DONE:
 126                 return hdr_size + sizeof(struct kib_completion_msg);
 127         default:
 128                 return -1;
 129         }
 130 }
 131
 132 static int kiblnd_unpack_rd(struct kib_msg *msg, bool flip)
 133 {
 134         struct kib_rdma_desc *rd;
 135         int nob;
 136         int n;
 137         int i;
 138
 139         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 140                 msg->ibm_type == IBLND_MSG_PUT_ACK);
 141
 142         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 143                 &msg->ibm_u.get.ibgm_rd :
 144                 &msg->ibm_u.putack.ibpam_rd;
 145
 146         if (flip) {
 147                 __swab32s(&rd->rd_key);
 148                 __swab32s(&rd->rd_nfrags);
 149         }
 150
 151         n = rd->rd_nfrags;
 152
 153         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 154                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 155                        n, IBLND_MAX_RDMA_FRAGS);
 156                 return 1;
 157         }
 158
 159         nob = offsetof(struct kib_msg, ibm_u) +
 160                 kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 161
 162         if (msg->ibm_nob < nob) {
 163                 CERROR("Short %s: %d(%d)\n",
 164                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 165                 return 1;
 166         }
 167
 168         if (!flip)
 169                 return 0;
 170
 171         for (i = 0; i < n; i++) {
 172                 __swab32s(&rd->rd_frags[i].rf_nob);
 173                 __swab64s(&rd->rd_frags[i].rf_addr);
 174         }
 175
 176         return 0;
 177 }
 178
 179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
 180                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         struct kib_net *net = ni->ni_data;
 183
 184         /* CAVEAT EMPTOR! all message fields not set here should have been
 185          * initialised previously.
 186          */
 187         msg->ibm_magic    = IBLND_MSG_MAGIC;
 188         msg->ibm_version  = version;
 189         /*   ibm_type */
 190         msg->ibm_credits  = credits;
 191         /*   ibm_nob */
 192         msg->ibm_cksum    = 0;
 193         msg->ibm_srcnid   = lnet_nid_to_nid4(&ni->ni_nid);
 194         msg->ibm_srcstamp = net->ibn_incarnation;
 195         msg->ibm_dstnid   = dstnid;
 196         msg->ibm_dststamp = dststamp;
 197
 198         if (*kiblnd_tunables.kib_cksum) {
 199                 /* NB ibm_cksum zero while computing cksum */
 200                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 201         }
 202 }
 203
 204 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 205 {
 206         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 207         __u32 msg_cksum;
 208         __u16 version;
 209         int msg_nob;
 210         bool flip;
 211
 212         /* 6 bytes are enough to have received magic + version */
 213         if (nob < 6) {
 214                 CERROR("Short message: %d\n", nob);
 215                 return -EPROTO;
 216         }
 217
 218         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 219                 flip = false;
 220         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 221                 flip = true;
 222         } else {
 223                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 224                 return -EPROTO;
 225         }
 226
 227         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 228         if (version != IBLND_MSG_VERSION &&
 229             version != IBLND_MSG_VERSION_1) {
 230                 CERROR("Bad version: %x\n", version);
 231                 return -EPROTO;
 232         }
 233
 234         if (nob < hdr_size) {
 235                 CERROR("Short message: %d\n", nob);
 236                 return -EPROTO;
 237         }
 238
 239         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 240         if (msg_nob > nob) {
 241                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 242                 return -EPROTO;
 243         }
 244
 245         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 246          * gets flipped
 247          */
 248         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 249         msg->ibm_cksum = 0;
 250         if (msg_cksum != 0 &&
 251             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 252                 CERROR("Bad checksum\n");
 253                 return -EPROTO;
 254         }
 255
 256         msg->ibm_cksum = msg_cksum;
 257
 258         if (flip) {
 259                 /* leave magic unflipped as a clue to peer_ni endianness */
 260                 msg->ibm_version = version;
 261                 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
 262                 BUILD_BUG_ON(sizeof(msg->ibm_credits) != 1);
 263                 msg->ibm_nob     = msg_nob;
 264                 __swab64s(&msg->ibm_srcnid);
 265                 __swab64s(&msg->ibm_srcstamp);
 266                 __swab64s(&msg->ibm_dstnid);
 267                 __swab64s(&msg->ibm_dststamp);
 268         }
 269
 270         if (msg->ibm_srcnid == LNET_NID_ANY) {
 271                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 272                 return -EPROTO;
 273         }
 274
 275         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 276                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 277                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 278                 return -EPROTO;
 279         }
 280
 281         switch (msg->ibm_type) {
 282         default:
 283                 CERROR("Unknown message type %x\n", msg->ibm_type);
 284                 return -EPROTO;
 285
 286         case IBLND_MSG_NOOP:
 287         case IBLND_MSG_IMMEDIATE:
 288         case IBLND_MSG_PUT_REQ:
 289                 break;
 290
 291         case IBLND_MSG_PUT_ACK:
 292         case IBLND_MSG_GET_REQ:
 293                 if (kiblnd_unpack_rd(msg, flip))
 294                         return -EPROTO;
 295                 break;
 296
 297         case IBLND_MSG_PUT_NAK:
 298         case IBLND_MSG_PUT_DONE:
 299         case IBLND_MSG_GET_DONE:
 300                 if (flip)
 301                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 302                 break;
 303
 304         case IBLND_MSG_CONNREQ:
 305         case IBLND_MSG_CONNACK:
 306                 if (flip) {
 307                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 308                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 309                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 310                 }
 311                 break;
 312         }
 313         return 0;
 314 }
 315
 316 int
 317 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
 318                    lnet_nid_t nid)
 319 {
 320         struct kib_peer_ni *peer_ni;
 321         struct kib_net *net = ni->ni_data;
 322         int cpt = lnet_cpt_of_nid(nid, ni);
 323         unsigned long flags;
 324
 325         LASSERT(net != NULL);
 326         LASSERT(nid != LNET_NID_ANY);
 327
 328         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 329         if (!peer_ni) {
 330                 CERROR("Cannot allocate peer_ni\n");
 331                 return -ENOMEM;
 332         }
 333
 334         peer_ni->ibp_ni = ni;
 335         peer_ni->ibp_nid = nid;
 336         peer_ni->ibp_error = 0;
 337         peer_ni->ibp_last_alive = 0;
 338         peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
 339         peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
 340         peer_ni->ibp_queue_depth_mod = 0;       /* try to use the default */
 341         kref_init(&peer_ni->ibp_kref);
 342         atomic_set(&peer_ni->ibp_nconns, 0);
 343
 344         INIT_HLIST_NODE(&peer_ni->ibp_list);
 345         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 346         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer_ni;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer(struct kref *kref)
 364 {
 365         struct kib_peer_ni *peer_ni = container_of(kref, struct kib_peer_ni,
 366                                                    ibp_kref);
 367         struct kib_net *net = peer_ni->ibp_ni->ni_data;
 368
 369         LASSERT(net != NULL);
 370         LASSERT(!kiblnd_peer_active(peer_ni));
 371         LASSERT(kiblnd_peer_idle(peer_ni));
 372         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 373
 374         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 375
 376         /* NB a peer_ni's connections keep a reference on their peer_ni until
 377          * they are destroyed, so we can be assured that _all_ state to do
 378          * with this peer_ni has been cleaned up when its refcount drops to
 379          * zero.
 380          */
 381         if (atomic_dec_and_test(&net->ibn_npeers))
 382                 wake_up_var(&net->ibn_npeers);
 383 }
 384
 385 struct kib_peer_ni *
 386 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 387 {
 388         /* the caller is responsible for accounting the additional reference
 389          * that this creates
 390          */
 391         struct kib_peer_ni *peer_ni;
 392
 393         hash_for_each_possible(kiblnd_data.kib_peers, peer_ni,
 394                                ibp_list, nid) {
 395                 LASSERT(!kiblnd_peer_idle(peer_ni));
 396
 397                 /*
 398                  * Match a peer if its NID and the NID of the local NI it
 399                  * communicates over are the same. Otherwise don't match
 400                  * the peer, which will result in a new lnd peer being
 401                  * created.
 402                  */
 403                 if (peer_ni->ibp_nid != nid ||
 404                     !nid_same(&peer_ni->ibp_ni->ni_nid, &ni->ni_nid))
 405                         continue;
 406
 407                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 408                        peer_ni, libcfs_nid2str(nid),
 409                        kref_read(&peer_ni->ibp_kref),
 410                        peer_ni->ibp_version);
 411                 return peer_ni;
 412         }
 413         return NULL;
 414 }
 415
 416 void
 417 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
 418 {
 419         LASSERT(list_empty(&peer_ni->ibp_conns));
 420
 421         LASSERT(kiblnd_peer_active(peer_ni));
 422         hlist_del_init(&peer_ni->ibp_list);
 423         /* lose peerlist's ref */
 424         kiblnd_peer_decref(peer_ni);
 425 }
 426
 427
 428 static void
 429 kiblnd_debug_rx(struct kib_rx *rx)
 430 {
 431         CDEBUG(D_CONSOLE, "      %p msg_type %x cred %d\n",
 432                rx, rx->rx_msg->ibm_type,
 433                rx->rx_msg->ibm_credits);
 434 }
 435
 436 static void
 437 kiblnd_debug_tx(struct kib_tx *tx)
 438 {
 439         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 440                "cookie %#llx msg %s%s type %x cred %d\n",
 441                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 442                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 443                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 444                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 445                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 446 }
 447
 448 static void
 449 kiblnd_debug_conn(struct kib_conn *conn)
 450 {
 451         struct list_head        *tmp;
 452         int                     i;
 453
 454         spin_lock(&conn->ibc_lock);
 455
 456         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 457                atomic_read(&conn->ibc_refcount), conn,
 458                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 459         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 460                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 461                conn->ibc_nsends_posted, conn->ibc_credits,
 462                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 463         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 464
 465         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 466         list_for_each(tmp, &conn->ibc_early_rxs)
 467                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 468
 469         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 470         list_for_each(tmp, &conn->ibc_tx_noops)
 471                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 472
 473         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 474         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 475                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 476
 477         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 478         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 479                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 480
 481         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 482         list_for_each(tmp, &conn->ibc_tx_queue)
 483                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 484
 485         CDEBUG(D_CONSOLE, "   active_txs:\n");
 486         list_for_each(tmp, &conn->ibc_active_txs)
 487                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 488
 489         CDEBUG(D_CONSOLE, "   rxs:\n");
 490         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 491                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 492
 493         spin_unlock(&conn->ibc_lock);
 494 }
 495
 496 static void
 497 kiblnd_dump_peer_debug_info(struct kib_peer_ni *peer_ni)
 498 {
 499         struct kib_conn *conn;
 500         struct kib_conn *cnxt;
 501         int count = 0;
 502
 503         CDEBUG(D_CONSOLE, "[last_alive, races, reconnected, error]: %lld, %d, %d, %d\n",
 504                peer_ni->ibp_last_alive,
 505                peer_ni->ibp_races,
 506                peer_ni->ibp_reconnected,
 507                peer_ni->ibp_error);
 508         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 509                                  ibc_list) {
 510                 CDEBUG(D_CONSOLE, "Conn %d:\n", count);
 511                 kiblnd_debug_conn(conn);
 512                 count++;
 513         }
 514 }
 515
 516
 517 static int
 518 kiblnd_get_peer_info(struct lnet_ni *ni, lnet_nid_t nid, int index,
 519                      lnet_nid_t *nidp, int *count)
 520 {
 521         struct kib_peer_ni              *peer_ni;
 522         int                      i;
 523         unsigned long            flags;
 524
 525         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 526
 527         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 528                 LASSERT(!kiblnd_peer_idle(peer_ni));
 529
 530                 if (peer_ni->ibp_ni != ni)
 531                         continue;
 532
 533                 if (peer_ni->ibp_nid == nid)
 534                         kiblnd_dump_peer_debug_info(peer_ni);
 535
 536                 if (index-- > 0)
 537                         continue;
 538
 539                 *nidp = peer_ni->ibp_nid;
 540                 *count = kref_read(&peer_ni->ibp_kref);
 541
 542                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 543                 return 0;
 544         }
 545
 546         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 547         return -ENOENT;
 548 }
 549
 550 static void
 551 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
 552 {
 553         struct kib_conn *cnxt;
 554         struct kib_conn *conn;
 555
 556         if (list_empty(&peer_ni->ibp_conns)) {
 557                 kiblnd_unlink_peer_locked(peer_ni);
 558         } else {
 559                 list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
 560                                          ibc_list)
 561                         kiblnd_close_conn_locked(conn, 0);
 562                 /* NB closing peer_ni's last conn unlinked it. */
 563         }
 564         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 565          * last ref on it. */
 566 }
 567
 568 static int
 569 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 570 {
 571         LIST_HEAD(zombies);
 572         struct hlist_node *pnxt;
 573         struct kib_peer_ni *peer_ni;
 574         int lo;
 575         int hi;
 576         int i;
 577         unsigned long flags;
 578         int rc = -ENOENT;
 579
 580         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 581
 582         if (nid != LNET_NID_ANY) {
 583                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
 584                 hi = lo;
 585         } else {
 586                 lo = 0;
 587                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
 588         }
 589
 590         for (i = lo; i <= hi; i++) {
 591                 hlist_for_each_entry_safe(peer_ni, pnxt,
 592                                           &kiblnd_data.kib_peers[i], ibp_list) {
 593                         LASSERT(!kiblnd_peer_idle(peer_ni));
 594
 595                         if (peer_ni->ibp_ni != ni)
 596                                 continue;
 597
 598                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 599                                 continue;
 600
 601                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 602                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 603
 604                                 list_splice_init(&peer_ni->ibp_tx_queue,
 605                                                  &zombies);
 606                         }
 607
 608                         kiblnd_del_peer_locked(peer_ni);
 609                         rc = 0;         /* matched something */
 610                 }
 611         }
 612
 613         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 614
 615         kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
 616
 617         return rc;
 618 }
 619
 620 static struct kib_conn *
 621 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 622 {
 623         struct kib_peer_ni *peer_ni;
 624         struct kib_conn *conn;
 625         int i;
 626         unsigned long flags;
 627
 628         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 629
 630         hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list) {
 631                 LASSERT(!kiblnd_peer_idle(peer_ni));
 632
 633                 if (peer_ni->ibp_ni != ni)
 634                         continue;
 635
 636                 list_for_each_entry(conn, &peer_ni->ibp_conns,
 637                                     ibc_list) {
 638                         if (index-- > 0)
 639                                 continue;
 640
 641                         kiblnd_conn_addref(conn);
 642                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 643                                                flags);
 644                         return conn;
 645                 }
 646         }
 647
 648         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 649         return NULL;
 650 }
 651
 652 static void
 653 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 654 {
 655         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 656         if (cmid->route.path_rec == NULL)
 657                 return;
 658
 659         if (*kiblnd_tunables.kib_ib_mtu)
 660                 cmid->route.path_rec->mtu =
 661                         ib_mtu_int_to_enum(*kiblnd_tunables.kib_ib_mtu);
 662 }
 663
 664 static int
 665 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 666 {
 667         cpumask_var_t   *mask;
 668         int             vectors;
 669         int             off;
 670         int             i;
 671         lnet_nid_t      ibp_nid;
 672
 673         vectors = conn->ibc_cmid->device->num_comp_vectors;
 674         if (vectors <= 1)
 675                 return 0;
 676
 677         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 678
 679         /* hash NID to CPU id in this partition... when targeting a single peer
 680          * with multiple QPs, to engage more cores in CQ processing to a single
 681          * peer, use ibp_nconns to salt the value the comp_vector value
 682          */
 683         ibp_nid = conn->ibc_peer->ibp_nid +
 684                 atomic_read(&conn->ibc_peer->ibp_nconns);
 685         off = do_div(ibp_nid, cpumask_weight(*mask));
 686         for_each_cpu(i, *mask) {
 687                 if (off-- == 0)
 688                         return i % vectors;
 689         }
 690
 691         LBUG();
 692         return 1;
 693 }
 694
 695 /*
 696  * Get the scheduler bound to this CPT. If the scheduler has no
 697  * threads, which means that the CPT has no CPUs, then grab the
 698  * next scheduler that we can use.
