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 void
1220 kiblnd_free_pages(struct kib_pages *p)
1221 {
1222         int     npages = p->ibp_npages;
1223         int     i;
1224
1225         for (i = 0; i < npages; i++) {
1226                 if (p->ibp_pages[i] != NULL)
1227                         __free_page(p->ibp_pages[i]);
1228         }
1229
1230         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1231 }
1232
1233 int
1234 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1235 {
1236         struct kib_pages *p;
1237         int i;
1238
1239         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1240                          offsetof(struct kib_pages, ibp_pages[npages]));
1241         if (p == NULL) {
1242                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1243                 return -ENOMEM;
1244         }
1245
1246         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1247         p->ibp_npages = npages;
1248
1249         for (i = 0; i < npages; i++) {
1250                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1251                                                      GFP_NOFS);
1252                 if (p->ibp_pages[i] == NULL) {
1253                         CERROR("Can't allocate page %d of %d\n", i, npages);
1254                         kiblnd_free_pages(p);
1255                         return -ENOMEM;
1256                 }
1257         }
1258
1259         *pp = p;
1260         return 0;
1261 }
1262
1263 void
1264 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1265 {
1266         struct kib_rx *rx;
1267         int       i;
1268
1269         LASSERT (conn->ibc_rxs != NULL);
1270         LASSERT (conn->ibc_hdev != NULL);
1271
1272         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1273                 rx = &conn->ibc_rxs[i];
1274
1275                 LASSERT(rx->rx_nob >= 0); /* not posted */
1276
1277                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1278                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1279                                                           rx->rx_msgaddr),
1280                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1281         }
1282
1283         kiblnd_free_pages(conn->ibc_rx_pages);
1284
1285         conn->ibc_rx_pages = NULL;
1286 }
1287
1288 void
1289 kiblnd_map_rx_descs(struct kib_conn *conn)
1290 {
1291         struct kib_rx *rx;
1292         struct page    *pg;
1293         int             pg_off;
1294         int             ipg;
1295         int             i;
1296
1297         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1298                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1299                 rx = &conn->ibc_rxs[i];
1300
1301                 rx->rx_conn = conn;
1302                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1303
1304                 rx->rx_msgaddr =
1305                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1306                                               rx->rx_msg, IBLND_MSG_SIZE,
1307                                               DMA_FROM_DEVICE);
1308                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1309                                                   rx->rx_msgaddr));
1310                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1311
1312                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1313                        i, rx->rx_msg, rx->rx_msgaddr,
1314                        (__u64)(page_to_phys(pg) + pg_off));
1315
1316                 pg_off += IBLND_MSG_SIZE;
1317                 LASSERT(pg_off <= PAGE_SIZE);
1318
1319                 if (pg_off == PAGE_SIZE) {
1320                         pg_off = 0;
1321                         ipg++;
1322                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1323                 }
1324         }
1325 }
1326
1327 static void
1328 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1329 {
1330         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1331         struct kib_tx *tx;
1332         int i;
1333
1334         LASSERT (tpo->tpo_pool.po_allocated == 0);
1335
1336         if (hdev == NULL)
1337                 return;
1338
1339         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1340                 tx = &tpo->tpo_tx_descs[i];
1341                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1342                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1343                                                           tx->tx_msgaddr),
1344                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1345         }
1346
1347         kiblnd_hdev_decref(hdev);
1348         tpo->tpo_hdev = NULL;
1349 }
1350
1351 static struct kib_hca_dev *
1352 kiblnd_current_hdev(struct kib_dev *dev)
1353 {
1354         struct kib_hca_dev *hdev;
1355         unsigned long  flags;
1356         int            i = 0;
1357
1358         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1359         while (dev->ibd_failover) {
1360                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1361                 if (i++ % 50 == 0)
1362                         CDEBUG(D_NET, "%s: Wait for failover\n",
1363                                dev->ibd_ifname);
1364                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1365
1366                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1367         }
1368
1369         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1370         hdev = dev->ibd_hdev;
1371
1372         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1373
1374         return hdev;
1375 }
1376
1377 static void
1378 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1379 {
1380         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1381         struct kib_pool *pool = &tpo->tpo_pool;
1382         struct kib_net      *net   = pool->po_owner->ps_net;
1383         struct kib_dev *dev;
1384         struct page *page;
1385         struct kib_tx *tx;
1386         int             page_offset;
1387         int             ipage;
1388         int             i;
1389
1390         LASSERT (net != NULL);
1391
1392         dev = net->ibn_dev;
1393
1394         /* pre-mapped messages are not bigger than 1 page */
1395         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1396
1397         /* No fancy arithmetic when we do the buffer calculations */
1398         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1399
1400         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1401
1402         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1403                 page = txpgs->ibp_pages[ipage];
1404                 tx = &tpo->tpo_tx_descs[i];
1405
1406                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1407                                                 page_offset);
1408
1409                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1410                                                        tx->tx_msg,
1411                                                        IBLND_MSG_SIZE,
1412                                                        DMA_TO_DEVICE);
1413                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1414                                                   tx->tx_msgaddr));
1415                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1416
1417                 list_add(&tx->tx_list, &pool->po_free_list);
1418
1419                 page_offset += IBLND_MSG_SIZE;
1420                 LASSERT(page_offset <= PAGE_SIZE);
1421
1422                 if (page_offset == PAGE_SIZE) {
1423                         page_offset = 0;
1424                         ipage++;
1425                         LASSERT(ipage <= txpgs->ibp_npages);
1426                 }
1427         }
1428 }
1429
1430 static void
1431 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1432 {
1433         LASSERT(fpo->fpo_map_count == 0);
1434
1435 #ifdef HAVE_FMR_POOL_API
1436         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1437                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1438         } else
1439 #endif /* HAVE_FMR_POOL_API */
1440         {
1441                 struct kib_fast_reg_descriptor *frd, *tmp;
1442                 int i = 0;
1443
1444                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1445                                          frd_list) {
1446                         list_del(&frd->frd_list);
1447 #ifndef HAVE_IB_MAP_MR_SG
1448                         ib_free_fast_reg_page_list(frd->frd_frpl);
1449 #endif
1450                         ib_dereg_mr(frd->frd_mr);
1451                         LIBCFS_FREE(frd, sizeof(*frd));
1452                         i++;
1453                 }
1454                 if (i < fpo->fast_reg.fpo_pool_size)
1455                         CERROR("FastReg pool still has %d regions registered\n",
1456                                 fpo->fast_reg.fpo_pool_size - i);
1457         }
1458
1459         if (fpo->fpo_hdev)
1460                 kiblnd_hdev_decref(fpo->fpo_hdev);
1461
1462         LIBCFS_FREE(fpo, sizeof(*fpo));
1463 }
1464
1465 static void
1466 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1467 {
1468         struct kib_fmr_pool *fpo, *tmp;
1469
1470         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1471                 list_del(&fpo->fpo_list);
1472                 kiblnd_destroy_fmr_pool(fpo);
1473         }
1474 }
1475
1476 static int
1477 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1478                      int ncpts)
1479 {
1480         int size = tunables->lnd_fmr_pool_size / ncpts;
1481
1482         return max(IBLND_FMR_POOL, size);
1483 }
1484
1485 static int
1486 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1487                          int ncpts)
1488 {
1489         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1490
1491         return max(IBLND_FMR_POOL_FLUSH, size);
1492 }
1493
1494 #ifdef HAVE_FMR_POOL_API
1495 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1496                                  struct kib_fmr_pool *fpo)
1497 {
1498         struct ib_fmr_pool_param param = {
1499                 .max_pages_per_fmr = IBLND_MAX_RDMA_FRAGS,
1500                 .page_shift        = PAGE_SHIFT,
1501                 .access            = (IB_ACCESS_LOCAL_WRITE |
1502                                       IB_ACCESS_REMOTE_WRITE),
1503                 .pool_size         = fps->fps_pool_size,
1504                 .dirty_watermark   = fps->fps_flush_trigger,
1505                 .flush_function    = NULL,
1506                 .flush_arg         = NULL,
1507                 .cache             = !!fps->fps_cache };
1508         int rc = 0;
1509
1510         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1511                                                    &param);
1512         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1513                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1514                 if (rc != -ENOSYS)
1515                         CERROR("Failed to create FMR pool: %d\n", rc);
1516                 else
1517                         CERROR("FMRs are not supported\n");
1518         }
1519         fpo->fpo_is_fmr = true;
1520
1521         return rc;
1522 }
1523 #endif /* HAVE_FMR_POOL_API */
1524
1525 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1526                                   struct kib_fmr_pool *fpo,
1527                                   enum kib_dev_caps dev_caps)
1528 {
1529         struct kib_fast_reg_descriptor *frd, *tmp;
1530         int i, rc;
1531
1532 #ifdef HAVE_FMR_POOL_API
1533         fpo->fpo_is_fmr = false;
1534 #endif
1535
1536         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1537         fpo->fast_reg.fpo_pool_size = 0;
1538         for (i = 0; i < fps->fps_pool_size; i++) {
1539                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1540                                  sizeof(*frd));
1541                 if (!frd) {
1542                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1543                         rc = -ENOMEM;
1544                         goto out;
1545                 }
1546                 frd->frd_mr = NULL;
1547
1548 #ifndef HAVE_IB_MAP_MR_SG
1549                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1550                                                             IBLND_MAX_RDMA_FRAGS);
1551                 if (IS_ERR(frd->frd_frpl)) {
1552                         rc = PTR_ERR(frd->frd_frpl);
1553                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1554                                 rc);
1555                         frd->frd_frpl = NULL;
1556                         goto out_middle;
1557                 }
1558 #endif
1559
1560 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1561                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1562                                                    IBLND_MAX_RDMA_FRAGS);
1563 #else
1564                 /*
1565                  * it is expected to get here if this is an MLX-5 card.
1566                  * MLX-4 cards will always use FMR and MLX-5 cards will
1567                  * always use fast_reg. It turns out that some MLX-5 cards
1568                  * (possibly due to older FW versions) do not natively support
1569                  * gaps. So we will need to track them here.
1570                  */
1571                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1572 #ifdef IB_MR_TYPE_SG_GAPS
1573                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1574                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1575                                                 IB_MR_TYPE_SG_GAPS :
1576                                                 IB_MR_TYPE_MEM_REG,
1577 #else
1578                                                 IB_MR_TYPE_MEM_REG,
1579 #endif
1580                                           IBLND_MAX_RDMA_FRAGS);
1581                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1582                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1583                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1584 #endif
1585                 if (IS_ERR(frd->frd_mr)) {
1586                         rc = PTR_ERR(frd->frd_mr);
1587                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1588                         frd->frd_mr = NULL;
1589                         goto out_middle;
1590                 }
1591
1592                 /* There appears to be a bug in MLX5 code where you must
1593                  * invalidate the rkey of a new FastReg pool before first
1594                  * using it. Thus, I am marking the FRD invalid here. */
1595                 frd->frd_valid = false;
1596
1597                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1598                 fpo->fast_reg.fpo_pool_size++;
1599         }
1600
1601         return 0;
1602
1603 out_middle:
1604         if (frd->frd_mr)
1605                 ib_dereg_mr(frd->frd_mr);
1606 #ifndef HAVE_IB_MAP_MR_SG
1607         if (frd->frd_frpl)
1608                 ib_free_fast_reg_page_list(frd->frd_frpl);
1609 #endif
1610         LIBCFS_FREE(frd, sizeof(*frd));
1611
1612 out:
1613         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1614                                  frd_list) {
1615                 list_del(&frd->frd_list);
1616 #ifndef HAVE_IB_MAP_MR_SG
1617                 ib_free_fast_reg_page_list(frd->frd_frpl);
1618 #endif
1619                 ib_dereg_mr(frd->frd_mr);
1620                 LIBCFS_FREE(frd, sizeof(*frd));
1621         }
1622
1623         return rc;
1624 }
1625
1626 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1627                                   struct kib_fmr_pool **pp_fpo)
1628 {
1629         struct kib_dev *dev = fps->fps_net->ibn_dev;
1630         struct kib_fmr_pool *fpo;
1631         int rc;
1632
1633         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1634         if (!fpo) {
1635                 return -ENOMEM;
1636         }
1637         memset(fpo, 0, sizeof(*fpo));
1638
1639         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1640
1641 #ifdef HAVE_FMR_POOL_API
1642         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1643                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1644         else
1645 #endif /* HAVE_FMR_POOL_API */
1646                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1647         if (rc)
1648                 goto out_fpo;
1649
1650         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1651         fpo->fpo_owner = fps;
1652         *pp_fpo = fpo;
1653
1654         return 0;
1655
1656 out_fpo:
1657         kiblnd_hdev_decref(fpo->fpo_hdev);
1658         LIBCFS_FREE(fpo, sizeof(*fpo));
1659         return rc;
1660 }
1661
1662 static void
1663 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1664 {
1665         struct kib_fmr_pool *fpo;
1666
1667         if (fps->fps_net == NULL) /* intialized? */
1668                 return;
1669
1670         spin_lock(&fps->fps_lock);
1671
1672         while ((fpo = list_first_entry_or_null(&fps->fps_pool_list,
1673                                                struct kib_fmr_pool,
1674                                                fpo_list)) != NULL) {
1675                 fpo->fpo_failed = 1;
1676                 if (fpo->fpo_map_count == 0)
1677                         list_move(&fpo->fpo_list, zombies);
1678                 else
1679                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1680         }
1681
1682         spin_unlock(&fps->fps_lock);
1683 }
1684
1685 static void
1686 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1687 {
1688         if (fps->fps_net != NULL) { /* initialized? */
1689                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1690                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1691         }
1692 }
1693
1694 static int
1695 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1696                         struct kib_net *net,
1697                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1698 {
1699         struct kib_fmr_pool *fpo;
1700         int rc;
1701
1702         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1703
1704         fps->fps_net = net;
1705         fps->fps_cpt = cpt;
1706
1707         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1708         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1709         fps->fps_cache = tunables->lnd_fmr_cache;
1710
1711         spin_lock_init(&fps->fps_lock);
1712         INIT_LIST_HEAD(&fps->fps_pool_list);
1713         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1714
1715         rc = kiblnd_create_fmr_pool(fps, &fpo);
1716         if (rc == 0)
1717                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1718
1719         return rc;
1720 }
1721
1722 static int
1723 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1724 {
1725         if (fpo->fpo_map_count != 0) /* still in use */
1726                 return 0;
1727         if (fpo->fpo_failed)
1728                 return 1;
1729         return now >= fpo->fpo_deadline;
1730 }
1731
1732 #if defined(HAVE_FMR_POOL_API) || !defined(HAVE_IB_MAP_MR_SG)
1733 static int
1734 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1735 {
1736         struct kib_hca_dev *hdev;
1737         __u64           *pages = tx->tx_pages;
1738         int             npages;
1739         int             size;
1740         int             i;
1741
1742         hdev = tx->tx_pool->tpo_hdev;
1743
1744         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1745                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1746                         size += hdev->ibh_page_size) {
1747                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1748                                            hdev->ibh_page_mask) + size;
1749                 }
1750         }
1751
1752         return npages;
1753 }
1754 #endif
1755
1756 void
1757 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1758 {
1759         LIST_HEAD(zombies);
1760         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1761         struct kib_fmr_poolset *fps;
1762         time64_t now = ktime_get_seconds();
1763         struct kib_fmr_pool *tmp;
1764
1765         if (!fpo)
1766                 return;
1767
1768         fps = fpo->fpo_owner;
1769
1770 #ifdef HAVE_FMR_POOL_API
1771         if (fpo->fpo_is_fmr) {
1772                 if (fmr->fmr_pfmr) {
1773                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1774                         fmr->fmr_pfmr = NULL;
1775                 }
1776
1777                 if (status) {
1778                         int rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1779                         LASSERT(!rc);
1780                 }
1781         } else
1782 #endif /* HAVE_FMR_POOL_API */
1783         {
1784                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1785
1786                 if (frd) {
1787                         frd->frd_valid = false;
1788                         frd->frd_posted = false;
1789                         fmr->fmr_frd = NULL;
1790                         spin_lock(&fps->fps_lock);
1791                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1792                         spin_unlock(&fps->fps_lock);
1793                 }
1794         }
1795         fmr->fmr_pool = NULL;
1796
1797         spin_lock(&fps->fps_lock);
1798         fpo->fpo_map_count--;   /* decref the pool */
1799
1800         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1801                 /* the first pool is persistent */
1802                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1803                         continue;
1804
1805                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1806                         list_move(&fpo->fpo_list, &zombies);
1807                         fps->fps_version++;
1808                 }
1809         }
1810         spin_unlock(&fps->fps_lock);
1811
1812         if (!list_empty(&zombies))
1813                 kiblnd_destroy_fmr_pool_list(&zombies);
1814 }
1815
1816 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1817                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1818                         struct kib_fmr *fmr)
1819 {
1820         struct kib_fmr_pool *fpo;
1821         __u64 version;
1822         bool is_rx = (rd != tx->tx_rd);
1823 #ifdef HAVE_FMR_POOL_API
1824         __u64 *pages = tx->tx_pages;
1825         bool tx_pages_mapped = false;
1826         int npages = 0;
1827 #endif
1828         int rc;
1829
1830 again:
1831         spin_lock(&fps->fps_lock);
1832         version = fps->fps_version;
1833         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1834                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1835                 fpo->fpo_map_count++;
1836
1837 #ifdef HAVE_FMR_POOL_API
1838                 fmr->fmr_pfmr = NULL;
1839                 if (fpo->fpo_is_fmr) {
1840                         struct ib_pool_fmr *pfmr;
1841
1842                         spin_unlock(&fps->fps_lock);
1843
1844                         if (!tx_pages_mapped) {
1845                                 npages = kiblnd_map_tx_pages(tx, rd);
1846                                 tx_pages_mapped = true;
1847                         }
1848
1849                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1850                                                 pages, npages, iov);
1851                         if (likely(!IS_ERR(pfmr))) {
1852                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1853                                         : pfmr->fmr->lkey;
1854                                 fmr->fmr_frd  = NULL;
1855                                 fmr->fmr_pfmr = pfmr;
1856                                 fmr->fmr_pool = fpo;
1857                                 return 0;
1858                         }
1859                         rc = PTR_ERR(pfmr);
1860                 } else
1861 #endif /* HAVE_FMR_POOL_API */
1862                 {
1863                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1864                                 struct kib_fast_reg_descriptor *frd;
1865 #ifdef HAVE_IB_MAP_MR_SG
1866                                 struct ib_reg_wr *wr;