 699  *
 700  * This case would be triggered if a NUMA node is configured with
 701  * no associated CPUs.
 702  */
 703 static struct kib_sched_info *
 704 kiblnd_get_scheduler(int cpt)
 705 {
 706         struct kib_sched_info *sched;
 707         int i;
 708
 709         sched = kiblnd_data.kib_scheds[cpt];
 710
 711         if (sched->ibs_nthreads > 0)
 712                 return sched;
 713
 714         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 715                 if (sched->ibs_nthreads > 0) {
 716                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 717                                         cpt, sched->ibs_cpt);
 718                         return sched;
 719                 }
 720         }
 721
 722         return NULL;
 723 }
 724
 725 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 726 {
 727         /*
 728          * One WR for the LNet message
 729          * And ibc_max_frags for the transfer WRs
 730          */
 731         int ret;
 732         int multiplier = 1 + conn->ibc_max_frags;
 733
 734         /* FastReg needs two extra WRs for map and invalidate */
 735         if (IS_FAST_REG_DEV(conn->ibc_hdev->ibh_dev))
 736                 multiplier += 2;
 737
 738         /* account for a maximum of ibc_queue_depth in-flight transfers */
 739         ret = multiplier * conn->ibc_queue_depth;
 740
 741         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 742                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 743                        "request size %d larger than maximum %d device "
 744                        "can handle\n", conn->ibc_queue_depth, ret,
 745                        conn->ibc_hdev->ibh_max_qp_wr);
 746                 conn->ibc_queue_depth =
 747                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 748         }
 749
 750         /* don't go beyond the maximum the device can handle */
 751         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 752 }
 753
 754 struct kib_conn *
 755 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 756                    int state, int version)
 757 {
 758         /* CAVEAT EMPTOR:
 759          * If the new conn is created successfully it takes over the caller's
 760          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 761          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 762          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 763          * to destroy 'cmid' here since I'm called from the CM which still has
 764          * its ref on 'cmid'). */
 765         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 766         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 767         struct kib_dev *dev;
 768         struct ib_qp_init_attr init_qp_attr = {};
 769         struct kib_sched_info   *sched;
 770 #ifdef HAVE_IB_CQ_INIT_ATTR
 771         struct ib_cq_init_attr  cq_attr = {};
 772 #endif
 773         struct kib_conn *conn;
 774         struct ib_cq            *cq;
 775         unsigned long           flags;
 776         int                     cpt;
 777         int                     rc;
 778         int                     i;
 779
 780         LASSERT(net != NULL);
 781         LASSERT(!in_interrupt());
 782
 783         dev = net->ibn_dev;
 784
 785         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 786         sched = kiblnd_get_scheduler(cpt);
 787
 788         if (sched == NULL) {
 789                 CERROR("no schedulers available. node is unhealthy\n");
 790                 goto failed_0;
 791         }
 792
 793         /*
 794          * The cpt might have changed if we ended up selecting a non cpt
 795          * native scheduler. So use the scheduler's cpt instead.
 796          */
 797         cpt = sched->ibs_cpt;
 798
 799         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 800         if (conn == NULL) {
 801                 CERROR("Can't allocate connection for %s\n",
 802                        libcfs_nid2str(peer_ni->ibp_nid));
 803                 goto failed_0;
 804         }
 805
 806         conn->ibc_state = IBLND_CONN_INIT;
 807         conn->ibc_version = version;
 808         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 809         cmid->context = conn;                   /* for future CM callbacks */
 810         conn->ibc_cmid = cmid;
 811         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 812         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 813         conn->ibc_rxs = NULL;
 814         conn->ibc_rx_pages = NULL;
 815
 816         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 817         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 818         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 819         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 820         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 821         INIT_LIST_HEAD(&conn->ibc_active_txs);
 822         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 823         spin_lock_init(&conn->ibc_lock);
 824
 825         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 826                          sizeof(*conn->ibc_connvars));
 827         if (conn->ibc_connvars == NULL) {
 828                 CERROR("Can't allocate in-progress connection state\n");
 829                 goto failed_2;
 830         }
 831
 832         write_lock_irqsave(glock, flags);
 833         if (dev->ibd_failover) {
 834                 write_unlock_irqrestore(glock, flags);
 835                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 836                 goto failed_2;
 837         }
 838
 839         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 840                 /* wakeup failover thread and teardown connection */
 841                 if (kiblnd_dev_can_failover(dev)) {
 842                         list_add_tail(&dev->ibd_fail_list,
 843                                       &kiblnd_data.kib_failed_devs);
 844                         wake_up(&kiblnd_data.kib_failover_waitq);
 845                 }
 846
 847                 write_unlock_irqrestore(glock, flags);
 848                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 849                        cmid->device->name, dev->ibd_ifname);
 850                 goto failed_2;
 851         }
 852
 853         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 854         conn->ibc_hdev = dev->ibd_hdev;
 855
 856         kiblnd_setup_mtu_locked(cmid);
 857
 858         write_unlock_irqrestore(glock, flags);
 859
 860 #ifdef HAVE_IB_CQ_INIT_ATTR
 861         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 862         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 863         cq = ib_create_cq(cmid->device,
 864                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 865                           &cq_attr);
 866 #else
 867         cq = ib_create_cq(cmid->device,
 868                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 869                           IBLND_CQ_ENTRIES(conn),
 870                           kiblnd_get_completion_vector(conn, cpt));
 871 #endif
 872         if (IS_ERR(cq)) {
 873                 /*
 874                  * on MLX-5 (possibly MLX-4 as well) this error could be
 875                  * hit if the concurrent_sends and/or peer_tx_credits is set
 876                  * too high. Or due to an MLX-5 bug which tries to
 877                  * allocate 256kb via kmalloc for WR cookie array
 878                  */
 879                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 880                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 881                 goto failed_2;
 882         }
 883
 884         conn->ibc_cq = cq;
 885
 886         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 887         if (rc != 0) {
 888                 CERROR("Can't request completion notification: %d\n", rc);
 889                 goto failed_2;
 890         }
 891
 892         init_qp_attr.event_handler = kiblnd_qp_event;
 893         init_qp_attr.qp_context = conn;
 894         init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 895         init_qp_attr.cap.max_recv_sge = 1;
 896         init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 897         init_qp_attr.qp_type = IB_QPT_RC;
 898         init_qp_attr.send_cq = cq;
 899         init_qp_attr.recv_cq = cq;
 900
 901         if (peer_ni->ibp_queue_depth_mod &&
 902             peer_ni->ibp_queue_depth_mod < peer_ni->ibp_queue_depth) {
 903                 conn->ibc_queue_depth = peer_ni->ibp_queue_depth_mod;
 904                 CDEBUG(D_NET, "Use reduced queue depth %u (from %u)\n",
 905                        peer_ni->ibp_queue_depth_mod,
 906                        peer_ni->ibp_queue_depth);
 907         }
 908
 909         do {
 910                 /* kiblnd_send_wrs() can change the connection's queue depth if
 911                  * the maximum work requests for the device is maxed out
 912                  */
 913                 init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
 914                 init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
 915                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd,
 916                                     &init_qp_attr);
 917                 if (rc != -ENOMEM || conn->ibc_queue_depth < 2)
 918                         break;
 919                 conn->ibc_queue_depth--;
 920         } while (rc);
 921
 922         if (rc) {
 923                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 924                        "send_sge: %d, recv_sge: %d\n",
 925                        rc, init_qp_attr.cap.max_send_wr,
 926                        init_qp_attr.cap.max_recv_wr,
 927                        init_qp_attr.cap.max_send_sge,
 928                        init_qp_attr.cap.max_recv_sge);
 929                 goto failed_2;
 930         }
 931
 932         conn->ibc_sched = sched;
 933
 934         if (!peer_ni->ibp_queue_depth_mod &&
 935             conn->ibc_queue_depth != peer_ni->ibp_queue_depth) {
 936                 CWARN("peer %s - queue depth reduced from %u to %u"
 937                       "  to allow for qp creation\n",
 938                       libcfs_nid2str(peer_ni->ibp_nid),
 939                       peer_ni->ibp_queue_depth,
 940                       conn->ibc_queue_depth);
 941                 peer_ni->ibp_queue_depth_mod = conn->ibc_queue_depth;
 942         }
 943
 944         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 945                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 946         if (conn->ibc_rxs == NULL) {
 947                 CERROR("Cannot allocate RX buffers\n");
 948                 goto failed_2;
 949         }
 950
 951         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 952                                 IBLND_RX_MSG_PAGES(conn));
 953         if (rc != 0)
 954                 goto failed_2;
 955
 956         kiblnd_map_rx_descs(conn);
 957
 958         /* 1 ref for caller and each rxmsg */
 959         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 960         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 961
 962         /* post receives */
 963         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 964                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 965                 if (rc != 0) {
 966                         CERROR("Can't post rxmsg: %d\n", rc);
 967
 968                         /* Make posted receives complete */
 969                         kiblnd_abort_receives(conn);
 970
 971                         /* correct # of posted buffers
 972                          * NB locking needed now I'm racing with completion */
 973                         spin_lock_irqsave(&sched->ibs_lock, flags);
 974                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 975                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 976
 977                         /* cmid will be destroyed by CM(ofed) after cm_callback
 978                          * returned, so we can't refer it anymore
 979                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 980                         rdma_destroy_qp(conn->ibc_cmid);
 981                         conn->ibc_cmid = NULL;
 982
 983                         /* Drop my own and unused rxbuffer refcounts */
 984                         while (i++ <= IBLND_RX_MSGS(conn))
 985                                 kiblnd_conn_decref(conn);
 986
 987                         return NULL;
 988                 }
 989         }
 990
 991         /* Init successful! */
 992         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 993                  state == IBLND_CONN_PASSIVE_WAIT);
 994         conn->ibc_state = state;
 995
 996         /* 1 more conn */
 997         atomic_inc(&peer_ni->ibp_nconns);
 998         atomic_inc(&net->ibn_nconns);
 999         return conn;
1000
1001  failed_2:
1002         kiblnd_destroy_conn(conn);
1003         LIBCFS_FREE(conn, sizeof(*conn));
1004  failed_0:
1005         return NULL;
1006 }
1007
1008 void
1009 kiblnd_destroy_conn(struct kib_conn *conn)
1010 {
1011         struct rdma_cm_id *cmid = conn->ibc_cmid;
1012         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1013
1014         LASSERT (!in_interrupt());
1015         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1016         LASSERT(list_empty(&conn->ibc_early_rxs));
1017         LASSERT(list_empty(&conn->ibc_tx_noops));
1018         LASSERT(list_empty(&conn->ibc_tx_queue));
1019         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1020         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1021         LASSERT(list_empty(&conn->ibc_active_txs));
1022         LASSERT (conn->ibc_noops_posted == 0);
1023         LASSERT (conn->ibc_nsends_posted == 0);
1024
1025         switch (conn->ibc_state) {
1026         default:
1027                 /* conn must be completely disengaged from the network */
1028                 LBUG();
1029
1030         case IBLND_CONN_DISCONNECTED:
1031                 /* connvars should have been freed already */
1032                 LASSERT (conn->ibc_connvars == NULL);
1033                 break;
1034
1035         case IBLND_CONN_INIT:
1036                 break;
1037         }
1038
1039         /* conn->ibc_cmid might be destroyed by CM already */
1040         if (cmid != NULL && cmid->qp != NULL)
1041                 rdma_destroy_qp(cmid);
1042
1043         if (conn->ibc_cq)
1044                 ib_destroy_cq(conn->ibc_cq);
1045
1046         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1047                            LNET_MSG_STATUS_OK);
1048
1049         if (conn->ibc_rx_pages != NULL)
1050                 kiblnd_unmap_rx_descs(conn);
1051
1052         if (conn->ibc_rxs != NULL)
1053                 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1054
1055         if (conn->ibc_connvars != NULL)
1056                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1057
1058         if (conn->ibc_hdev != NULL)
1059                 kiblnd_hdev_decref(conn->ibc_hdev);
1060
1061         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1062         if (conn->ibc_state != IBLND_CONN_INIT) {
1063                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1064
1065                 kiblnd_peer_decref(peer_ni);
1066                 rdma_destroy_id(cmid);
1067                 atomic_dec(&peer_ni->ibp_nconns);
1068                 atomic_dec(&net->ibn_nconns);
1069         }
1070 }
1071
1072 int
1073 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1074 {
1075         struct kib_conn *conn;
1076         struct kib_conn *cnxt;
1077         int count = 0;
1078
1079         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1080                                  ibc_list) {
1081                 CDEBUG(D_NET, "Closing conn -> %s, "
1082                               "version: %x, reason: %d\n",
1083                        libcfs_nid2str(peer_ni->ibp_nid),
1084                        conn->ibc_version, why);
1085
1086                 kiblnd_close_conn_locked(conn, why);
1087                 count++;
1088         }
1089
1090         return count;
1091 }
1092
1093 int
1094 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1095                                 int version, __u64 incarnation)
1096 {
1097         struct kib_conn *conn;
1098         struct kib_conn *cnxt;
1099         int count = 0;
1100
1101         list_for_each_entry_safe(conn, cnxt, &peer_ni->ibp_conns,
1102                                  ibc_list) {
1103                 if (conn->ibc_version     == version &&
1104                     conn->ibc_incarnation == incarnation)
1105                         continue;
1106
1107                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1108                               "incarnation:%#llx(%x, %#llx)\n",
1109                        libcfs_nid2str(peer_ni->ibp_nid),
1110                        conn->ibc_version, conn->ibc_incarnation,
1111                        version, incarnation);
1112
1113                 kiblnd_close_conn_locked(conn, -ESTALE);
1114                 count++;
1115         }
1116
1117         return count;
1118 }
1119
1120 static int
1121 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1122 {
1123         struct kib_peer_ni *peer_ni;
1124         struct hlist_node *pnxt;
1125         int lo;
1126         int hi;
1127         int i;
1128         unsigned long flags;
1129         int count = 0;
1130
1131         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1132
1133         if (nid != LNET_NID_ANY) {
1134                 lo = hash_min(nid, HASH_BITS(kiblnd_data.kib_peers));
1135                 hi = lo;
1136         } else {
1137                 lo = 0;
1138                 hi = HASH_SIZE(kiblnd_data.kib_peers) - 1;
1139         }
1140
1141         for (i = lo; i <= hi; i++) {