1867                                 int n;
1868 #else
1869                                 struct ib_rdma_wr *wr;
1870                                 struct ib_fast_reg_page_list *frpl;
1871 #endif
1872                                 struct ib_mr *mr;
1873
1874                                 frd = list_first_entry(
1875                                         &fpo->fast_reg.fpo_pool_list,
1876                                         struct kib_fast_reg_descriptor,
1877                                         frd_list);
1878                                 list_del(&frd->frd_list);
1879                                 spin_unlock(&fps->fps_lock);
1880
1881 #ifndef HAVE_IB_MAP_MR_SG
1882                                 frpl = frd->frd_frpl;
1883 #endif
1884                                 mr   = frd->frd_mr;
1885
1886                                 if (!frd->frd_valid) {
1887                                         struct ib_rdma_wr *inv_wr;
1888                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1889
1890                                         inv_wr = &frd->frd_inv_wr;
1891                                         memset(inv_wr, 0, sizeof(*inv_wr));
1892
1893                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1894                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1895                                         inv_wr->wr.ex.invalidate_rkey = key;
1896
1897                                         /* Bump the key */
1898                                         key = ib_inc_rkey(key);
1899                                         ib_update_fast_reg_key(mr, key);
1900                                 }
1901
1902 #ifdef HAVE_IB_MAP_MR_SG
1903 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1904                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1905                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
1906 #else
1907                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1908                                                  rd->rd_nfrags, PAGE_SIZE);
1909 #endif /* HAVE_IB_MAP_MR_SG_5ARGS */
1910                                 if (unlikely(n != rd->rd_nfrags)) {
1911                                         CERROR("Failed to map mr %d/%d elements\n",
1912                                                n, rd->rd_nfrags);
1913                                         return n < 0 ? n : -EINVAL;
1914                                 }
1915
1916                                 wr = &frd->frd_fastreg_wr;
1917                                 memset(wr, 0, sizeof(*wr));
1918
1919                                 wr->wr.opcode = IB_WR_REG_MR;
1920                                 wr->wr.wr_id  = IBLND_WID_MR;
1921                                 wr->wr.num_sge = 0;
1922                                 wr->wr.send_flags = 0;
1923                                 wr->mr = mr;
1924                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1925                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1926                                               IB_ACCESS_REMOTE_WRITE);
1927 #else /* HAVE_IB_MAP_MR_SG */
1928                                 if (!tx_pages_mapped) {
1929                                         npages = kiblnd_map_tx_pages(tx, rd);
1930                                         tx_pages_mapped = true;
1931                                 }
1932
1933                                 LASSERT(npages <= frpl->max_page_list_len);
1934                                 memcpy(frpl->page_list, pages,
1935                                        sizeof(*pages) * npages);
1936
1937                                 /* Prepare FastReg WR */
1938                                 wr = &frd->frd_fastreg_wr;
1939                                 memset(wr, 0, sizeof(*wr));
1940
1941                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1942                                 wr->wr.wr_id  = IBLND_WID_MR;
1943
1944                                 wr->wr.wr.fast_reg.iova_start = iov;
1945                                 wr->wr.wr.fast_reg.page_list  = frpl;
1946                                 wr->wr.wr.fast_reg.page_list_len = npages;
1947                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1948                                 wr->wr.wr.fast_reg.length = nob;
1949                                 wr->wr.wr.fast_reg.rkey =
1950                                         is_rx ? mr->rkey : mr->lkey;
1951                                 wr->wr.wr.fast_reg.access_flags =
1952                                         (IB_ACCESS_LOCAL_WRITE |
1953                                          IB_ACCESS_REMOTE_WRITE);
1954 #endif /* HAVE_IB_MAP_MR_SG */
1955
1956                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1957                                 fmr->fmr_frd  = frd;
1958                                 fmr->fmr_pool = fpo;
1959                                 frd->frd_posted = false;
1960                                 return 0;
1961                         }
1962                         spin_unlock(&fps->fps_lock);
1963                         rc = -EAGAIN;
1964                 }
1965
1966                 spin_lock(&fps->fps_lock);
1967                 fpo->fpo_map_count--;
1968                 if (rc != -EAGAIN) {
1969                         spin_unlock(&fps->fps_lock);
1970                         return rc;
1971                 }
1972
1973                 /* EAGAIN and ... */
1974                 if (version != fps->fps_version) {
1975                         spin_unlock(&fps->fps_lock);
1976                         goto again;
1977                 }
1978         }
1979
1980         if (fps->fps_increasing) {
1981                 spin_unlock(&fps->fps_lock);
1982                 CDEBUG(D_NET, "Another thread is allocating new "
1983                        "FMR pool, waiting for her to complete\n");
1984                 wait_var_event(fps, !fps->fps_increasing);
1985                 goto again;
1986
1987         }
1988
1989         if (ktime_get_seconds() < fps->fps_next_retry) {
1990                 /* someone failed recently */
1991                 spin_unlock(&fps->fps_lock);
1992                 return -EAGAIN;
1993         }
1994
1995         fps->fps_increasing = 1;
1996         spin_unlock(&fps->fps_lock);
1997
1998         CDEBUG(D_NET, "Allocate new FMR pool\n");
1999         rc = kiblnd_create_fmr_pool(fps, &fpo);
2000         spin_lock(&fps->fps_lock);
2001         fps->fps_increasing = 0;
2002         wake_up_var(fps);
2003         if (rc == 0) {
2004                 fps->fps_version++;
2005                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2006         } else {
2007                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2008         }
2009         spin_unlock(&fps->fps_lock);
2010
2011         goto again;
2012 }
2013
2014 static void
2015 kiblnd_fini_pool(struct kib_pool *pool)
2016 {
2017         LASSERT(list_empty(&pool->po_free_list));
2018         LASSERT(pool->po_allocated == 0);
2019
2020         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2021 }
2022
2023 static void
2024 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2025 {
2026         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2027
2028         memset(pool, 0, sizeof(struct kib_pool));
2029         INIT_LIST_HEAD(&pool->po_free_list);
2030         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2031         pool->po_owner = ps;
2032         pool->po_size = size;
2033 }
2034
2035 static void
2036 kiblnd_destroy_pool_list(struct list_head *head)
2037 {
2038         struct kib_pool *pool;
2039
2040         while ((pool = list_first_entry_or_null(head,
2041                                                 struct kib_pool,
2042                                                 po_list)) != NULL) {
2043                 list_del(&pool->po_list);
2044
2045                 LASSERT(pool->po_owner != NULL);
2046                 pool->po_owner->ps_pool_destroy(pool);
2047         }
2048 }
2049
2050 static void
2051 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2052 {
2053         struct kib_pool *po;
2054
2055         if (ps->ps_net == NULL) /* intialized? */
2056                 return;
2057
2058         spin_lock(&ps->ps_lock);
2059         while ((po = list_first_entry_or_null(&ps->ps_pool_list,
2060                                               struct kib_pool,
2061                                               po_list)) != NULL) {
2062                 po->po_failed = 1;
2063                 if (po->po_allocated == 0)
2064                         list_move(&po->po_list, zombies);
2065                 else
2066                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2067         }
2068         spin_unlock(&ps->ps_lock);
2069 }
2070
2071 static void
2072 kiblnd_fini_poolset(struct kib_poolset *ps)
2073 {
2074         if (ps->ps_net != NULL) { /* initialized? */
2075                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2076                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2077         }
2078 }
2079
2080 static int
2081 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2082                     struct kib_net *net, char *name, int size,
2083                     kib_ps_pool_create_t po_create,
2084                     kib_ps_pool_destroy_t po_destroy,
2085                     kib_ps_node_init_t nd_init,
2086                     kib_ps_node_fini_t nd_fini)
2087 {
2088         struct kib_pool *pool;
2089         int rc;
2090
2091         memset(ps, 0, sizeof(struct kib_poolset));
2092
2093         ps->ps_cpt          = cpt;
2094         ps->ps_net          = net;
2095         ps->ps_pool_create  = po_create;
2096         ps->ps_pool_destroy = po_destroy;
2097         ps->ps_node_init    = nd_init;
2098         ps->ps_node_fini    = nd_fini;
2099         ps->ps_pool_size    = size;
2100         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2101             >= sizeof(ps->ps_name))
2102                 return -E2BIG;
2103         spin_lock_init(&ps->ps_lock);
2104         INIT_LIST_HEAD(&ps->ps_pool_list);
2105         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2106
2107         rc = ps->ps_pool_create(ps, size, &pool);
2108         if (rc == 0)
2109                 list_add(&pool->po_list, &ps->ps_pool_list);
2110         else
2111                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2112
2113         return rc;
2114 }
2115
2116 static int
2117 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2118 {
2119         if (pool->po_allocated != 0) /* still in use */
2120                 return 0;
2121         if (pool->po_failed)
2122                 return 1;
2123         return now >= pool->po_deadline;
2124 }
2125
2126 void
2127 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2128 {
2129         LIST_HEAD(zombies);
2130         struct kib_poolset *ps = pool->po_owner;
2131         struct kib_pool *tmp;
2132         time64_t now = ktime_get_seconds();
2133
2134         spin_lock(&ps->ps_lock);
2135
2136         if (ps->ps_node_fini != NULL)
2137                 ps->ps_node_fini(pool, node);
2138
2139         LASSERT(pool->po_allocated > 0);
2140         list_add(node, &pool->po_free_list);
2141         pool->po_allocated--;
2142
2143         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2144                 /* the first pool is persistent */
2145                 if (ps->ps_pool_list.next == &pool->po_list)
2146                         continue;
2147
2148                 if (kiblnd_pool_is_idle(pool, now))
2149                         list_move(&pool->po_list, &zombies);
2150         }
2151         spin_unlock(&ps->ps_lock);
2152
2153         if (!list_empty(&zombies))
2154                 kiblnd_destroy_pool_list(&zombies);
2155 }
2156
2157 struct list_head *
2158 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2159 {
2160         struct list_head        *node;
2161         struct kib_pool *pool;
2162         int                     rc;
2163         unsigned int            interval = 1;
2164         ktime_t time_before;
2165         unsigned int trips = 0;
2166
2167 again:
2168         spin_lock(&ps->ps_lock);
2169         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2170                 if (list_empty(&pool->po_free_list))
2171                         continue;
2172
2173                 pool->po_allocated++;
2174                 pool->po_deadline = ktime_get_seconds() +
2175                                     IBLND_POOL_DEADLINE;
2176                 node = pool->po_free_list.next;
2177                 list_del(node);
2178
2179                 if (ps->ps_node_init != NULL) {
2180                         /* still hold the lock */
2181                         ps->ps_node_init(pool, node);
2182                 }
2183                 spin_unlock(&ps->ps_lock);
2184                 return node;
2185         }
2186
2187         /* no available tx pool and ... */
2188         if (ps->ps_increasing) {
2189                 /* another thread is allocating a new pool */
2190                 spin_unlock(&ps->ps_lock);
2191                 trips++;
2192                 CDEBUG(D_NET,
2193                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2194                        ps->ps_name, interval, trips);
2195
2196                 schedule_timeout_interruptible(interval);
2197                 if (interval < cfs_time_seconds(1))
2198                         interval *= 2;
2199
2200                 goto again;
2201         }
2202
2203         if (ktime_get_seconds() < ps->ps_next_retry) {
2204                 /* someone failed recently */
2205                 spin_unlock(&ps->ps_lock);
2206                 return NULL;
2207         }
2208
2209         ps->ps_increasing = 1;
2210         spin_unlock(&ps->ps_lock);
2211
2212         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2213         time_before = ktime_get();
2214         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2215         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete\n",
2216                ktime_ms_delta(ktime_get(), time_before));
2217
2218         spin_lock(&ps->ps_lock);
2219         ps->ps_increasing = 0;
2220         if (rc == 0) {
2221                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2222         } else {
2223                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2224                 CERROR("Can't allocate new %s pool because out of memory\n",
2225                        ps->ps_name);
2226         }
2227         spin_unlock(&ps->ps_lock);
2228
2229         goto again;
2230 }
2231
2232 static void
2233 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2234 {
2235         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2236                                                tpo_pool);
2237         int i;
2238
2239         LASSERT (pool->po_allocated == 0);
2240
2241         if (tpo->tpo_tx_pages != NULL) {
2242                 kiblnd_unmap_tx_pool(tpo);
2243                 kiblnd_free_pages(tpo->tpo_tx_pages);
2244         }
2245
2246         if (tpo->tpo_tx_descs == NULL)
2247                 goto out;
2248
2249         for (i = 0; i < pool->po_size; i++) {
2250                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2251                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2252
2253                 list_del(&tx->tx_list);
2254                 if (tx->tx_pages != NULL)
2255                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2256                 if (tx->tx_frags != NULL)
2257                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2258                                            IBLND_MAX_RDMA_FRAGS);
2259                 if (tx->tx_wrq != NULL)
2260                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2261                                            IBLND_MAX_RDMA_FRAGS);
2262                 if (tx->tx_sge != NULL) {
2263                         /* +1 is for the lnet header/message itself */
2264                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2265                                            (IBLND_MAX_RDMA_FRAGS *
2266                                            wrq_sge + 1));
2267                 }
2268                 if (tx->tx_rd != NULL)
2269                         LIBCFS_FREE(tx->tx_rd,
2270                                     offsetof(struct kib_rdma_desc,
2271                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2272         }
2273
2274         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2275 out:
2276         kiblnd_fini_pool(pool);
2277         CFS_FREE_PTR(tpo);
2278 }
2279
2280 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2281 {
2282         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2283         int ntx;
2284
2285         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2286         ntx = tunables->lnd_ntx / ncpts;
2287
2288         return max(IBLND_TX_POOL, ntx);
2289 }
2290
2291 static int
2292 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2293 {
2294         int            i;
2295         int            npg;
2296         struct kib_pool *pool;
2297         struct kib_tx_pool *tpo;
2298
2299         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2300         if (tpo == NULL) {
2301                 CERROR("Failed to allocate TX pool\n");
2302                 return -ENOMEM;
2303         }
2304
2305         pool = &tpo->tpo_pool;
2306         kiblnd_init_pool(ps, pool, size);
2307         tpo->tpo_tx_descs = NULL;
2308         tpo->tpo_tx_pages = NULL;
2309
2310         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2311         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2312                 CERROR("Can't allocate tx pages: %d\n", npg);
2313                 CFS_FREE_PTR(tpo);
2314                 return -ENOMEM;
2315         }
2316
2317         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2318                          size * sizeof(struct kib_tx));
2319         if (tpo->tpo_tx_descs == NULL) {
2320                 CERROR("Can't allocate %d tx descriptors\n", size);
2321                 ps->ps_pool_destroy(pool);
2322                 return -ENOMEM;
2323         }
2324
2325         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2326
2327         for (i = 0; i < size; i++) {
2328                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2329                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2330
2331                 tx->tx_pool = tpo;
2332                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2333                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2334                                          lnet_cpt_table(), ps->ps_cpt,
2335                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2336                         if (tx->tx_pages == NULL)
2337                                 break;
2338                 }
2339
2340                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2341                                  IBLND_MAX_RDMA_FRAGS *
2342                                  sizeof(*tx->tx_frags));
2343                 if (tx->tx_frags == NULL)
2344                         break;
2345
2346                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2347
2348                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2349                                  IBLND_MAX_RDMA_FRAGS *
2350                                  sizeof(*tx->tx_wrq));
2351                 if (tx->tx_wrq == NULL)
2352                         break;
2353
2354                 /* +1 is for the lnet header/message itself */
2355                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2356                                  (IBLND_MAX_RDMA_FRAGS * wrq_sge + 1) *
2357                                  sizeof(*tx->tx_sge));
2358                 if (tx->tx_sge == NULL)
2359                         break;
2360
2361                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2362                                  offsetof(struct kib_rdma_desc,
2363                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2364                 if (tx->tx_rd == NULL)
2365                         break;
2366         }
2367
2368         if (i == size) {
2369                 kiblnd_map_tx_pool(tpo);
2370                 *pp_po = pool;
2371                 return 0;
2372         }
2373
2374         ps->ps_pool_destroy(pool);
2375         return -ENOMEM;
2376 }
2377
2378 static void
2379 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2380 {
2381         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2382                                                   struct kib_tx_poolset,
2383                                                   tps_poolset);
2384         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2385
2386         tx->tx_cookie = tps->tps_next_tx_cookie++;
2387 }
2388
2389 static void
2390 kiblnd_net_fini_pools(struct kib_net *net)
2391 {
2392         int     i;
2393
2394         cfs_cpt_for_each(i, lnet_cpt_table()) {
2395                 struct kib_tx_poolset *tps;
2396                 struct kib_fmr_poolset *fps;
2397
2398                 if (net->ibn_tx_ps != NULL) {
2399                         tps = net->ibn_tx_ps[i];
2400                         kiblnd_fini_poolset(&tps->tps_poolset);
2401                 }
2402
2403                 if (net->ibn_fmr_ps != NULL) {
2404                         fps = net->ibn_fmr_ps[i];
2405                         kiblnd_fini_fmr_poolset(fps);
2406                 }
2407         }
2408
2409         if (net->ibn_tx_ps != NULL) {
2410                 cfs_percpt_free(net->ibn_tx_ps);
2411                 net->ibn_tx_ps = NULL;
2412         }
2413
2414         if (net->ibn_fmr_ps != NULL) {
2415                 cfs_percpt_free(net->ibn_fmr_ps);
2416                 net->ibn_fmr_ps = NULL;
2417         }
2418 }
2419
2420 static int
2421 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2422                       int ncpts)
2423 {
2424         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2425 #ifdef HAVE_IB_GET_DMA_MR
2426         unsigned long   flags;
2427 #endif
2428         int             cpt;
2429         int             rc;
2430         int             i;
2431
2432         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2433
2434 #ifdef HAVE_IB_GET_DMA_MR
2435         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2436         /*
2437          * if lnd_map_on_demand is zero then we have effectively disabled
2438          * FMR or FastReg and we're using global memory regions
2439          * exclusively.