1142                 hlist_for_each_entry_safe(peer_ni, pnxt,
1143                                           &kiblnd_data.kib_peers[i], ibp_list) {
1144                         LASSERT(!kiblnd_peer_idle(peer_ni));
1145
1146                         if (peer_ni->ibp_ni != ni)
1147                                 continue;
1148
1149                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1150                                 continue;
1151
1152                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1153                 }
1154         }
1155
1156         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1157
1158         /* wildcards always succeed */
1159         if (nid == LNET_NID_ANY)
1160                 return 0;
1161
1162         return (count == 0) ? -ENOENT : 0;
1163 }
1164
1165 static int
1166 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1167 {
1168         struct libcfs_ioctl_data *data = arg;
1169         int                       rc = -EINVAL;
1170
1171         switch(cmd) {
1172         case IOC_LIBCFS_GET_PEER: {
1173                 lnet_nid_t   nid = 0;
1174                 int          count = 0;
1175
1176                 rc = kiblnd_get_peer_info(ni, data->ioc_nid, data->ioc_count,
1177                                           &nid, &count);
1178                 data->ioc_nid    = nid;
1179                 data->ioc_count  = count;
1180                 break;
1181         }
1182
1183         case IOC_LIBCFS_DEL_PEER: {
1184                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1185                 break;
1186         }
1187         case IOC_LIBCFS_GET_CONN: {
1188                 struct kib_conn *conn;
1189
1190                 rc = 0;
1191                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1192                 if (conn == NULL) {
1193                         rc = -ENOENT;
1194                         break;
1195                 }
1196
1197                 LASSERT(conn->ibc_cmid != NULL);
1198                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1199                 if (conn->ibc_cmid->route.path_rec == NULL)
1200                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1201                 else
1202                         data->ioc_u32[0] =
1203                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1204                 kiblnd_conn_decref(conn);
1205                 break;
1206         }
1207         case IOC_LIBCFS_CLOSE_CONNECTION: {
1208                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1209                 break;
1210         }
1211
1212         default:
1213                 break;
1214         }
1215
1216         return rc;
1217 }
1218
1219 static const struct ln_key_list kiblnd_tunables_keys = {
1220         .lkl_maxattr                    = LNET_NET_O2IBLND_TUNABLES_ATTR_MAX,
1221         .lkl_list                       = {
1222                 [LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS]  = {
1223                         .lkp_value      = "peercredits_hiw",
1224                         .lkp_data_type  = NLA_U32
1225                 },
1226                 [LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND]  = {
1227                         .lkp_value      = "map_on_demand",
1228                         .lkp_data_type  = NLA_FLAG
1229                 },
1230                 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS]  = {
1231                         .lkp_value      = "concurrent_sends",
1232                         .lkp_data_type  = NLA_U32
1233                 },
1234                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE]  = {
1235                         .lkp_value      = "fmr_pool_size",
1236                         .lkp_data_type  = NLA_U32
1237                 },
1238                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER]  = {
1239                         .lkp_value      = "fmr_flush_trigger",
1240                         .lkp_data_type  = NLA_U32
1241                 },
1242                 [LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE]  = {
1243                         .lkp_value      = "fmr_cache",
1244                         .lkp_data_type  = NLA_U32
1245                 },
1246                 [LNET_NET_O2IBLND_TUNABLES_ATTR_NTX]  = {
1247                         .lkp_value      = "ntx",
1248                         .lkp_data_type  = NLA_U16
1249                 },
1250                 [LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER]  = {
1251                         .lkp_value      = "conns_per_peer",
1252                         .lkp_data_type  = NLA_U16
1253                 },
1254         },
1255 };
1256
1257 static int
1258 kiblnd_nl_set(int cmd, struct nlattr *attr, int type, void *data)
1259 {
1260         struct lnet_lnd_tunables *tunables = data;
1261         s64 num;
1262
1263         if (cmd != LNET_CMD_NETS)
1264                 return -EOPNOTSUPP;
1265
1266         if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
1267                 return -EINVAL;
1268
1269         switch (type) {
1270         case LNET_NET_O2IBLND_TUNABLES_ATTR_HIW_PEER_CREDITS:
1271                 tunables->lnd_tun_u.lnd_o2ib.lnd_peercredits_hiw = nla_get_s64(attr);
1272                 break;
1273         case LNET_NET_O2IBLND_TUNABLES_ATTR_MAP_ON_DEMAND:
1274                 tunables->lnd_tun_u.lnd_o2ib.lnd_map_on_demand = nla_get_s64(attr);
1275                 break;
1276         case LNET_NET_O2IBLND_TUNABLES_ATTR_CONCURRENT_SENDS:
1277                 tunables->lnd_tun_u.lnd_o2ib.lnd_concurrent_sends = nla_get_s64(attr);
1278                 break;
1279         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_POOL_SIZE:
1280                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_pool_size = nla_get_s64(attr);
1281                 break;
1282         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_FLUSH_TRIGGER:
1283                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_flush_trigger = nla_get_s64(attr);
1284                 break;
1285         case LNET_NET_O2IBLND_TUNABLES_ATTR_FMR_CACHE:
1286                 tunables->lnd_tun_u.lnd_o2ib.lnd_fmr_cache = nla_get_s64(attr);
1287                 break;
1288         case LNET_NET_O2IBLND_TUNABLES_ATTR_NTX:
1289                 tunables->lnd_tun_u.lnd_o2ib.lnd_ntx = nla_get_s64(attr);
1290                 break;
1291         case LNET_NET_O2IBLND_TUNABLES_ATTR_CONNS_PER_PEER:
1292                 num = nla_get_s64(attr);
1293                 clamp_t(s64, num, 1, 127);
1294                 tunables->lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = num;
1295                 fallthrough;
1296         default:
1297                 break;
1298         }
1299
1300         return 0;
1301 }
1302
1303 static void
1304 kiblnd_free_pages(struct kib_pages *p)
1305 {
1306         int     npages = p->ibp_npages;
1307         int     i;
1308
1309         for (i = 0; i < npages; i++) {
1310                 if (p->ibp_pages[i] != NULL)
1311                         __free_page(p->ibp_pages[i]);
1312         }
1313
1314         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1315 }
1316
1317 int
1318 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1319 {
1320         struct kib_pages *p;
1321         int i;
1322
1323         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1324                          offsetof(struct kib_pages, ibp_pages[npages]));
1325         if (p == NULL) {
1326                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1327                 return -ENOMEM;
1328         }
1329
1330         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1331         p->ibp_npages = npages;
1332
1333         for (i = 0; i < npages; i++) {
1334                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1335                                                      GFP_NOFS);
1336                 if (p->ibp_pages[i] == NULL) {
1337                         CERROR("Can't allocate page %d of %d\n", i, npages);
1338                         kiblnd_free_pages(p);
1339                         return -ENOMEM;
1340                 }
1341         }
1342
1343         *pp = p;
1344         return 0;
1345 }
1346
1347 void
1348 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1349 {
1350         struct kib_rx *rx;
1351         int       i;
1352
1353         LASSERT (conn->ibc_rxs != NULL);
1354         LASSERT (conn->ibc_hdev != NULL);
1355
1356         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1357                 rx = &conn->ibc_rxs[i];
1358
1359                 LASSERT(rx->rx_nob >= 0); /* not posted */
1360
1361                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1362                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1363                                                           rx->rx_msgaddr),
1364                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1365         }
1366
1367         kiblnd_free_pages(conn->ibc_rx_pages);
1368
1369         conn->ibc_rx_pages = NULL;
1370 }
1371
1372 void
1373 kiblnd_map_rx_descs(struct kib_conn *conn)
1374 {
1375         struct kib_rx *rx;
1376         struct page    *pg;
1377         int             pg_off;
1378         int             ipg;
1379         int             i;
1380
1381         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1382                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1383                 rx = &conn->ibc_rxs[i];
1384
1385                 rx->rx_conn = conn;
1386                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1387
1388                 rx->rx_msgaddr =
1389                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1390                                               rx->rx_msg, IBLND_MSG_SIZE,
1391                                               DMA_FROM_DEVICE);
1392                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1393                                                   rx->rx_msgaddr));
1394                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1395
1396                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1397                        i, rx->rx_msg, rx->rx_msgaddr,
1398                        (__u64)(page_to_phys(pg) + pg_off));
1399
1400                 pg_off += IBLND_MSG_SIZE;
1401                 LASSERT(pg_off <= PAGE_SIZE);
1402
1403                 if (pg_off == PAGE_SIZE) {
1404                         pg_off = 0;
1405                         ipg++;
1406                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1407                 }
1408         }
1409 }
1410
1411 static void
1412 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1413 {
1414         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1415         struct kib_tx *tx;
1416         int i;
1417
1418         LASSERT (tpo->tpo_pool.po_allocated == 0);
1419
1420         if (hdev == NULL)
1421                 return;
1422
1423         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1424                 tx = &tpo->tpo_tx_descs[i];
1425                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1426                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1427                                                           tx->tx_msgaddr),
1428                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1429         }
1430
1431         kiblnd_hdev_decref(hdev);
1432         tpo->tpo_hdev = NULL;
1433 }
1434
1435 static struct kib_hca_dev *
1436 kiblnd_current_hdev(struct kib_dev *dev)
1437 {
1438         struct kib_hca_dev *hdev;
1439         unsigned long  flags;
1440         int            i = 0;
1441
1442         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1443         while (dev->ibd_failover) {
1444                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1445                 if (i++ % 50 == 0)
1446                         CDEBUG(D_NET, "%s: Wait for failover\n",
1447                                dev->ibd_ifname);
1448                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1449
1450                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1451         }
1452
1453         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1454         hdev = dev->ibd_hdev;
1455
1456         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1457
1458         return hdev;
1459 }
1460
1461 static void
1462 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1463 {
1464         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1465         struct kib_pool *pool = &tpo->tpo_pool;
1466         struct kib_net      *net   = pool->po_owner->ps_net;
1467         struct kib_dev *dev;
1468         struct page *page;
1469         struct kib_tx *tx;
1470         int             page_offset;
1471         int             ipage;
1472         int             i;
1473
1474         LASSERT (net != NULL);
1475
1476         dev = net->ibn_dev;
1477
1478         /* pre-mapped messages are not bigger than 1 page */
1479         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1480
1481         /* No fancy arithmetic when we do the buffer calculations */
1482         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1483
1484         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1485
1486         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1487                 page = txpgs->ibp_pages[ipage];
1488                 tx = &tpo->tpo_tx_descs[i];
1489
1490                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1491                                                 page_offset);
1492
1493                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1494                                                        tx->tx_msg,
1495                                                        IBLND_MSG_SIZE,
1496                                                        DMA_TO_DEVICE);
1497                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1498                                                   tx->tx_msgaddr));
1499                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1500
1501                 list_add(&tx->tx_list, &pool->po_free_list);
1502
1503                 page_offset += IBLND_MSG_SIZE;
1504                 LASSERT(page_offset <= PAGE_SIZE);
1505
1506                 if (page_offset == PAGE_SIZE) {
1507                         page_offset = 0;
1508                         ipage++;
1509                         LASSERT(ipage <= txpgs->ibp_npages);
1510                 }
1511         }
1512 }
1513
1514 static void
1515 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1516 {
1517         LASSERT(fpo->fpo_map_count == 0);
1518
1519 #ifdef HAVE_FMR_POOL_API
1520         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1521                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1522         } else
1523 #endif /* HAVE_FMR_POOL_API */
1524         {
1525                 struct kib_fast_reg_descriptor *frd, *tmp;
1526                 int i = 0;
1527
1528                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1529                                          frd_list) {
1530                         list_del(&frd->frd_list);
1531 #ifndef HAVE_IB_MAP_MR_SG
1532                         ib_free_fast_reg_page_list(frd->frd_frpl);
1533 #endif
1534                         ib_dereg_mr(frd->frd_mr);
1535                         LIBCFS_FREE(frd, sizeof(*frd));
1536                         i++;
1537                 }
1538                 if (i < fpo->fast_reg.fpo_pool_size)
1539                         CERROR("FastReg pool still has %d regions registered\n",
1540                                 fpo->fast_reg.fpo_pool_size - i);
1541         }
1542
1543         if (fpo->fpo_hdev)
1544                 kiblnd_hdev_decref(fpo->fpo_hdev);
1545
1546         LIBCFS_FREE(fpo, sizeof(*fpo));
1547 }
1548
1549 static void
1550 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1551 {
1552         struct kib_fmr_pool *fpo, *tmp;
1553
1554         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1555                 list_del(&fpo->fpo_list);
1556                 kiblnd_destroy_fmr_pool(fpo);
1557         }
1558 }
1559
1560 static int
1561 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1562                      int ncpts)
1563 {
1564         int size = tunables->lnd_fmr_pool_size / ncpts;
1565
1566         return max(IBLND_FMR_POOL, size);
1567 }
1568
1569 static int
1570 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1571                          int ncpts)
1572 {
1573         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1574
1575         return max(IBLND_FMR_POOL_FLUSH, size);
1576 }
1577
1578 #ifdef HAVE_FMR_POOL_API
1579 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1580                                  struct kib_fmr_pool *fpo)
1581 {
1582         struct ib_fmr_pool_param param = {
1583                 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1584                 .page_shift        = PAGE_SHIFT,
1585                 .access            = (IB_ACCESS_LOCAL_WRITE |
1586                                       IB_ACCESS_REMOTE_WRITE),
1587                 .pool_size         = fps->fps_pool_size,
1588                 .dirty_watermark   = fps->fps_flush_trigger,
1589                 .flush_function    = NULL,
1590                 .flush_arg         = NULL,
1591                 .cache             = !!fps->fps_cache };
1592         int rc = 0;
1593
1594         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1595                                                    &param);
1596         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1597                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1598                 if (rc != -ENOSYS)
1599                         CERROR("Failed to create FMR pool: %d\n", rc);
1600                 else
1601                         CERROR("FMRs are not supported\n");