2440          */
2441         if (!tunables->lnd_map_on_demand) {
2442                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2443                                            flags);
2444                 goto create_tx_pool;
2445         }
2446
2447         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2448 #endif
2449
2450         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2451                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2452                        tunables->lnd_fmr_pool_size,
2453                        tunables->lnd_ntx / 4);
2454                 rc = -EINVAL;
2455                 goto failed;
2456         }
2457
2458         /* TX pool must be created later than FMR, see LU-2268
2459          * for details */
2460         LASSERT(net->ibn_tx_ps == NULL);
2461
2462         /* premapping can fail if ibd_nmr > 1, so we always create
2463          * FMR pool and map-on-demand if premapping failed */
2464
2465         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2466                                            sizeof(struct kib_fmr_poolset));
2467         if (net->ibn_fmr_ps == NULL) {
2468                 CERROR("Failed to allocate FMR pool array\n");
2469                 rc = -ENOMEM;
2470                 goto failed;
2471         }
2472
2473         for (i = 0; i < ncpts; i++) {
2474                 cpt = (cpts == NULL) ? i : cpts[i];
2475                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2476                                              net, tunables);
2477                 if (rc != 0) {
2478                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2479                                cpt, rc);
2480                         goto failed;
2481                 }
2482         }
2483
2484         if (i > 0)
2485                 LASSERT(i == ncpts);
2486
2487 #ifdef HAVE_IB_GET_DMA_MR
2488  create_tx_pool:
2489 #endif
2490         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2491                                           sizeof(struct kib_tx_poolset));
2492         if (net->ibn_tx_ps == NULL) {
2493                 CERROR("Failed to allocate tx pool array\n");
2494                 rc = -ENOMEM;
2495                 goto failed;
2496         }
2497
2498         for (i = 0; i < ncpts; i++) {
2499                 cpt = (cpts == NULL) ? i : cpts[i];
2500                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2501                                          cpt, net, "TX",
2502                                          kiblnd_tx_pool_size(ni, ncpts),
2503                                          kiblnd_create_tx_pool,
2504                                          kiblnd_destroy_tx_pool,
2505                                          kiblnd_tx_init, NULL);
2506                 if (rc != 0) {
2507                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2508                                cpt, rc);
2509                         goto failed;
2510                 }
2511         }
2512
2513         return 0;
2514  failed:
2515         kiblnd_net_fini_pools(net);
2516         LASSERT(rc != 0);
2517         return rc;
2518 }
2519
2520 static int
2521 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2522 {
2523         struct ib_port_attr *port_attr;
2524         int rc;
2525         unsigned long flags;
2526         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2527
2528         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2529         if (port_attr == NULL) {
2530                 CDEBUG(D_NETERROR, "Out of memory\n");
2531                 return -ENOMEM;
2532         }
2533
2534         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2535
2536         write_lock_irqsave(g_lock, flags);
2537
2538         if (rc == 0)
2539                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2540                                  ? IBLND_DEV_PORT_ACTIVE
2541                                  : IBLND_DEV_PORT_DOWN;
2542
2543         write_unlock_irqrestore(g_lock, flags);
2544         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2545
2546         if (rc != 0) {
2547                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2548                 return rc;
2549         }
2550         return 0;
2551 }
2552
2553 static inline void
2554 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2555 {
2556         struct kib_net  *net;
2557
2558         /* for health check */
2559         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2560                 if (val)
2561                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2562                                         libcfs_nidstr(&net->ibn_ni->ni_nid));
2563                 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2564         }
2565 }
2566
2567 void
2568 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2569 {
2570         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2571         struct kib_hca_dev  *hdev;
2572         unsigned long flags;
2573
2574         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2575
2576         write_lock_irqsave(g_lock, flags);
2577
2578         switch (event->event) {
2579         case IB_EVENT_DEVICE_FATAL:
2580                 CDEBUG(D_NET, "IB device fatal\n");
2581                 hdev->ibh_state = IBLND_DEV_FATAL;
2582                 kiblnd_set_ni_fatal_on(hdev, 1);
2583                 break;
2584         case IB_EVENT_PORT_ACTIVE:
2585                 CDEBUG(D_NET, "IB port active\n");
2586                 if (event->element.port_num == hdev->ibh_port) {
2587                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2588                         kiblnd_set_ni_fatal_on(hdev, 0);
2589                 }
2590                 break;
2591         case IB_EVENT_PORT_ERR:
2592                 CDEBUG(D_NET, "IB port err\n");
2593                 if (event->element.port_num == hdev->ibh_port) {
2594                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2595                         kiblnd_set_ni_fatal_on(hdev, 1);
2596                 }
2597                 break;
2598         default:
2599                 break;
2600         }
2601         write_unlock_irqrestore(g_lock, flags);
2602 }
2603
2604 static int
2605 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2606 {
2607         struct ib_device_attr *dev_attr;
2608         int rc = 0;
2609         int rc2 = 0;
2610
2611         /* It's safe to assume a HCA can handle a page size
2612          * matching that of the native system */
2613         hdev->ibh_page_shift = PAGE_SHIFT;
2614         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2615         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2616
2617 #ifndef HAVE_IB_DEVICE_ATTRS
2618         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2619         if (dev_attr == NULL) {
2620                 CERROR("Out of memory\n");
2621                 return -ENOMEM;
2622         }
2623
2624         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2625         if (rc != 0) {
2626                 CERROR("Failed to query IB device: %d\n", rc);
2627                 goto out_clean_attr;
2628         }
2629 #else
2630         dev_attr = &hdev->ibh_ibdev->attrs;
2631 #endif
2632
2633         hdev->ibh_mr_size = dev_attr->max_mr_size;
2634         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2635
2636         /* Setup device Memory Registration capabilities */
2637 #ifdef HAVE_FMR_POOL_API
2638 #ifdef HAVE_IB_DEVICE_OPS
2639         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2640             hdev->ibh_ibdev->ops.dealloc_fmr &&
2641             hdev->ibh_ibdev->ops.map_phys_fmr &&
2642             hdev->ibh_ibdev->ops.unmap_fmr) {
2643 #else
2644         if (hdev->ibh_ibdev->alloc_fmr &&
2645             hdev->ibh_ibdev->dealloc_fmr &&
2646             hdev->ibh_ibdev->map_phys_fmr &&
2647             hdev->ibh_ibdev->unmap_fmr) {
2648 #endif
2649                 LCONSOLE_INFO("Using FMR for registration\n");
2650                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2651         } else
2652 #endif /* HAVE_FMR_POOL_API */
2653         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2654                 LCONSOLE_INFO("Using FastReg for registration\n");
2655                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2656 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2657 #ifdef IB_DEVICE_SG_GAPS_REG
2658                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2659                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2660 #endif
2661 #endif
2662         } else {
2663                 rc = -ENOSYS;
2664         }
2665
2666         rc2 = kiblnd_port_get_attr(hdev);
2667         if (rc2 != 0)
2668                 return rc2;
2669
2670         if (rc != 0)
2671                 rc = -EINVAL;
2672
2673 #ifndef HAVE_IB_DEVICE_ATTRS
2674 out_clean_attr:
2675         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2676 #endif
2677
2678         if (rc == -ENOSYS)
2679                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2680                        "register memory: %d\n", rc);
2681         else if (rc == -EINVAL)
2682                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2683         return rc;
2684 }
2685
2686 #ifdef HAVE_IB_GET_DMA_MR
2687 static void
2688 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2689 {
2690         if (hdev->ibh_mrs == NULL)
2691                 return;
2692
2693         ib_dereg_mr(hdev->ibh_mrs);
2694
2695         hdev->ibh_mrs = NULL;
2696 }
2697 #endif
2698
2699 void
2700 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2701 {
2702         if (hdev->ibh_event_handler.device != NULL)
2703                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2704
2705 #ifdef HAVE_IB_GET_DMA_MR
2706         kiblnd_hdev_cleanup_mrs(hdev);
2707 #endif
2708
2709         if (hdev->ibh_pd != NULL)
2710                 ib_dealloc_pd(hdev->ibh_pd);
2711
2712         if (hdev->ibh_cmid != NULL)
2713                 rdma_destroy_id(hdev->ibh_cmid);
2714
2715         LIBCFS_FREE(hdev, sizeof(*hdev));
2716 }
2717
2718 #ifdef HAVE_IB_GET_DMA_MR
2719 static int
2720 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2721 {
2722         struct ib_mr *mr;
2723         int           acflags = IB_ACCESS_LOCAL_WRITE |
2724                                 IB_ACCESS_REMOTE_WRITE;
2725
2726         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2727         if (IS_ERR(mr)) {
2728                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2729                 kiblnd_hdev_cleanup_mrs(hdev);
2730                 return PTR_ERR(mr);
2731         }
2732
2733         hdev->ibh_mrs = mr;
2734
2735         return 0;
2736 }
2737 #endif
2738
2739 static int
2740 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2741 {       /* DUMMY */
2742         return 0;
2743 }
2744
2745 static int kiblnd_get_link_status(struct net_device *dev)
2746 {
2747         int ret = -1;
2748
2749         LASSERT(dev);
2750
2751         if (!netif_running(dev))
2752                 ret = 0;
2753         /* Some devices may not be providing link settings */
2754         else if (dev->ethtool_ops->get_link)
2755                 ret = dev->ethtool_ops->get_link(dev);
2756
2757         return ret;
2758 }
2759
2760 static int
2761 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2762 {
2763         struct rdma_cm_id  *cmid;
2764         struct sockaddr_in  srcaddr;
2765         struct sockaddr_in  dstaddr;
2766         int                 rc;
2767
2768         if (dev->ibd_hdev == NULL || /* initializing */
2769             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2770             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2771                 return 1;
2772
2773         /* XXX: it's UGLY, but I don't have better way to find
2774          * ib-bonding HCA failover because:
2775          *
2776          * a. no reliable CM event for HCA failover...
2777          * b. no OFED API to get ib_device for current net_device...
2778          *
2779          * We have only two choices at this point:
2780          *
2781          * a. rdma_bind_addr(), it will conflict with listener cmid
2782          * b. rdma_resolve_addr() to zero addr */
2783         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2784                                      RDMA_PS_TCP, IB_QPT_RC);
2785         if (IS_ERR(cmid)) {
2786                 rc = PTR_ERR(cmid);
2787                 CERROR("Failed to create cmid for failover: %d\n", rc);
2788                 return rc;
2789         }
2790
2791         memset(&srcaddr, 0, sizeof(srcaddr));
2792         srcaddr.sin_family      = AF_INET;
2793         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2794
2795         memset(&dstaddr, 0, sizeof(dstaddr));
2796         dstaddr.sin_family = AF_INET;
2797         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2798                                (struct sockaddr *)&dstaddr, 1);
2799         if (rc != 0 || cmid->device == NULL) {
2800                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2801                        dev->ibd_ifname, &dev->ibd_ifip,
2802                        cmid->device, rc);
2803                 rdma_destroy_id(cmid);
2804                 return rc;
2805         }
2806
2807         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2808         rdma_destroy_id(cmid);
2809         return rc;
2810 }
2811
2812 int
2813 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2814 {
2815         LIST_HEAD(zombie_tpo);
2816         LIST_HEAD(zombie_ppo);
2817         LIST_HEAD(zombie_fpo);
2818         struct rdma_cm_id  *cmid  = NULL;
2819         struct kib_hca_dev *hdev  = NULL;
2820         struct kib_hca_dev *old;
2821         struct ib_pd       *pd;
2822         struct kib_net *net;
2823         struct sockaddr_in  addr;
2824         struct net_device *netdev;
2825         unsigned long       flags;
2826         int                 rc = 0;
2827         int                 i;
2828         bool                set_fatal = true;
2829
2830         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2831                 dev->ibd_can_failover ||
2832                 dev->ibd_hdev == NULL);
2833
2834         rc = kiblnd_dev_need_failover(dev, ns);
2835         if (rc <= 0)
2836                 goto out;
2837
2838         if (dev->ibd_hdev != NULL &&
2839             dev->ibd_hdev->ibh_cmid != NULL) {
2840                 /* XXX it's not good to close old listener at here,
2841                  * because we can fail to create new listener.
2842                  * But we have to close it now, otherwise rdma_bind_addr
2843                  * will return EADDRINUSE... How crap! */
2844                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2845
2846                 cmid = dev->ibd_hdev->ibh_cmid;
2847                 /* make next schedule of kiblnd_dev_need_failover()
2848                  * return 1 for me */
2849                 dev->ibd_hdev->ibh_cmid  = NULL;
2850                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2851
2852                 rdma_destroy_id(cmid);
2853         }
2854
2855         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2856                                      IB_QPT_RC);
2857         if (IS_ERR(cmid)) {
2858                 rc = PTR_ERR(cmid);
2859                 CERROR("Failed to create cmid for failover: %d\n", rc);
2860                 goto out;
2861         }
2862
2863         memset(&addr, 0, sizeof(addr));
2864         addr.sin_family      = AF_INET;
2865         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2866         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2867
2868         /* Bind to failover device or port */
2869         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2870         if (rc != 0 || cmid->device == NULL) {
2871                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2872                        dev->ibd_ifname, &dev->ibd_ifip,
2873                        cmid->device, rc);
2874                 if (!rc && !cmid->device)
2875                         set_fatal = false;
2876                 rdma_destroy_id(cmid);
2877                 goto out;
2878         }
2879
2880         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2881         if (hdev == NULL) {
2882                 CERROR("Failed to allocate kib_hca_dev\n");
2883                 rdma_destroy_id(cmid);
2884                 rc = -ENOMEM;
2885                 goto out;
2886         }
2887
2888         atomic_set(&hdev->ibh_ref, 1);
2889         hdev->ibh_dev   = dev;
2890         hdev->ibh_cmid  = cmid;
2891         hdev->ibh_ibdev = cmid->device;
2892         hdev->ibh_port  = cmid->port_num;
2893
2894 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2895         pd = ib_alloc_pd(cmid->device, 0);
2896 #else
2897         pd = ib_alloc_pd(cmid->device);
2898 #endif
2899         if (IS_ERR(pd)) {
2900                 rc = PTR_ERR(pd);
2901                 CERROR("Can't allocate PD: %d\n", rc);
2902                 goto out;
2903         }
2904
2905         hdev->ibh_pd = pd;
2906
2907         rc = rdma_listen(cmid, 0);
2908         if (rc != 0) {
2909                 CERROR("Can't start new listener: %d\n", rc);
2910                 goto out;
2911         }
2912
2913         rc = kiblnd_hdev_get_attr(hdev);
2914         if (rc != 0) {
2915                 CERROR("Can't get device attributes: %d\n", rc);
2916                 goto out;
2917         }
2918
2919 #ifdef HAVE_IB_GET_DMA_MR
2920         rc = kiblnd_hdev_setup_mrs(hdev);
2921         if (rc != 0) {
2922                 CERROR("Can't setup device: %d\n", rc);
2923                 goto out;
2924         }
2925 #endif
2926
2927         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2928                                 hdev->ibh_ibdev, kiblnd_event_handler);
2929         ib_register_event_handler(&hdev->ibh_event_handler);
2930
2931         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2932
2933         old = dev->ibd_hdev;
2934         dev->ibd_hdev = hdev;   /* take over the refcount */
2935         hdev = old;
2936
2937         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2938                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2939                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2940                                             &zombie_tpo);
2941
2942                         if (net->ibn_fmr_ps != NULL)
2943                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2944                                                         &zombie_fpo);
2945                 }
2946         }
2947
2948         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2949  out:
2950         if (!list_empty(&zombie_tpo))
2951                 kiblnd_destroy_pool_list(&zombie_tpo);
2952         if (!list_empty(&zombie_ppo))
2953                 kiblnd_destroy_pool_list(&zombie_ppo);
2954         if (!list_empty(&zombie_fpo))
2955                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2956         if (hdev != NULL)
2957                 kiblnd_hdev_decref(hdev);
2958
2959         if (rc != 0) {
2960                 dev->ibd_failed_failover++;
2961         } else {
2962                 dev->ibd_failed_failover = 0;
2963
2964                 if (set_fatal) {
2965                         rcu_read_lock();
2966                         netdev = dev_get_by_name_rcu(ns, dev->ibd_ifname);
2967                         if (netdev && (kiblnd_get_link_status(netdev) == 1))
2968                                 kiblnd_set_ni_fatal_on(dev->ibd_hdev, 0);
2969                         rcu_read_unlock();
2970                 }
2971         }
2972
2973         return rc;
2974 }
2975
2976 void
2977 kiblnd_destroy_dev(struct kib_dev *dev)
2978 {
2979         LASSERT(dev->ibd_nnets == 0);
2980         LASSERT(list_empty(&dev->ibd_nets));
2981
2982         list_del(&dev->ibd_fail_list);
2983         list_del(&dev->ibd_list);
2984
2985         if (dev->ibd_hdev != NULL)
2986                 kiblnd_hdev_decref(dev->ibd_hdev);
2987
2988         LIBCFS_FREE(dev, sizeof(*dev));
2989 }
2990
2991 static struct kib_dev *
2992 kiblnd_dev_search(char *ifname)
2993 {
2994         struct kib_dev *alias = NULL;
2995         struct kib_dev *dev;
2996         char *colon;
2997         char *colon2;
2998
2999         colon = strchr(ifname, ':');
3000         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3001                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3002                         return dev;
3003
3004                 if (alias != NULL)
3005                         continue;
3006
3007                 colon2 = strchr(dev->ibd_ifname, ':');
3008                 if (colon != NULL)
3009                         *colon = 0;
3010                 if (colon2 != NULL)
3011                         *colon2 = 0;
3012
3013                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3014                         alias = dev;
3015
3016                 if (colon != NULL)
3017                         *colon = ':';
3018                 if (colon2 != NULL)
3019                         *colon2 = ':';
3020         }
3021         return alias;
3022 }
3023
3024 static int
3025 kiblnd_handle_link_state_change(struct net_device *dev,
3026                                 unsigned char operstate)
3027 {
3028         struct lnet_ni *ni = NULL;
3029         struct kib_dev *event_kibdev;
3030         struct kib_net *net;
3031         struct kib_net *cnxt;
3032         bool link_down = !(operstate == IF_OPER_UP);
3033         struct in_device *in_dev;
3034         bool found_ip = false;
3035         DECLARE_CONST_IN_IFADDR(ifa);
3036
3037         event_kibdev = kiblnd_dev_search(dev->name);
3038
3039         if (!event_kibdev)
3040                 goto out;
3041
3042         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets, ibn_list) {
3043                 found_ip = false;
3044
3045                 ni = net->ibn_ni;
3046
3047                 in_dev = __in_dev_get_rtnl(dev);
3048                 if (!in_dev) {
3049                         CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
3050                                dev->name);
3051                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3052                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3053                         atomic_set(&ni->ni_fatal_error_on, 1);
3054                         continue;
3055                 }
3056                 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
3057                         if (htonl(event_kibdev->ibd_ifip) == ifa->ifa_local)
3058                                 found_ip = true;
3059                 }
3060                 endfor_ifa(in_dev);
3061
3062                 if (!found_ip) {