1602         }
1603         fpo->fpo_is_fmr = true;
1604
1605         return rc;
1606 }
1607 #endif /* HAVE_FMR_POOL_API */
1608
1609 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1610                                   struct kib_fmr_pool *fpo,
1611                                   enum kib_dev_caps dev_caps)
1612 {
1613         struct kib_fast_reg_descriptor *frd, *tmp;
1614         int i, rc;
1615
1616 #ifdef HAVE_FMR_POOL_API
1617         fpo->fpo_is_fmr = false;
1618 #endif
1619
1620         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1621         fpo->fast_reg.fpo_pool_size = 0;
1622         for (i = 0; i < fps->fps_pool_size; i++) {
1623                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1624                                  sizeof(*frd));
1625                 if (!frd) {
1626                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1627                         rc = -ENOMEM;
1628                         goto out;
1629                 }
1630                 frd->frd_mr = NULL;
1631
1632 #ifndef HAVE_IB_MAP_MR_SG
1633                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1634                                                             IBLND_MAX_RDMA_FRAGS);
1635                 if (IS_ERR(frd->frd_frpl)) {
1636                         rc = PTR_ERR(frd->frd_frpl);
1637                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1638                                 rc);
1639                         frd->frd_frpl = NULL;
1640                         goto out_middle;
1641                 }
1642 #endif
1643
1644 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1645                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1646                                                    IBLND_MAX_RDMA_FRAGS);
1647 #else
1648                 /*
1649                  * it is expected to get here if this is an MLX-5 card.
1650                  * MLX-4 cards will always use FMR and MLX-5 cards will
1651                  * always use fast_reg. It turns out that some MLX-5 cards
1652                  * (possibly due to older FW versions) do not natively support
1653                  * gaps. So we will need to track them here.
1654                  */
1655                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1656 #ifdef IB_MR_TYPE_SG_GAPS
1657                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1658                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1659                                                 IB_MR_TYPE_SG_GAPS :
1660                                                 IB_MR_TYPE_MEM_REG,
1661 #else
1662                                                 IB_MR_TYPE_MEM_REG,
1663 #endif
1664                                           IBLND_MAX_RDMA_FRAGS);
1665                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1666                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1667                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1668 #endif
1669                 if (IS_ERR(frd->frd_mr)) {
1670                         rc = PTR_ERR(frd->frd_mr);
1671                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1672                         frd->frd_mr = NULL;
1673                         goto out_middle;
1674                 }
1675
1676                 /* There appears to be a bug in MLX5 code where you must
1677                  * invalidate the rkey of a new FastReg pool before first
1678                  * using it. Thus, I am marking the FRD invalid here. */
1679                 frd->frd_valid = false;
1680
1681                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1682                 fpo->fast_reg.fpo_pool_size++;
1683         }
1684
1685         return 0;
1686
1687 out_middle:
1688         if (frd->frd_mr)
1689                 ib_dereg_mr(frd->frd_mr);
1690 #ifndef HAVE_IB_MAP_MR_SG
1691         if (frd->frd_frpl)
1692                 ib_free_fast_reg_page_list(frd->frd_frpl);
1693 #endif
1694         LIBCFS_FREE(frd, sizeof(*frd));
1695
1696 out:
1697         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1698                                  frd_list) {
1699                 list_del(&frd->frd_list);
1700 #ifndef HAVE_IB_MAP_MR_SG
1701                 ib_free_fast_reg_page_list(frd->frd_frpl);
1702 #endif
1703                 ib_dereg_mr(frd->frd_mr);
1704                 LIBCFS_FREE(frd, sizeof(*frd));
1705         }
1706
1707         return rc;
1708 }
1709
1710 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1711                                   struct kib_fmr_pool **pp_fpo)
1712 {
1713         struct kib_dev *dev = fps->fps_net->ibn_dev;
1714         struct kib_fmr_pool *fpo;
1715         int rc;
1716
1717         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1718         if (!fpo) {
1719                 return -ENOMEM;
1720         }
1721         memset(fpo, 0, sizeof(*fpo));
1722
1723         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1724
1725 #ifdef HAVE_FMR_POOL_API
1726         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1727                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1728         else
1729 #endif /* HAVE_FMR_POOL_API */
1730                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1731         if (rc)
1732                 goto out_fpo;
1733
1734         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1735         fpo->fpo_owner = fps;
1736         *pp_fpo = fpo;
1737
1738         return 0;
1739
1740 out_fpo:
1741         kiblnd_hdev_decref(fpo->fpo_hdev);
1742         LIBCFS_FREE(fpo, sizeof(*fpo));
1743         return rc;
1744 }
1745
1746 static void
1747 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1748 {
1749         struct kib_fmr_pool *fpo;
1750
1751         if (fps->fps_net == NULL) /* intialized? */
1752                 return;
1753
1754         spin_lock(&fps->fps_lock);
1755
1756         while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1757                                                struct kib_fmr_pool,
1758                                                fpo_list)) != NULL) {
1759                 fpo->fpo_failed = 1;
1760                 if (fpo->fpo_map_count == 0)
1761                         list_move(&fpo->fpo_list, zombies);
1762                 else
1763                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1764         }
1765
1766         spin_unlock(&fps->fps_lock);
1767 }
1768
1769 static void
1770 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1771 {
1772         if (fps->fps_net != NULL) { /* initialized? */
1773                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1774                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1775         }
1776 }
1777
1778 static int
1779 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1780                         struct kib_net *net,
1781                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1782 {
1783         struct kib_fmr_pool *fpo;
1784         int rc;
1785
1786         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1787
1788         fps->fps_net = net;
1789         fps->fps_cpt = cpt;
1790
1791         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1792         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1793         fps->fps_cache = tunables->lnd_fmr_cache;
1794
1795         spin_lock_init(&fps->fps_lock);
1796         INIT_LIST_HEAD(&fps->fps_pool_list);
1797         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1798
1799         rc = kiblnd_create_fmr_pool(fps, &fpo);
1800         if (rc == 0)
1801                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1802
1803         return rc;
1804 }
1805
1806 static int
1807 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1808 {
1809         if (fpo->fpo_map_count != 0) /* still in use */
1810                 return 0;
1811         if (fpo->fpo_failed)
1812                 return 1;
1813         return now >= fpo->fpo_deadline;
1814 }
1815
1816 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1817 static int
1818 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1819 {
1820         struct kib_hca_dev *hdev;
1821         __u64           *pages = tx->tx_pages;
1822         int             npages;
1823         int             size;
1824         int             i;
1825
1826         hdev = tx->tx_pool->tpo_hdev;
1827
1828         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1829                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1830                         size += hdev->ibh_page_size) {
1831                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1832                                            hdev->ibh_page_mask) + size;
1833                 }
1834         }
1835
1836         return npages;
1837 }
1838 #endif
1839
1840 void
1841 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1842 {
1843         LIST_HEAD(zombies);
1844         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1845         struct kib_fmr_poolset *fps;
1846         time64_t now = ktime_get_seconds();
1847         struct kib_fmr_pool *tmp;
1848
1849         if (!fpo)
1850                 return;
1851
1852         fps = fpo->fpo_owner;
1853
1854 #ifdef HAVE_FMR_POOL_API
1855         if (fpo->fpo_is_fmr) {
1856                 if (fmr->fmr_pfmr) {
1857                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1858                         fmr->fmr_pfmr = NULL;
1859                 }
1860
1861                 if (status) {
1862                         int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1863                         LASSERT(!rc);
1864                 }
1865         } else
1866 #endif /* HAVE_FMR_POOL_API */
1867         {
1868                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1869
1870                 if (frd) {
1871                         frd->frd_valid = false;
1872                         frd->frd_posted = false;
1873                         fmr->fmr_frd = NULL;
1874                         spin_lock(&fps->fps_lock);
1875                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1876                         spin_unlock(&fps->fps_lock);
1877                 }
1878         }
1879         fmr->fmr_pool = NULL;
1880
1881         spin_lock(&fps->fps_lock);
1882         fpo->fpo_map_count--;   /* decref the pool */
1883
1884         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1885                 /* the first pool is persistent */
1886                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1887                         continue;
1888
1889                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1890                         list_move(&fpo->fpo_list, &zombies);
1891                         fps->fps_version++;
1892                 }
1893         }
1894         spin_unlock(&fps->fps_lock);
1895
1896         if (!list_empty(&zombies))
1897                 kiblnd_destroy_fmr_pool_list(&zombies);
1898 }
1899
1900 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1901                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1902                         struct kib_fmr *fmr)
1903 {
1904         struct kib_fmr_pool *fpo;
1905         __u64 version;
1906         bool is_rx = (rd != tx->tx_rd);
1907 #ifdef HAVE_FMR_POOL_API
1908         __u64 *pages = tx->tx_pages;
1909         bool tx_pages_mapped = false;
1910         int npages = 0;
1911 #endif
1912         int rc;
1913
1914 again:
1915         spin_lock(&fps->fps_lock);
1916         version = fps->fps_version;
1917         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1918                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1919                 fpo->fpo_map_count++;
1920
1921 #ifdef HAVE_FMR_POOL_API
1922                 fmr->fmr_pfmr = NULL;
1923                 if (fpo->fpo_is_fmr) {
1924                         struct ib_pool_fmr *pfmr;
1925
1926                         spin_unlock(&fps->fps_lock);
1927
1928                         if (!tx_pages_mapped) {
1929                                 npages = kiblnd_map_tx_pages(tx, rd);
1930                                 tx_pages_mapped = true;
1931                         }
1932
1933                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1934                                                 pages, npages, iov);
1935                         if (likely(!IS_ERR(pfmr))) {
1936                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1937                                         : pfmr->fmr->lkey;
1938                                 fmr->fmr_frd  = NULL;
1939                                 fmr->fmr_pfmr = pfmr;
1940                                 fmr->fmr_pool = fpo;
1941                                 return 0;
1942                         }
1943                         rc = PTR_ERR(pfmr);
1944                 } else
1945 #endif /* HAVE_FMR_POOL_API */
1946                 {
1947                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1948                                 struct kib_fast_reg_descriptor *frd;
1949 #ifdef HAVE_IB_MAP_MR_SG
1950                                 struct ib_reg_wr *wr;
1951                                 int n;
1952 #else
1953                                 struct ib_rdma_wr *wr;
1954                                 struct ib_fast_reg_page_list *frpl;
1955 #endif
1956                                 struct ib_mr *mr;
1957
1958                                 frd = list_first_entry(
1959                                         &fpo->fast_reg.fpo_pool_list,
1960                                         struct kib_fast_reg_descriptor,
1961                                         frd_list);
1962                                 list_del(&frd->frd_list);
1963                                 spin_unlock(&fps->fps_lock);
1964
1965 #ifndef HAVE_IB_MAP_MR_SG
1966                                 frpl = frd->frd_frpl;
1967 #endif
1968                                 mr   = frd->frd_mr;
1969
1970                                 if (!frd->frd_valid) {
1971                                         struct ib_rdma_wr *inv_wr;
1972                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1973
1974                                         inv_wr = &frd->frd_inv_wr;
1975                                         memset(inv_wr, 0, sizeof(*inv_wr));
1976
1977                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1978                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1979                                         inv_wr->wr.ex.invalidate_rkey = key;
1980
1981                                         /* Bump the key */
1982                                         key = ib_inc_rkey(key);
1983                                         ib_update_fast_reg_key(mr, key);
1984                                 }
1985
1986 #ifdef HAVE_IB_MAP_MR_SG
1987 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1988                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1989                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
1990 #else
1991                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1992                                                  rd->rd_nfrags, PAGE_SIZE);
1993 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1994                                 if (unlikely(n != rd->rd_nfrags)) {
1995                                         CERROR("Failed to map mr %d/%d elements\n",
1996                                                n, rd->rd_nfrags);
1997                                         return n < 0 ? n : -EINVAL;
1998                                 }
1999
2000                                 wr = &frd->frd_fastreg_wr;
2001                                 memset(wr, 0, sizeof(*wr));
2002
2003                                 wr->wr.opcode = IB_WR_REG_MR;
2004                                 wr->wr.wr_id  = IBLND_WID_MR;
2005                                 wr->wr.num_sge = 0;
2006                                 wr->wr.send_flags = 0;
2007                                 wr->mr = mr;
2008                                 wr->key = is_rx ? mr->rkey : mr->lkey;
2009                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
2010                                               IB_ACCESS_REMOTE_WRITE);
2011 #else /* HAVE_IB_MAP_MR_SG */
2012                                 if (!tx_pages_mapped) {
2013                                         npages = kiblnd_map_tx_pages(tx, rd);
2014                                         tx_pages_mapped = true;
2015                                 }
2016
2017                                 LASSERT(npages <= frpl->max_page_list_len);
2018                                 memcpy(frpl->page_list, pages,
2019                                        sizeof(*pages) * npages);
2020
2021                                 /* Prepare FastReg WR */
2022                                 wr = &frd->frd_fastreg_wr;
2023                                 memset(wr, 0, sizeof(*wr));
2024
2025                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
2026                                 wr->wr.wr_id  = IBLND_WID_MR;
2027
2028                                 wr->wr.wr.fast_reg.iova_start = iov;
2029                                 wr->wr.wr.fast_reg.page_list  = frpl;
2030                                 wr->wr.wr.fast_reg.page_list_len = npages;
2031                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2032                                 wr->wr.wr.fast_reg.length = nob;
2033                                 wr->wr.wr.fast_reg.rkey =
2034                                         is_rx ? mr->rkey : mr->lkey;
2035                                 wr->wr.wr.fast_reg.access_flags =
2036                                         (IB_ACCESS_LOCAL_WRITE |
2037                                          IB_ACCESS_REMOTE_WRITE);
2038 #endif /* HAVE_IB_MAP_MR_SG */
2039
2040                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
2041                                 fmr->fmr_frd  = frd;
2042                                 fmr->fmr_pool = fpo;
2043                                 frd->frd_posted = false;
2044                                 return 0;
2045                         }
2046                         spin_unlock(&fps->fps_lock);
2047                         rc = -EAGAIN;
2048                 }
2049
2050                 spin_lock(&fps->fps_lock);
2051                 fpo->fpo_map_count--;
2052                 if (rc != -EAGAIN) {
2053                         spin_unlock(&fps->fps_lock);
2054                         return rc;
2055                 }
2056
2057                 /* EAGAIN and ... */
2058                 if (version != fps->fps_version) {
2059                         spin_unlock(&fps->fps_lock);
2060                         goto again;
2061                 }
2062         }
2063
2064         if (fps->fps_increasing) {
2065                 spin_unlock(&fps->fps_lock);
2066                 CDEBUG(D_NET, "Another thread is allocating new "
2067                        "FMR pool, waiting for her to complete\n");
2068                 wait_var_event(fps, !fps->fps_increasing);
2069                 goto again;
2070
2071         }
2072
2073         if (ktime_get_seconds() < fps->fps_next_retry) {
2074                 /* someone failed recently */
2075                 spin_unlock(&fps->fps_lock);
2076                 return -EAGAIN;
2077         }
2078
2079         fps->fps_increasing = 1;
2080         spin_unlock(&fps->fps_lock);
2081
2082         CDEBUG(D_NET, "Allocate new FMR pool\n");
2083         rc = kiblnd_create_fmr_pool(fps, &fpo);
2084         spin_lock(&fps->fps_lock);
2085         fps->fps_increasing = 0;
2086         wake_up_var(fps);
2087         if (rc == 0) {
2088                 fps->fps_version++;
2089                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2090         } else {
2091                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2092         }
2093         spin_unlock(&fps->fps_lock);
2094
2095         goto again;
2096 }
2097
2098 static void
2099 kiblnd_fini_pool(struct kib_pool *pool)
2100 {
2101         LASSERT(list_empty(&pool->po_free_list));
2102         LASSERT(pool->po_allocated == 0);
2103
2104         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2105 }
2106
2107 static void
2108 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2109 {
2110         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2111
2112         memset(pool, 0, sizeof(struct kib_pool));
2113         INIT_LIST_HEAD(&pool->po_free_list);
2114         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2115         pool->po_owner = ps;
2116         pool->po_size = size;
2117 }
2118
2119 static void
2120 kiblnd_destroy_pool_list(struct list_head *head)
2121 {
2122         struct kib_pool *pool;
2123
2124         while ((pool = list_first_entry_or_null(head,
2125                                                 struct kib_pool,
2126                                                 po_list)) != NULL) {
2127                 list_del(&pool->po_list);
2128
2129                 LASSERT(pool->po_owner != NULL);
2130                 pool->po_owner->ps_pool_destroy(pool);
2131         }
2132 }
2133
2134 static void
2135 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2136 {
2137         struct kib_pool *po;
2138
2139         if (ps->ps_net == NULL) /* intialized? */
2140                 return;
2141
2142         spin_lock(&ps->ps_lock);
2143         while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2144                                               struct kib_pool,
2145                                               po_list)) != NULL) {
2146                 po->po_failed = 1;
2147                 if (po->po_allocated == 0)
2148                         list_move(&po->po_list, zombies);
2149                 else
2150                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2151         }
2152         spin_unlock(&ps->ps_lock);
2153 }
2154
2155 static void
2156 kiblnd_fini_poolset(struct kib_poolset *ps)
2157 {
2158         if (ps->ps_net != NULL) { /* initialized? */
2159                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2160                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2161         }
2162 }
2163
2164 static int
2165 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2166                     struct kib_net *net, char *name, int size,
2167                     kib_ps_pool_create_t po_create,
2168                     kib_ps_pool_destroy_t po_destroy,
2169                     kib_ps_node_init_t nd_init,
2170                     kib_ps_node_fini_t nd_fini)
2171 {
2172         struct kib_pool *pool;
2173         int rc;
2174
2175         memset(ps, 0, sizeof(struct kib_poolset));
2176
2177         ps->ps_cpt          = cpt;
2178         ps->ps_net          = net;
2179         ps->ps_pool_create  = po_create;
2180         ps->ps_pool_destroy = po_destroy;
2181         ps->ps_node_init    = nd_init;
2182         ps->ps_node_fini    = nd_fini;
2183         ps->ps_pool_size    = size;
2184         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2185             >= sizeof(ps->ps_name))
2186                 return -E2BIG;
2187         spin_lock_init(&ps->ps_lock);
2188         INIT_LIST_HEAD(&ps->ps_pool_list);
2189         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2190
2191         rc = ps->ps_pool_create(ps, size, &pool);
2192         if (rc == 0)
2193                 list_add(&pool->po_list, &ps->ps_pool_list);
2194         else
2195                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2196
2197         return rc;
2198 }
2199
2200 static int
2201 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2202 {
2203         if (pool->po_allocated != 0) /* still in use */
2204                 return 0;
2205         if (pool->po_failed)
2206                 return 1;
2207         return now >= pool->po_deadline;
2208 }
2209
2210 void
2211 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2212 {
2213         LIST_HEAD(zombies);
2214         struct kib_poolset *ps = pool->po_owner;
2215         struct kib_pool *tmp;
2216         time64_t now = ktime_get_seconds();
2217
2218         spin_lock(&ps->ps_lock);
2219
2220         if (ps->ps_node_fini != NULL)
2221                 ps->ps_node_fini(pool, node);
2222
2223         LASSERT(pool->po_allocated > 0);
2224         list_add(node, &pool->po_free_list);
2225         pool->po_allocated--;
2226
2227         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2228                 /* the first pool is persistent */
2229                 if (ps->ps_pool_list.next == &pool->po_list)
2230                         continue;
2231
2232                 if (kiblnd_pool_is_idle(pool, now))
2233                         list_move(&pool->po_list, &zombies);
2234         }
2235         spin_unlock(&ps->ps_lock);
2236
2237         if (!list_empty(&zombies))
2238                 kiblnd_destroy_pool_list(&zombies);
2239 }
2240
2241 struct list_head *
2242 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2243 {
2244         struct list_head        *node;
2245         struct kib_pool *pool;
2246         int                     rc;
2247         unsigned int            interval = 1;
2248         ktime_t time_before;
2249         unsigned int trips = 0;
2250
2251 again:
2252         spin_lock(&ps->ps_lock);
2253         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2254                 if (list_empty(&pool->po_free_list))
2255                         continue;
2256
2257                 pool->po_allocated++;
2258                 pool->po_deadline = ktime_get_seconds() +
2259                                     IBLND_POOL_DEADLINE;
2260                 node = pool->po_free_list.next;
2261                 list_del(node);
2262
2263                 if (ps->ps_node_init != NULL) {
2264                         /* still hold the lock */
2265                         ps->ps_node_init(pool, node);
2266                 }
2267                 spin_unlock(&ps->ps_lock);
2268                 return node;
2269         }
2270
2271         /* no available tx pool and ... */
2272         if (ps->ps_increasing) {
2273                 /* another thread is allocating a new pool */
2274                 spin_unlock(&ps->ps_lock);
2275                 trips++;
2276                 CDEBUG(D_NET,
2277                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2278                        ps->ps_name, interval, trips);
2279
2280                 schedule_timeout_interruptible(interval);
2281                 if (interval < cfs_time_seconds(1))
2282                         interval *= 2;
2283
2284                 goto again;
2285         }
2286
2287         if (ktime_get_seconds() < ps->ps_next_retry) {
2288                 /* someone failed recently */
2289                 spin_unlock(&ps->ps_lock);
2290                 return NULL;
2291         }
2292
2293         ps->ps_increasing = 1;
2294         spin_unlock(&ps->ps_lock);
2295
2296         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2297         time_before = ktime_get();
2298         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2299         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2300                ktime_ms_delta(ktime_get(), time_before));
2301
2302         spin_lock(&ps->ps_lock);
2303         ps->ps_increasing = 0;
2304         if (rc == 0) {
2305                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2306         } else {
2307                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2308                 CERROR("Can't allocate new %s pool because out of memory\n",
2309                        ps->ps_name);
2310         }
2311         spin_unlock(&ps->ps_lock);
2312
2313         goto again;
2314 }
2315
2316 static void
2317 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2318 {
2319         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2320                                                tpo_pool);
2321         int i;
2322
2323         LASSERT (pool->po_allocated == 0);
2324
2325         if (tpo->tpo_tx_pages != NULL) {
2326                 kiblnd_unmap_tx_pool(tpo);
2327                 kiblnd_free_pages(tpo->tpo_tx_pages);
2328         }
2329
2330         if (tpo->tpo_tx_descs == NULL)
2331                 goto out;
2332
2333         for (i = 0; i < pool->po_size; i++) {
2334                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2335                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2336
2337                 list_del(&tx->tx_list);
2338                 if (tx->tx_pages != NULL)
2339                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2340                 if (tx->tx_frags != NULL)
2341                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2342                                            IBLND_MAX_RDMA_FRAGS);
2343                 if (tx->tx_wrq != NULL)
2344                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2345                                            IBLND_MAX_RDMA_FRAGS);
2346                 if (tx->tx_sge != NULL) {
2347                         /* +1 is for the lnet header/message itself */
2348                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2349                                            (IBLND_MAX_RDMA_FRAGS *
2350                                            wrq_sge + 1));
2351                 }
2352                 if (tx->tx_rd != NULL)
2353                         LIBCFS_FREE(tx->tx_rd,
2354                                     offsetof(struct kib_rdma_desc,
2355                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2356         }
2357
2358         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2359 out:
2360         kiblnd_fini_pool(pool);
2361         CFS_FREE_PTR(tpo);
2362 }
2363
2364 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2365 {
2366         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2367         int ntx;
2368
2369         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2370         ntx = tunables->lnd_ntx / ncpts;
2371
2372         return max(IBLND_TX_POOL, ntx);
2373 }
2374
2375 static int
2376 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2377 {
2378         int            i;
2379         int            npg;
2380         struct kib_pool *pool;
2381         struct kib_tx_pool *tpo;
2382
2383         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2384         if (tpo == NULL) {
2385                 CERROR("Failed to allocate TX pool\n");
2386                 return -ENOMEM;
2387         }
2388
2389         pool = &tpo->tpo_pool;
2390         kiblnd_init_pool(ps, pool, size);
2391         tpo->tpo_tx_descs = NULL;
2392         tpo->tpo_tx_pages = NULL;
2393
2394         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2395         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2396                 CERROR("Can't allocate tx pages: %d\n", npg);
2397                 CFS_FREE_PTR(tpo);
2398                 return -ENOMEM;
2399         }
2400
2401         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2402                          size * sizeof(struct kib_tx));
2403         if (tpo->tpo_tx_descs == NULL) {
2404                 CERROR("Can't allocate %d tx descriptors\n", size);
2405                 ps->ps_pool_destroy(pool);
2406                 return -ENOMEM;
2407         }
2408
2409         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2410
2411         for (i = 0; i < size; i++) {
2412                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2413                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2414
2415                 tx->tx_pool = tpo;
2416                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2417                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2418                                          lnet_cpt_table(), ps->ps_cpt,
2419                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2420                         if (tx->tx_pages == NULL)
2421                                 break;
2422                 }
2423
2424                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2425                                  IBLND_MAX_RDMA_FRAGS *
2426                                  sizeof(*tx->tx_frags));
2427                 if (tx->tx_frags == NULL)
2428                         break;
2429
2430                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2431
2432                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2433                                  IBLND_MAX_RDMA_FRAGS *
2434                                  sizeof(*tx->tx_wrq));
2435                 if (tx->tx_wrq == NULL)
2436                         break;
2437
2438                 /* +1 is for the lnet header/message itself */
2439                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2440                                  (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2441                                  sizeof(*tx->tx_sge));
2442                 if (tx->tx_sge == NULL)
2443                         break;
2444
2445                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2446                                  offsetof(struct kib_rdma_desc,
2447                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2448                 if (tx->tx_rd == NULL)
2449                         break;
2450         }
2451
2452         if (i == size) {
2453                 kiblnd_map_tx_pool(tpo);
2454                 *pp_po = pool;
2455                 return 0;
2456         }
2457
2458         ps->ps_pool_destroy(pool);
2459         return -ENOMEM;
2460 }
2461
2462 static void
2463 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2464 {
2465         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2466                                                   struct kib_tx_poolset,
2467                                                   tps_poolset);
2468         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2469
2470         tx->tx_cookie = tps->tps_next_tx_cookie++;
2471 }
2472
2473 static void
2474 kiblnd_net_fini_pools(struct kib_net *net)
2475 {
2476         int     i;
2477
2478         cfs_cpt_for_each(i, lnet_cpt_table()) {
2479                 struct kib_tx_poolset *tps;
2480                 struct kib_fmr_poolset *fps;
2481
2482                 if (net->ibn_tx_ps != NULL) {
2483                         tps = net->ibn_tx_ps[i];
2484                         kiblnd_fini_poolset(&tps->tps_poolset);
2485                 }
2486
2487                 if (net->ibn_fmr_ps != NULL) {
2488                         fps = net->ibn_fmr_ps[i];
2489                         kiblnd_fini_fmr_poolset(fps);
2490                 }
2491         }
2492
2493         if (net->ibn_tx_ps != NULL) {
2494                 cfs_percpt_free(net->ibn_tx_ps);
2495                 net->ibn_tx_ps = NULL;
2496         }
2497
2498         if (net->ibn_fmr_ps != NULL) {
2499                 cfs_percpt_free(net->ibn_fmr_ps);
2500                 net->ibn_fmr_ps = NULL;
2501         }
2502 }
2503
2504 static int
2505 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2506                       int ncpts)
2507 {
2508         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2509 #ifdef HAVE_IB_GET_DMA_MR
2510         unsigned long   flags;
2511 #endif
2512         int             cpt;
2513         int             rc;
2514         int             i;
2515
2516         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2517
2518 #ifdef HAVE_IB_GET_DMA_MR
2519         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2520         /*
2521          * if lnd_map_on_demand is zero then we have effectively disabled
2522          * FMR or FastReg and we're using global memory regions
2523          * exclusively.