3063                         CDEBUG(D_NET, "Interface %s has no matching ip\n",
3064                                dev->name);
3065                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3066                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3067                         atomic_set(&ni->ni_fatal_error_on, 1);
3068                         continue;
3069                 }
3070
3071                 if (link_down) {
3072                         CDEBUG(D_NET, "%s: set link fatal state to 1\n",
3073                                libcfs_nidstr(&net->ibn_ni->ni_nid));
3074                         atomic_set(&ni->ni_fatal_error_on, link_down);
3075                 } else {
3076                         CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3077                                libcfs_nidstr(&net->ibn_ni->ni_nid),
3078                                (kiblnd_get_link_status(dev) == 0));
3079                         atomic_set(&ni->ni_fatal_error_on,
3080                                    (kiblnd_get_link_status(dev) == 0));
3081                 }
3082         }
3083 out:
3084         return 0;
3085 }
3086
3087 static int
3088 kiblnd_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
3089 {
3090         struct kib_dev *event_kibdev;
3091         struct kib_net *net;
3092         struct kib_net *cnxt;
3093         struct net_device *event_netdev = ifa->ifa_dev->dev;
3094
3095         event_kibdev = kiblnd_dev_search(event_netdev->name);
3096
3097         if (!event_kibdev)
3098                 goto out;
3099
3100         if (htonl(event_kibdev->ibd_ifip) != ifa->ifa_local)
3101                 goto out;
3102
3103         list_for_each_entry_safe(net, cnxt, &event_kibdev->ibd_nets,
3104                                  ibn_list) {
3105                 CDEBUG(D_NET, "%s: set link fatal state to %u\n",
3106                        libcfs_nidstr(&net->ibn_ni->ni_nid),
3107                        (event == NETDEV_DOWN));
3108                 atomic_set(&net->ibn_ni->ni_fatal_error_on,
3109                            (event == NETDEV_DOWN));
3110         }
3111 out:
3112         return 0;
3113 }
3114
3115
3116 /************************************
3117  * Net device notifier event handler
3118  ************************************/
3119 static int kiblnd_device_event(struct notifier_block *unused,
3120                                  unsigned long event, void *ptr)
3121 {
3122         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
3123         unsigned char operstate;
3124
3125         operstate = dev->operstate;
3126
3127         CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
3128                event, dev->name, dev->ifindex, operstate);
3129
3130         switch (event) {
3131         case NETDEV_UP:
3132         case NETDEV_DOWN:
3133         case NETDEV_CHANGE:
3134                 kiblnd_handle_link_state_change(dev, operstate);
3135                 break;
3136         }
3137
3138         return NOTIFY_OK;
3139 }
3140
3141 /************************************
3142  * Inetaddr notifier event handler
3143  ************************************/
3144 static int kiblnd_inetaddr_event(struct notifier_block *unused,
3145                                  unsigned long event, void *ptr)
3146 {
3147         struct in_ifaddr *ifa = ptr;
3148
3149         CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
3150                event, &ifa->ifa_address, &ifa->ifa_mask);
3151
3152         switch (event) {
3153         case NETDEV_UP:
3154         case NETDEV_DOWN:
3155         case NETDEV_CHANGE:
3156                 kiblnd_handle_inetaddr_change(ifa, event);
3157                 break;
3158
3159         }
3160         return NOTIFY_OK;
3161 }
3162
3163 static struct notifier_block kiblnd_dev_notifier_block = {
3164         .notifier_call = kiblnd_device_event,
3165 };
3166
3167 static struct notifier_block kiblnd_inetaddr_notifier_block = {
3168         .notifier_call = kiblnd_inetaddr_event,
3169 };
3170
3171 static void
3172 kiblnd_base_shutdown(void)
3173 {
3174         struct kib_sched_info *sched;
3175         struct kib_peer_ni *peer_ni;
3176         int i;
3177
3178         LASSERT(list_empty(&kiblnd_data.kib_devs));
3179
3180         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %lld\n",
3181                libcfs_kmem_read());
3182
3183         if (kiblnd_data.kib_init == IBLND_INIT_ALL) {
3184                 unregister_netdevice_notifier(&kiblnd_dev_notifier_block);
3185                 unregister_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3186         }
3187
3188         switch (kiblnd_data.kib_init) {
3189         default:
3190                 LBUG();
3191
3192         case IBLND_INIT_ALL:
3193         case IBLND_INIT_DATA:
3194                 hash_for_each(kiblnd_data.kib_peers, i, peer_ni, ibp_list)
3195                         LASSERT(0);
3196                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
3197                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
3198                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
3199                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
3200
3201                 /* flag threads to terminate; wake and wait for them to die */
3202                 kiblnd_data.kib_shutdown = 1;
3203
3204                 /* NB: we really want to stop scheduler threads net by net
3205                  * instead of the whole module, this should be improved
3206                  * with dynamic configuration LNet.
3207                  */
3208                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
3209                         wake_up_all(&sched->ibs_waitq);
3210
3211                 wake_up(&kiblnd_data.kib_connd_waitq);
3212                 wake_up(&kiblnd_data.kib_failover_waitq);
3213
3214                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
3215                                        !atomic_read(&kiblnd_data.kib_nthreads),
3216                                        "Waiting for %d threads to terminate\n",
3217                                        atomic_read(&kiblnd_data.kib_nthreads));
3218                 fallthrough;
3219
3220         case IBLND_INIT_NOTHING:
3221                 break;
3222         }
3223
3224         if (kiblnd_data.kib_scheds != NULL)
3225                 cfs_percpt_free(kiblnd_data.kib_scheds);
3226
3227         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %lld\n",
3228                libcfs_kmem_read());
3229
3230         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
3231         module_put(THIS_MODULE);
3232 }
3233
3234 static void
3235 kiblnd_shutdown(struct lnet_ni *ni)
3236 {
3237         struct kib_net *net = ni->ni_data;
3238         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3239         unsigned long     flags;
3240
3241         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3242
3243         if (net == NULL)
3244                 goto out;
3245
3246         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %lld\n",
3247                libcfs_kmem_read());
3248
3249         write_lock_irqsave(g_lock, flags);
3250         net->ibn_shutdown = 1;
3251         write_unlock_irqrestore(g_lock, flags);
3252
3253         switch (net->ibn_init) {
3254         default:
3255                 LBUG();
3256
3257         case IBLND_INIT_ALL:
3258                 /* nuke all existing peers within this net */
3259                 kiblnd_del_peer(ni, LNET_NID_ANY);
3260
3261                 /* Wait for all peer_ni state to clean up */
3262                 wait_var_event_warning(&net->ibn_npeers,
3263                                        atomic_read(&net->ibn_npeers) == 0,
3264                                        "%s: waiting for %d peers to disconnect\n",
3265                                        libcfs_nidstr(&ni->ni_nid),
3266                                        atomic_read(&net->ibn_npeers));
3267
3268                 kiblnd_net_fini_pools(net);
3269
3270                 write_lock_irqsave(g_lock, flags);
3271                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3272                 net->ibn_dev->ibd_nnets--;
3273                 list_del(&net->ibn_list);
3274                 write_unlock_irqrestore(g_lock, flags);
3275
3276                 wake_up_all(&kiblnd_data.kib_connd_waitq);
3277                 wait_var_event_warning(&net->ibn_nconns,
3278                                        atomic_read(&net->ibn_nconns) == 0,
3279                                        "%s: waiting for %d conns to clean\n",
3280                                        libcfs_nidstr(&ni->ni_nid),
3281                                        atomic_read(&net->ibn_nconns));
3282                 fallthrough;
3283
3284         case IBLND_INIT_NOTHING:
3285                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3286
3287                 if (net->ibn_dev != NULL &&
3288                     net->ibn_dev->ibd_nnets == 0)
3289                         kiblnd_destroy_dev(net->ibn_dev);
3290
3291                 break;
3292         }
3293
3294         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %lld\n",
3295                libcfs_kmem_read());
3296
3297         net->ibn_init = IBLND_INIT_NOTHING;
3298         ni->ni_data = NULL;
3299
3300         LIBCFS_FREE(net, sizeof(*net));
3301
3302 out:
3303         if (list_empty(&kiblnd_data.kib_devs))
3304                 kiblnd_base_shutdown();
3305 }
3306
3307 static int
3308 kiblnd_base_startup(struct net *ns)
3309 {
3310         struct kib_sched_info *sched;
3311         int rc;
3312         int i;
3313
3314         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3315
3316         if (!try_module_get(THIS_MODULE))
3317                 goto failed;
3318
3319         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3320
3321         rwlock_init(&kiblnd_data.kib_global_lock);
3322
3323         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3324         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3325
3326         hash_init(kiblnd_data.kib_peers);
3327
3328         spin_lock_init(&kiblnd_data.kib_connd_lock);
3329         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3330         INIT_LIST_HEAD(&kiblnd_data.kib_connd_waits);
3331         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3332         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3333         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3334
3335         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3336         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3337
3338         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3339                                                   sizeof(*sched));
3340         if (kiblnd_data.kib_scheds == NULL)
3341                 goto failed;
3342
3343         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3344                 int     nthrs;
3345