2524          */
2525         if (!tunables->lnd_map_on_demand) {
2526                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2527                                            flags);
2528                 goto create_tx_pool;
2529         }
2530
2531         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2532 #endif
2533
2534         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2535                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2536                        tunables->lnd_fmr_pool_size,
2537                        tunables->lnd_ntx / 4);
2538                 rc = -EINVAL;
2539                 goto failed;
2540         }
2541
2542         /* TX pool must be created later than FMR, see LU-2268
2543          * for details */
2544         LASSERT(net->ibn_tx_ps == NULL);
2545
2546         /* premapping can fail if ibd_nmr > 1, so we always create
2547          * FMR pool and map-on-demand if premapping failed */
2548
2549         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2550                                            sizeof(struct kib_fmr_poolset));
2551         if (net->ibn_fmr_ps == NULL) {
2552                 CERROR("Failed to allocate FMR pool array\n");
2553                 rc = -ENOMEM;
2554                 goto failed;
2555         }
2556
2557         for (i = 0; i < ncpts; i++) {
2558                 cpt = (cpts == NULL) ? i : cpts[i];
2559                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2560                                              net, tunables);
2561                 if (rc != 0) {
2562                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2563                                cpt, rc);
2564                         goto failed;
2565                 }
2566         }
2567
2568         if (i > 0)
2569                 LASSERT(i == ncpts);
2570
2571 #ifdef HAVE_IB_GET_DMA_MR
2572  create_tx_pool:
2573 #endif
2574         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2575                                           sizeof(struct kib_tx_poolset));
2576         if (net->ibn_tx_ps == NULL) {
2577                 CERROR("Failed to allocate tx pool array\n");
2578                 rc = -ENOMEM;
2579                 goto failed;
2580         }
2581
2582         for (i = 0; i < ncpts; i++) {
2583                 cpt = (cpts == NULL) ? i : cpts[i];
2584                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2585                                          cpt, net, "TX",
2586                                          kiblnd_tx_pool_size(ni, ncpts),
2587                                          kiblnd_create_tx_pool,
2588                                          kiblnd_destroy_tx_pool,
2589                                          kiblnd_tx_init, NULL);
2590                 if (rc != 0) {
2591                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2592                                cpt, rc);
2593                         goto failed;
2594                 }
2595         }
2596
2597         return 0;
2598  failed:
2599         kiblnd_net_fini_pools(net);
2600         LASSERT(rc != 0);
2601         return rc;
2602 }
2603
2604 static int
2605 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2606 {
2607         struct ib_port_attr *port_attr;
2608         int rc;
2609         unsigned long flags;
2610         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2611
2612         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2613         if (port_attr == NULL) {
2614                 CDEBUG(D_NETERROR, "Out of memory\n");
2615                 return -ENOMEM;
2616         }
2617
2618         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2619
2620         write_lock_irqsave(g_lock, flags);
2621
2622         if (rc == 0)
2623                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2624                                  ? IBLND_DEV_PORT_ACTIVE
2625                                  : IBLND_DEV_PORT_DOWN;
2626
2627         write_unlock_irqrestore(g_lock, flags);
2628         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2629
2630         if (rc != 0) {
2631                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2632                 return rc;
2633         }
2634         return 0;
2635 }
2636
2637 static inline void
2638 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2639 {
2640         struct kib_net  *net;
2641
2642         /* for health check */
2643         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2644                 if (val)
2645                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2646                                         libcfs_nidstr(&net->ibn_ni->ni_nid));
2647                 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2648         }
2649 }
2650
2651 void
2652 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2653 {
2654         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2655         struct kib_hca_dev  *hdev;
2656         unsigned long flags;
2657
2658         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2659
2660         write_lock_irqsave(g_lock, flags);
2661
2662         switch (event->event) {
2663         case IB_EVENT_DEVICE_FATAL:
2664                 CDEBUG(D_NET, "IB device fatal\n");
2665                 hdev->ibh_state = IBLND_DEV_FATAL;
2666                 kiblnd_set_ni_fatal_on(hdev, 1);
2667                 break;
2668         case IB_EVENT_PORT_ACTIVE:
2669                 CDEBUG(D_NET, "IB port active\n");
2670                 if (event->element.port_num == hdev->ibh_port) {
2671                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2672                         kiblnd_set_ni_fatal_on(hdev, 0);
2673                 }
2674                 break;
2675         case IB_EVENT_PORT_ERR:
2676                 CDEBUG(D_NET, "IB port err\n");
2677                 if (event->element.port_num == hdev->ibh_port) {
2678                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2679                         kiblnd_set_ni_fatal_on(hdev, 1);
2680                 }
2681                 break;
2682         default:
2683                 break;
2684         }
2685         write_unlock_irqrestore(g_lock, flags);
2686 }
2687
2688 static int
2689 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2690 {
2691         struct ib_device_attr *dev_attr;
2692         int rc = 0;
2693         int rc2 = 0;
2694
2695         /* It's safe to assume a HCA can handle a page size
2696          * matching that of the native system */
2697         hdev->ibh_page_shift = PAGE_SHIFT;
2698         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2699         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2700
2701 #ifndef HAVE_IB_DEVICE_ATTRS
2702         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2703         if (dev_attr == NULL) {
2704                 CERROR("Out of memory\n");
2705                 return -ENOMEM;
2706         }
2707
2708         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2709         if (rc != 0) {
2710                 CERROR("Failed to query IB device: %d\n", rc);
2711                 goto out_clean_attr;
2712         }
2713 #else
2714         dev_attr = &hdev->ibh_ibdev->attrs;
2715 #endif
2716
2717         hdev->ibh_mr_size = dev_attr->max_mr_size;
2718         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2719
2720         /* Setup device Memory Registration capabilities */
2721 #ifdef HAVE_FMR_POOL_API
2722 #ifdef HAVE_IB_DEVICE_OPS
2723         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2724             hdev->ibh_ibdev->ops.dealloc_fmr &&
2725             hdev->ibh_ibdev->ops.map_phys_fmr &&
2726             hdev->ibh_ibdev->ops.unmap_fmr) {
2727 #else
2728         if (hdev->ibh_ibdev->alloc_fmr &&
2729             hdev->ibh_ibdev->dealloc_fmr &&
2730             hdev->ibh_ibdev->map_phys_fmr &&
2731             hdev->ibh_ibdev->unmap_fmr) {
2732 #endif
2733                 LCONSOLE_INFO("Using FMR for registration\n");
2734                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2735         } else
2736 #endif /* HAVE_FMR_POOL_API */
2737         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2738                 LCONSOLE_INFO("Using FastReg for registration\n");
2739                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2740 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2741 #ifdef IB_DEVICE_SG_GAPS_REG
2742                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2743                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2744 #endif
2745 #endif
2746         } else {
2747                 rc = -ENOSYS;
2748         }
2749
2750         rc2 = kiblnd_port_get_attr(hdev);
2751         if (rc2 != 0)
2752                 return rc2;
2753
2754         if (rc != 0)
2755                 rc = -EINVAL;
2756
2757 #ifndef HAVE_IB_DEVICE_ATTRS
2758 out_clean_attr:
2759         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2760 #endif
2761
2762         if (rc == -ENOSYS)
2763                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2764                        "register memory: %d\n", rc);
2765         else if (rc == -EINVAL)
2766                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2767         return rc;
2768 }
2769
2770 #ifdef HAVE_IB_GET_DMA_MR
2771 static void
2772 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2773 {
2774         if (hdev->ibh_mrs == NULL)
2775                 return;
2776
2777         ib_dereg_mr(hdev->ibh_mrs);
2778
2779         hdev->ibh_mrs = NULL;
2780 }
2781 #endif
2782
2783 void
2784 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2785 {
2786         if (hdev->ibh_event_handler.device != NULL)
2787                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2788
2789 #ifdef HAVE_IB_GET_DMA_MR
2790         kiblnd_hdev_cleanup_mrs(hdev);
2791 #endif
2792
2793         if (hdev->ibh_pd != NULL)
2794                 ib_dealloc_pd(hdev->ibh_pd);
2795
2796         if (hdev->ibh_cmid != NULL)
2797                 rdma_destroy_id(hdev->ibh_cmid);
2798
2799         LIBCFS_FREE(hdev, sizeof(*hdev));
2800 }
2801
2802 #ifdef HAVE_IB_GET_DMA_MR
2803 static int
2804 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2805 {
2806         struct ib_mr *mr;
2807         int           acflags = IB_ACCESS_LOCAL_WRITE |
2808                                 IB_ACCESS_REMOTE_WRITE;
2809
2810         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2811         if (IS_ERR(mr)) {
2812                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2813                 kiblnd_hdev_cleanup_mrs(hdev);
2814                 return PTR_ERR(mr);
2815         }
2816
2817         hdev->ibh_mrs = mr;
2818
2819         return 0;
2820 }
2821 #endif
2822
2823 static int
2824 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2825 {       /* DUMMY */
2826         return 0;
2827 }
2828
2829 static int kiblnd_get_link_status(struct net_device *dev)
2830 {
2831         int ret = -1;
2832
2833         LASSERT(dev);
2834
2835         if (!netif_running(dev))
2836                 ret = 0;
2837         /* Some devices may not be providing link settings */
2838         else if (dev->ethtool_ops->get_link)
2839                 ret = dev->ethtool_ops->get_link(dev);
2840
2841         return ret;
2842 }
2843
2844 static int
2845 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2846 {
2847         struct rdma_cm_id  *cmid;
2848         struct sockaddr_in  srcaddr;
2849         struct sockaddr_in  dstaddr;
2850         int                 rc;
2851
2852         if (dev->ibd_hdev == NULL || /* initializing */
2853             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2854             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2855                 return 1;
2856
2857         /* XXX: it's UGLY, but I don't have better way to find
2858          * ib-bonding HCA failover because:
2859          *
2860          * a. no reliable CM event for HCA failover...
2861          * b. no OFED API to get ib_device for current net_device...
2862          *
2863          * We have only two choices at this point:
2864          *
2865          * a. rdma_bind_addr(), it will conflict with listener cmid
2866          * b. rdma_resolve_addr() to zero addr */
2867         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2868                                      RDMA_PS_TCP, IB_QPT_RC);
2869         if (IS_ERR(cmid)) {
2870                 rc = PTR_ERR(cmid);
2871                 CERROR("Failed to create cmid for failover: %d\n", rc);
2872                 return rc;
2873         }
2874
2875         memset(&srcaddr, 0, sizeof(srcaddr));
2876         srcaddr.sin_family      = AF_INET;
2877         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2878
2879         memset(&dstaddr, 0, sizeof(dstaddr));
2880         dstaddr.sin_family = AF_INET;
2881         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2882                                (struct sockaddr *)&dstaddr, 1);
2883         if (rc != 0 || cmid->device == NULL) {
2884                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2885                        dev->ibd_ifname, &dev->ibd_ifip,
2886                        cmid->device, rc);
2887                 rdma_destroy_id(cmid);
2888                 return rc;
2889         }
2890
2891         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2892         rdma_destroy_id(cmid);
2893         return rc;
2894 }
2895
2896 int
2897 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2898 {
2899         LIST_HEAD(zombie_tpo);
2900         LIST_HEAD(zombie_ppo);
2901         LIST_HEAD(zombie_fpo);
2902         struct rdma_cm_id  *cmid  = NULL;
2903         struct kib_hca_dev *hdev  = NULL;
2904         struct kib_hca_dev *old;
2905         struct ib_pd       *pd;
2906         struct kib_net *net;
2907         struct sockaddr_in  addr;
2908         struct net_device *netdev;
2909         unsigned long       flags;
2910         int                 rc = 0;
2911         int                 i;
2912         bool                set_fatal = true;
2913
2914         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2915                 dev->ibd_can_failover ||
2916                 dev->ibd_hdev == NULL);
2917
2918         rc = kiblnd_dev_need_failover(dev, ns);
2919         if (rc <= 0)
2920                 goto out;
2921
2922         if (dev->ibd_hdev != NULL &&
2923             dev->ibd_hdev->ibh_cmid != NULL) {
2924                 /* XXX it's not good to close old listener at here,
2925                  * because we can fail to create new listener.
2926                  * But we have to close it now, otherwise rdma_bind_addr
2927                  * will return EADDRINUSE... How crap! */
2928                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2929
2930                 cmid = dev->ibd_hdev->ibh_cmid;
2931                 /* make next schedule of kiblnd_dev_need_failover()
2932                  * return 1 for me */
2933                 dev->ibd_hdev->ibh_cmid  = NULL;
2934                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2935
2936                 rdma_destroy_id(cmid);
2937         }
2938
2939         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2940                                      IB_QPT_RC);
2941         if (IS_ERR(cmid)) {
2942                 rc = PTR_ERR(cmid);
2943                 CERROR("Failed to create cmid for failover: %d\n", rc);
2944                 goto out;
2945         }
2946
2947         memset(&addr, 0, sizeof(addr));
2948         addr.sin_family      = AF_INET;
2949         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2950         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2951
2952         /* Bind to failover device or port */
2953         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2954         if (rc != 0 || cmid->device == NULL) {
2955                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2956                        dev->ibd_ifname, &dev->ibd_ifip,
2957                        cmid->device, rc);
2958                 if (!rc && !cmid->device)
2959                         set_fatal = false;
2960                 rdma_destroy_id(cmid);
2961                 goto out;
2962         }
2963
2964         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2965         if (hdev == NULL) {
2966                 CERROR("Failed to allocate kib_hca_dev\n");
2967                 rdma_destroy_id(cmid);
2968                 rc = -ENOMEM;
2969                 goto out;
2970         }
2971
2972         atomic_set(&hdev->ibh_ref, 1);
2973         hdev->ibh_dev   = dev;
2974         hdev->ibh_cmid  = cmid;
2975         hdev->ibh_ibdev = cmid->device;
2976         hdev->ibh_port  = cmid->port_num;
2977
2978 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2979         pd = ib_alloc_pd(cmid->device, 0);
2980 #else
2981         pd = ib_alloc_pd(cmid->device);
2982 #endif
2983         if (IS_ERR(pd)) {
2984                 rc = PTR_ERR(pd);
2985                 CERROR("Can't allocate PD: %d\n", rc);
2986                 goto out;
2987         }
2988
2989         hdev->ibh_pd = pd;
2990
2991         rc = rdma_listen(cmid, 0);
2992         if (rc != 0) {
2993                 CERROR("Can't start new listener: %d\n", rc);
2994                 goto out;
2995         }
2996
2997         rc = kiblnd_hdev_get_attr(hdev);
2998         if (rc != 0) {
2999                 CERROR("Can't get device attributes: %d\n", rc);
3000                 goto out;
3001         }
3002
3003 #ifdef HAVE_IB_GET_DMA_MR
3004         rc = kiblnd_hdev_setup_mrs(hdev);
3005         if (rc != 0) {
3006                 CERROR("Can't setup device: %d\n", rc);
3007                 goto out;
3008         }
3009 #endif
3010
3011         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
3012                                 hdev->ibh_ibdev, kiblnd_event_handler);
3013         ib_register_event_handler(&hdev->ibh_event_handler);
3014
3015         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3016
3017         old = dev->ibd_hdev;
3018         dev->ibd_hdev = hdev;   /* take over the refcount */
3019         hdev = old;
3020
3021         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
3022                 cfs_cpt_for_each(i, lnet_cpt_table()) {
3023                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
3024                                             &zombie_tpo);
3025
3026                         if (net->ibn_fmr_ps != NULL)
3027                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
3028                                                         &zombie_fpo);
3029                 }
3030         }
3031
3032         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3033  out:
3034         if (!list_empty(&zombie_tpo))
3035                 kiblnd_destroy_pool_list(&zombie_tpo);
3036         if (!list_empty(&zombie_ppo))
3037                 kiblnd_destroy_pool_list(&zombie_ppo);
3038         if (!list_empty(&zombie_fpo))
3039                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
3040         if (hdev != NULL)
3041                 kiblnd_hdev_decref(hdev);
3042
3043         if (rc != 0) {
3044                 dev->ibd_failed_failover++;
3045         } else {
3046                 dev->ibd_failed_failover = 0;
3047
3048                 if (set_fatal) {
3049                         rcu_read_lock();
3050                         netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
3051                         if (netdev && (kiblnd_get_link_status(netdev) == 1))
3052                                 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
3053                         rcu_read_unlock();
3054                 }
3055         }
3056
3057         return rc;
3058 }
3059
3060 void
3061 kiblnd_destroy_dev(struct kib_dev *dev)
3062 {
3063         LASSERT(dev->ibd_nnets == 0);
3064         LASSERT(list_empty(&dev->ibd_nets));
3065
3066         list_del(&dev->ibd_fail_list);
3067         list_del(&dev->ibd_list);
3068
3069         if (dev->ibd_hdev != NULL)
3070                 kiblnd_hdev_decref(dev->ibd_hdev);
3071
3072         LIBCFS_FREE(dev, sizeof(*dev));
3073 }
3074
3075 static struct kib_dev *
3076 kiblnd_dev_search(char *ifname)
3077 {
3078         struct kib_dev *alias = NULL;
3079         struct kib_dev *dev;
3080         char *colon;
3081         char *colon2;
3082
3083         colon = strchr(ifname, ':');
3084         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3085                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3086                         return dev;
3087
3088                 if (alias != NULL)
3089                         continue;
3090
3091                 colon2 = strchr(dev->ibd_ifname, ':');
3092                 if (colon != NULL)
3093                         *colon = 0;
3094                 if (colon2 != NULL)
3095                         *colon2 = 0;
3096
3097                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3098                         alias = dev;
3099
3100                 if (colon != NULL)
3101                         *colon = ':';
3102                 if (colon2 != NULL)
3103                         *colon2 = ':';
3104         }
3105         return alias;
3106 }
3107
3108 static int
3109 kiblnd_handle_link_state_change(struct net_device *dev,
3110                                 unsigned char operstate)
3111 {
3112         struct lnet_ni *ni = NULL;
3113         struct kib_dev *event_kibdev;
3114         struct kib_net *net;
3115         struct kib_net *cnxt;
3116         bool link_down = !(operstate == IF_OPER_UP);
3117         struct in_device *in_dev;
3118         bool found_ip = false;
3119         DECLARE_CONST_IN_IFADDR(ifa);
3120
3121         event_kibdev = kiblnd_dev_search(dev->name);
3122
3123         if (!event_kibdev)
3124                 goto out;
3125
3126         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3127                 found_ip = false;
3128
3129                 ni = net->ibn_ni;
3130
3131                 in_dev = __in_dev_get_rtnl(dev);
3132                 if (!in_dev) {
3133                         CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3134                                dev->name);
3135                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3136                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3137                         atomic_set(&ni->ni_fatal_error_on, 1);
3138                         continue;
3139                 }
3140                 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3141                         if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3142                                 found_ip = true;
3143                 }
3144                 endfor_ifa(in_dev);
3145
3146                 if (!found_ip) {
3147                         CDEBUG(D_NET, "Interface %s has no matching ip\n",
3148                                dev->name);
3149                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3150                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3151                         atomic_set(&ni->ni_fatal_error_on, 1);
3152                         continue;
3153                 }
3154
3155                 if (link_down) {
3156                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3157                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3158                         atomic_set(&ni->ni_fatal_error_on, link_down);
3159                 } else {
3160                         CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3161                                libcfs_nidstr(&net->ibn_ni->ni_nid),
3162                                (kiblnd_get_link_status(dev) == 0));
3163                         atomic_set(&ni->ni_fatal_error_on,
3164                                    (kiblnd_get_link_status(dev) == 0));
3165                 }
3166         }
3167 out:
3168         return 0;
3169 }
3170
3171 static int
3172 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3173 {
3174         struct kib_dev *event_kibdev;
3175         struct kib_net *net;
3176         struct kib_net *cnxt;
3177         struct net_device *event_netdev = ifa->ifa_dev->dev;
3178
3179         event_kibdev = kiblnd_dev_search(event_netdev->name);
3180
3181         if (!event_kibdev)
3182                 goto out;
3183
3184         if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3185                 goto out;
3186
3187         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3188                                  ibn_list) {
3189                 CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3190                        libcfs_nidstr(&net->ibn_ni->ni_nid),
3191                        (event == NETDEV_DOWN));
3192                 atomic_set(&net->ibn_ni->ni_fatal_error_on,
3193                            (event == NETDEV_DOWN));
3194         }
3195 out:
3196         return 0;
3197 }
3198
3199
3200 /************************************
3201  * Net device notifier event handler
3202  ************************************/
3203 static int kiblnd_device_event(struct notifier_block *unused,
3204                                  unsigned long event, void *ptr)
3205 {
3206         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3207         unsigned char operstate;
3208
3209         operstate = dev->operstate;
3210
3211         CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3212                event, dev->name, dev->ifindex, operstate);
3213
3214         switch (event) {
3215         case NETDEV_UP:
3216         case NETDEV_DOWN:
3217         case NETDEV_CHANGE:
3218                 kiblnd_handle_link_state_change(dev, operstate);
3219                 break;
3220         }
3221
3222         return NOTIFY_OK;
3223 }
3224
3225 /************************************
3226  * Inetaddr notifier event handler
3227  ************************************/
3228 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3229                                  unsigned long event, void *ptr)
3230 {
3231         struct in_ifaddr *ifa = ptr;
3232
3233         CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3234                event, &ifa->ifa_address, &ifa->ifa_mask);
3235
3236         switch (event) {
3237         case NETDEV_UP:
3238         case NETDEV_DOWN:
3239         case NETDEV_CHANGE:
3240                 kiblnd_handle_inetaddr_change(ifa, event);
3241                 break;
3242
3243         }
3244         return NOTIFY_OK;
3245 }
3246
3247 static struct notifier_block kiblnd_dev_notifier_block = {
3248         .notifier_call = kiblnd_device_event,
3249 };
3250
3251 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3252         .notifier_call = kiblnd_inetaddr_event,
3253 };
3254
3255 static void
3256 kiblnd_base_shutdown(void)
3257 {
3258         struct kib_sched_info *sched;
3259         struct kib_peer_ni *peer_ni;
3260         int i;
3261
3262         LASSERT(list_empty(&kiblnd_data.kib_devs));
3263
3264         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3265                libcfs_kmem_read());
3266
3267         if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3268                 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3269                 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3270         }
3271
3272         switch (kiblnd_data.kib_init) {
3273         default:
3274                 LBUG();
3275
3276         case IBLND_INIT_ALL:
3277         case IBLND_INIT_DATA:
3278                 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3279                         LASSERT(0);
3280                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3281                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3282                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3283                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3284
3285                 /* flag threads to terminate; wake and wait for them to die */
3286                 kiblnd_data.kib_shutdown = 1;
3287
3288                 /* NB: we really want to stop scheduler threads net by net
3289                  * instead of the whole module, this should be improved
3290                  * with dynamic configuration LNet.
3291                  */
3292                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3293                         wake_up_all(&sched->ibs_waitq);
3294
3295                 wake_up(&kiblnd_data.kib_connd_waitq);
3296                 wake_up(&kiblnd_data.kib_failover_waitq);
3297
3298                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3299                                        !atomic_read(&kiblnd_data.kib_nthreads),
3300                                        "Waiting for %d threads to terminate\n",
3301                                        atomic_read(&kiblnd_data.kib_nthreads));
3302                 fallthrough;
3303
3304         case IBLND_INIT_NOTHING:
3305                 break;
3306         }
3307
3308         if (kiblnd_data.kib_scheds != NULL)
3309                 cfs_percpt_free(kiblnd_data.kib_scheds);
3310
3311         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3312                libcfs_kmem_read());
3313
3314         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3315         module_put(THIS_MODULE);
3316 }
3317
3318 static void
3319 kiblnd_shutdown(struct lnet_ni *ni)
3320 {
3321         struct kib_net *net = ni->ni_data;
3322         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3323         unsigned long     flags;
3324
3325         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3326
3327         if (net == NULL)
3328                 goto out;
3329
3330         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3331                libcfs_kmem_read());
3332
3333         write_lock_irqsave(g_lock, flags);
3334         net->ibn_shutdown = 1;
3335         write_unlock_irqrestore(g_lock, flags);
3336
3337         switch (net->ibn_init) {
3338         default:
3339                 LBUG();
3340
3341         case IBLND_INIT_ALL:
3342                 /* nuke all existing peers within this net */
3343                 kiblnd_del_peer(ni, LNET_NID_ANY);
3344
3345                 /* Wait for all peer_ni state to clean up */
3346                 wait_var_event_warning(&net->ibn_npeers,
3347                                        atomic_read(&net->ibn_npeers) == 0,
3348                                        "%s: waiting for %d peers to disconnect\n",
3349                                        libcfs_nidstr(&ni->ni_nid),
3350                                        atomic_read(&net->ibn_npeers));
3351
3352                 kiblnd_net_fini_pools(net);
3353
3354                 write_lock_irqsave(g_lock, flags);
3355                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3356                 net->ibn_dev->ibd_nnets--;
3357                 list_del(&net->ibn_list);
3358                 write_unlock_irqrestore(g_lock, flags);
3359
3360                 wake_up_all(&kiblnd_data.kib_connd_waitq);
3361                 wait_var_event_warning(&net->ibn_nconns,
3362                                        atomic_read(&net->ibn_nconns) == 0,
3363                                        "%s: waiting for %d conns to clean\n",
3364                                        libcfs_nidstr(&ni->ni_nid),
3365                                        atomic_read(&net->ibn_nconns));
3366                 fallthrough;
3367
3368         case IBLND_INIT_NOTHING:
3369                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3370
3371                 if (net->ibn_dev != NULL &&
3372                     net->ibn_dev->ibd_nnets == 0)
3373                         kiblnd_destroy_dev(net->ibn_dev);
3374
3375                 break;
3376         }
3377
3378         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3379                libcfs_kmem_read());
3380
3381         net->ibn_init = IBLND_INIT_NOTHING;
3382         ni->ni_data = NULL;
3383
3384         LIBCFS_FREE(net, sizeof(*net));
3385
3386 out:
3387         if (list_empty(&kiblnd_data.kib_devs))
3388                 kiblnd_base_shutdown();
3389 }
3390
3391 static int
3392 kiblnd_base_startup(struct net *ns)
3393 {
3394         struct kib_sched_info *sched;
3395         int rc;
3396         int i;
3397
3398         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3399
3400         if (!try_module_get(THIS_MODULE))
3401                 goto failed;
3402
3403         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3404
3405         rwlock_init(&kiblnd_data.kib_global_lock);
3406
3407         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3408         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3409
3410         hash_init(kiblnd_data.kib_peers);
3411
3412         spin_lock_init(&kiblnd_data.kib_connd_lock);
3413         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3414         INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3415         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3416         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3417         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3418
3419         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3420         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3421
3422         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3423                                                   sizeof(*sched));
3424         if (kiblnd_data.kib_scheds == NULL)
3425                 goto failed;
3426
3427         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3428                 int     nthrs;
3429
3430                 spin_lock_init(&sched->ibs_lock);
3431                 INIT_LIST_HEAD(&sched->ibs_conns);
3432                 init_waitqueue_head(&sched->ibs_waitq);
3433
3434                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3435                 if (*kiblnd_tunables.kib_nscheds > 0) {
3436                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3437                 } else {
3438                         /* max to half of CPUs, another half is reserved for
3439                          * upper layer modules */
3440                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3441                 }
3442
3443                 sched->ibs_nthreads_max = nthrs;
3444                 sched->ibs_cpt = i;
3445         }
3446
3447         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3448
3449         /* lists/ptrs/locks initialised */
3450         kiblnd_data.kib_init = IBLND_INIT_DATA;
3451         /*****************************************************/
3452
3453         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3454         if (rc != 0) {
3455                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3456                 goto failed;
3457         }
3458
3459         if (*kiblnd_tunables.kib_dev_failover != 0)
3460                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3461                                          "kiblnd_failover");
3462
3463         if (rc != 0) {
3464                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3465                 goto failed;
3466         }
3467
3468         register_netdevice_notifier(&kiblnd_dev_notifier_block);
3469         register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3470
3471         /* flag everything initialised */
3472         kiblnd_data.kib_init = IBLND_INIT_ALL;
3473         /*****************************************************/
3474
3475         return 0;
3476
3477  failed:
3478         kiblnd_base_shutdown();
3479         return -ENETDOWN;
3480 }
3481
3482 static int
3483 kiblnd_start_schedulers(struct kib_sched_info *sched)
3484 {
3485         int     rc = 0;
3486         int     nthrs;
3487         int     i;
3488
3489         if (sched->ibs_nthreads == 0) {
3490                 if (*kiblnd_tunables.kib_nscheds > 0) {
3491                         nthrs = sched->ibs_nthreads_max;
3492                 } else {
3493                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3494                                                sched->ibs_cpt);
3495                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3496                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3497                 }
3498         } else {
3499                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3500                 /* increase one thread if there is new interface */
3501                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3502         }
3503
3504         for (i = 0; i < nthrs; i++) {
3505                 long    id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3506
3507                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3508                                          "kiblnd_sd_%02ld_%02ld",
3509                                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3510                 if (rc == 0)
3511                         continue;
3512
3513                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3514                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3515                 break;
3516         }
3517
3518         sched->ibs_nthreads += i;
3519         return rc;
3520 }
3521
3522 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3523                                     int ncpts)
3524 {
3525         int     cpt;
3526         int     rc;
3527         int     i;
3528
3529         for (i = 0; i < ncpts; i++) {
3530                 struct kib_sched_info *sched;
3531
3532                 cpt = (cpts == NULL) ? i : cpts[i];
3533                 sched = kiblnd_data.kib_scheds[cpt];
3534
3535                 if (!newdev && sched->ibs_nthreads > 0)
3536                         continue;
3537
3538                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3539                 if (rc != 0) {
3540                         CERROR("Failed to start scheduler threads for %s\n",
3541                                dev->ibd_ifname);
3542                         return rc;
3543                 }
3544         }
3545         return 0;
3546 }
3547
3548 static int
3549 kiblnd_startup(struct lnet_ni *ni)
3550 {
3551         char *ifname = NULL;
3552         struct lnet_inetdev *ifaces = NULL;
3553         struct kib_dev *ibdev = NULL;
3554         struct kib_net *net = NULL;
3555         unsigned long flags;
3556         int rc;
3557         int i;
3558         bool newdev;
3559
3560         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3561
3562         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3563                 rc = kiblnd_base_startup(ni->ni_net_ns);
3564                 if (rc != 0)
3565                         return rc;
3566         }
3567
3568         LIBCFS_ALLOC(net, sizeof(*net));
3569         ni->ni_data = net;
3570         if (net == NULL) {
3571                 rc = -ENOMEM;
3572                 goto failed;
3573         }
3574
3575         net->ibn_ni = ni;
3576         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3577
3578         kiblnd_tunables_setup(ni);
3579
3580         /* Multi-Rail wants each secondary
3581          * IP to be treated as an unique 'struct ni' interface.
3582          */
3583         if (ni->ni_interface != NULL) {
3584                 /* Use the IPoIB interface specified in 'networks=' */
3585                 ifname = ni->ni_interface;
3586         } else {
3587                 ifname = *kiblnd_tunables.kib_default_ipif;
3588                 rc = libcfs_strnid(&ni->ni_nid, ifname);
3589                 if (rc < 0 || ni->ni_nid.nid_type != O2IBLND)
3590                         memset(&ni->ni_nid, 0, sizeof(ni->ni_nid));
3591         }
3592
3593         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3594                 CERROR("IPoIB interface name too long: %s\n", ifname);
3595                 rc = -E2BIG;
3596                 goto failed;
3597         }
3598
3599         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3600         if (rc < 0)
3601                 goto failed;
3602
3603         i = lnet_inet_select(ni, ifaces, rc);
3604         if (i < 0)
3605                 goto failed;
3606
3607         if (nid_addr_is_set(&ni->ni_nid)) {
3608                 strscpy(ifname, ifaces[i].li_name, sizeof(ifname));
3609         } else if (strcmp(ifname, ifaces[i].li_name) != 0) {
3610                 CERROR("ko2iblnd: No matching interfaces\n");
3611                 rc = -ENOENT;
3612                 goto failed;
3613         }
3614
3615         ibdev = kiblnd_dev_search(ifname);
3616         newdev = ibdev == NULL;
3617         /* hmm...create kib_dev even for alias */
3618         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3619                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3620                 if (!ibdev) {
3621                         rc = -ENOMEM;
3622                         goto failed;
3623                 }
3624
3625                 ibdev->ibd_ifip = ntohl(ifaces[i].li_ipaddr);
3626                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3627                         sizeof(ibdev->ibd_ifname));
3628                 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3629
3630                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3631                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3632                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3633
3634                 /* initialize the device */
3635                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3636                 if (rc) {
3637                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3638                                rc);
3639                         goto failed;
3640                 }
3641
3642                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3643         }
3644
3645         net->ibn_dev = ibdev;
3646         ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3647         if (!ni->ni_interface) {
3648                 rc = lnet_ni_add_interface(ni, ifaces[i].li_name);
3649                 if (rc < 0)
3650                         CWARN("ko2iblnd failed to allocate ni_interface\n");
3651         }
3652         ni->ni_dev_cpt = ifaces[i].li_cpt;
3653
3654         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3655         if (rc != 0)
3656                 goto failed;
3657
3658         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3659         if (rc != 0) {
3660                 CERROR("Failed to initialize NI pools: %d\n", rc);
3661                 goto failed;
3662         }
3663
3664         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3665         ibdev->ibd_nnets++;
3666         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3667         /* for health check */
3668         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3669                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3670         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3671
3672         net->ibn_init = IBLND_INIT_ALL;
3673
3674         return 0;
3675
3676 failed:
3677         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3678                 kiblnd_destroy_dev(ibdev);
3679
3680         kfree(ifaces);
3681         kiblnd_shutdown(ni);
3682
3683         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3684                ifname ? ifname : "", rc);
3685
3686         return -ENETDOWN;
3687 }
3688
3689 static const struct lnet_lnd the_o2iblnd = {
3690         .lnd_type       = O2IBLND,
3691         .lnd_startup    = kiblnd_startup,
3692         .lnd_shutdown   = kiblnd_shutdown,
3693         .lnd_ctl        = kiblnd_ctl,
3694         .lnd_send       = kiblnd_send,
3695         .lnd_recv       = kiblnd_recv,
3696         .lnd_get_dev_prio = kiblnd_get_dev_prio,
3697         .lnd_nl_set     = kiblnd_nl_set,
3698         .lnd_keys       = &kiblnd_tunables_keys,
3699 };
3700
3701 static void ko2inlnd_assert_wire_constants(void)
3702 {
3703         BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3704         BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3705         BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3706         BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3707
3708         BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3709         BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3710         BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3711         BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3712         BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3713         BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3714         BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3715         BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3716         BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3717         BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3718
3719         BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3720         BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3721         BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3722         BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3723         BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3724         BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3725         BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3726         BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3727
3728         BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3729         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3730         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3731         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3732         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3733         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3734         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3735
3736         BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3737         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3738         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3739         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3740         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3741
3742         BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3743         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3744         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3745         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3746         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3747
3748         BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3749         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3750         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3751         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3752         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3753         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3754         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3755
3756         BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3757         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3758         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3759         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3760         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3761
3762         BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3763         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3764         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3765         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3766         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3767         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3768         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3769
3770         BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3771         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3772         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3773         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3774         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3775         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3776         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3777
3778         BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3779         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3780         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3781         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3782         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3783
3784         /* Checks for struct kib_msg */
3785         //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3786         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3787         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3788         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3789         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3790         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3791         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3792         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3793         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3794         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3795         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3796         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3797         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3798         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3799         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3800         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3801         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3802         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3803         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3804         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3805         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3806
3807         /* Connparams */
3808         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3809         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3810         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3811         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3812         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3813         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3814
3815         /* Immediate message */
3816         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3817         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3818         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3819         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3820
3821         /* PUT req message */
3822         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3823         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3824         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3825         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3826
3827         /* Put ACK */
3828         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3829         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3830         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3831         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3832         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3833         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3834
3835         /* GET message */
3836         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3837         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3838         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3839         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3840         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3841         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3842
3843         /* Completion message */
3844         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3845         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3846         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3847         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3848
3849         /* Sanity checks */
3850         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3851         BUILD_BUG_ON(offsetof(struct kib_msg,
3852                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3853                      IBLND_MSG_SIZE);
3854         BUILD_BUG_ON(offsetof(struct kib_msg,
3855                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3856                      IBLND_MSG_SIZE);
3857 }
3858
3859 static void __exit ko2iblnd_exit(void)
3860 {
3861         lnet_unregister_lnd(&the_o2iblnd);
3862 }
3863
3864 static int __init ko2iblnd_init(void)
3865 {
3866         int rc;
3867
3868         ko2inlnd_assert_wire_constants();
3869
3870         rc = kiblnd_tunables_init();
3871         if (rc != 0)
3872                 return rc;
3873
3874         lnet_register_lnd(&the_o2iblnd);
3875
3876         return 0;
3877 }
3878
3879 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3880 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3881 MODULE_VERSION("2.8.0");
3882 MODULE_LICENSE("GPL");
3883
3884 module_init(ko2iblnd_init);
3885 module_exit(ko2iblnd_exit);