3346                 spin_lock_init(&sched->ibs_lock);
3347                 INIT_LIST_HEAD(&sched->ibs_conns);
3348                 init_waitqueue_head(&sched->ibs_waitq);
3349
3350                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3351                 if (*kiblnd_tunables.kib_nscheds > 0) {
3352                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3353                 } else {
3354                         /* max to half of CPUs, another half is reserved for
3355                          * upper layer modules */
3356                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3357                 }
3358
3359                 sched->ibs_nthreads_max = nthrs;
3360                 sched->ibs_cpt = i;
3361         }
3362
3363         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3364
3365         /* lists/ptrs/locks initialised */
3366         kiblnd_data.kib_init = IBLND_INIT_DATA;
3367         /*****************************************************/
3368
3369         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3370         if (rc != 0) {
3371                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3372                 goto failed;
3373         }
3374
3375         if (*kiblnd_tunables.kib_dev_failover != 0)
3376                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3377                                          "kiblnd_failover");
3378
3379         if (rc != 0) {
3380                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3381                 goto failed;
3382         }
3383
3384         register_netdevice_notifier(&kiblnd_dev_notifier_block);
3385         register_inetaddr_notifier(&kiblnd_inetaddr_notifier_block);
3386
3387         /* flag everything initialised */
3388         kiblnd_data.kib_init = IBLND_INIT_ALL;
3389         /*****************************************************/
3390
3391         return 0;
3392
3393  failed:
3394         kiblnd_base_shutdown();
3395         return -ENETDOWN;
3396 }
3397
3398 static int
3399 kiblnd_start_schedulers(struct kib_sched_info *sched)
3400 {
3401         int     rc = 0;
3402         int     nthrs;
3403         int     i;
3404
3405         if (sched->ibs_nthreads == 0) {
3406                 if (*kiblnd_tunables.kib_nscheds > 0) {
3407                         nthrs = sched->ibs_nthreads_max;
3408                 } else {
3409                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3410                                                sched->ibs_cpt);
3411                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3412                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3413                 }
3414         } else {
3415                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3416                 /* increase one thread if there is new interface */
3417                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3418         }
3419
3420         for (i = 0; i < nthrs; i++) {
3421                 long    id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3422
3423                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id,
3424                                          "kiblnd_sd_%02ld_%02ld",
3425                                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3426                 if (rc == 0)
3427                         continue;
3428
3429                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3430                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3431                 break;
3432         }
3433
3434         sched->ibs_nthreads += i;
3435         return rc;
3436 }
3437
3438 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3439                                     int ncpts)
3440 {
3441         int     cpt;
3442         int     rc;
3443         int     i;
3444
3445         for (i = 0; i < ncpts; i++) {
3446                 struct kib_sched_info *sched;
3447
3448                 cpt = (cpts == NULL) ? i : cpts[i];
3449                 sched = kiblnd_data.kib_scheds[cpt];
3450
3451                 if (!newdev && sched->ibs_nthreads > 0)
3452                         continue;
3453
3454                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3455                 if (rc != 0) {
3456                         CERROR("Failed to start scheduler threads for %s\n",
3457                                dev->ibd_ifname);
3458                         return rc;
3459                 }
3460         }
3461         return 0;
3462 }
3463
3464 static int
3465 kiblnd_startup(struct lnet_ni *ni)
3466 {
3467         char *ifname = NULL;
3468         struct lnet_inetdev *ifaces = NULL;
3469         struct kib_dev *ibdev = NULL;
3470         struct kib_net *net = NULL;
3471         unsigned long flags;
3472         int rc;
3473         int i;
3474         bool newdev;
3475
3476         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3477
3478         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3479                 rc = kiblnd_base_startup(ni->ni_net_ns);
3480                 if (rc != 0)
3481                         return rc;
3482         }
3483
3484         LIBCFS_ALLOC(net, sizeof(*net));
3485         ni->ni_data = net;
3486         if (net == NULL) {
3487                 rc = -ENOMEM;
3488                 goto failed;
3489         }
3490
3491         net->ibn_ni = ni;
3492         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3493
3494         kiblnd_tunables_setup(ni);
3495
3496         /*
3497          * Multi-Rail wants each secondary
3498          * IP to be treated as an unique 'struct ni' interface.
3499          */
3500         if (ni->ni_interface != NULL) {
3501                 /* Use the IPoIB interface specified in 'networks=' */
3502                 ifname = ni->ni_interface;
3503         } else {
3504                 ifname = *kiblnd_tunables.kib_default_ipif;
3505         }
3506
3507         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3508                 CERROR("IPoIB interface name too long: %s\n", ifname);
3509                 rc = -E2BIG;
3510                 goto failed;
3511         }
3512
3513         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, false);
3514         if (rc < 0)
3515                 goto failed;
3516
3517         for (i = 0; i < rc; i++) {
3518                 if (strcmp(ifname, ifaces[i].li_name) == 0)
3519                         break;
3520         }
3521
3522         if (i == rc) {
3523                 CERROR("ko2iblnd: No matching interfaces\n");
3524                 rc = -ENOENT;
3525                 goto failed;
3526         }
3527
3528         ibdev = kiblnd_dev_search(ifname);
3529         newdev = ibdev == NULL;
3530         /* hmm...create kib_dev even for alias */
3531         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3532                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3533                 if (!ibdev) {
3534                         rc = -ENOMEM;
3535                         goto failed;
3536                 }
3537
3538                 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3539                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3540                         sizeof(ibdev->ibd_ifname));
3541                 ibdev->ibd_can_failover = ifaces[i].li_iff_master;
3542
3543                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3544                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3545                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3546
3547                 /* initialize the device */
3548                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3549                 if (rc) {
3550                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3551                                rc);
3552                         goto failed;
3553                 }
3554
3555                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3556         }
3557
3558         net->ibn_dev = ibdev;
3559         ni->ni_nid.nid_addr[0] = cpu_to_be32(ibdev->ibd_ifip);
3560
3561         ni->ni_dev_cpt = ifaces[i].li_cpt;
3562
3563         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3564         if (rc != 0)
3565                 goto failed;
3566
3567         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3568         if (rc != 0) {
3569                 CERROR("Failed to initialize NI pools: %d\n", rc);
3570                 goto failed;
3571         }
3572
3573         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3574         ibdev->ibd_nnets++;
3575         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3576         /* for health check */
3577         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3578                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3579         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3580
3581         net->ibn_init = IBLND_INIT_ALL;
3582
3583         return 0;
3584
3585 failed:
3586         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3587                 kiblnd_destroy_dev(ibdev);
3588
3589         kfree(ifaces);
3590         kiblnd_shutdown(ni);
3591
3592         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3593                ifname ? ifname : "", rc);
3594
3595         return -ENETDOWN;
3596 }
3597
3598 static const struct lnet_lnd the_o2iblnd = {
3599         .lnd_type       = O2IBLND,
3600         .lnd_startup    = kiblnd_startup,
3601         .lnd_shutdown   = kiblnd_shutdown,
3602         .lnd_ctl        = kiblnd_ctl,
3603         .lnd_send       = kiblnd_send,
3604         .lnd_recv       = kiblnd_recv,
3605         .lnd_get_dev_prio = kiblnd_get_dev_prio,
3606 };
3607
3608 static void ko2inlnd_assert_wire_constants(void)
3609 {
3610         BUILD_BUG_ON(IBLND_MSG_MAGIC != 0x0be91b91);
3611         BUILD_BUG_ON(IBLND_MSG_VERSION_1 != 0x11);
3612         BUILD_BUG_ON(IBLND_MSG_VERSION_2 != 0x12);
3613         BUILD_BUG_ON(IBLND_MSG_VERSION != IBLND_MSG_VERSION_2);
3614
3615         BUILD_BUG_ON(IBLND_MSG_CONNREQ != 0xc0);
3616         BUILD_BUG_ON(IBLND_MSG_CONNACK != 0xc1);
3617         BUILD_BUG_ON(IBLND_MSG_NOOP != 0xd0);
3618         BUILD_BUG_ON(IBLND_MSG_IMMEDIATE != 0xd1);
3619         BUILD_BUG_ON(IBLND_MSG_PUT_REQ != 0xd2);
3620         BUILD_BUG_ON(IBLND_MSG_PUT_NAK != 0xd3);
3621         BUILD_BUG_ON(IBLND_MSG_PUT_ACK != 0xd4);
3622         BUILD_BUG_ON(IBLND_MSG_PUT_DONE != 0xd5);
3623         BUILD_BUG_ON(IBLND_MSG_GET_REQ != 0xd6);
3624         BUILD_BUG_ON(IBLND_MSG_GET_DONE != 0xd7);
3625
3626         BUILD_BUG_ON(IBLND_REJECT_CONN_RACE != 1);
3627         BUILD_BUG_ON(IBLND_REJECT_NO_RESOURCES != 2);
3628         BUILD_BUG_ON(IBLND_REJECT_FATAL != 3);
3629         BUILD_BUG_ON(IBLND_REJECT_CONN_UNCOMPAT != 4);
3630         BUILD_BUG_ON(IBLND_REJECT_CONN_STALE != 5);
3631         BUILD_BUG_ON(IBLND_REJECT_RDMA_FRAGS != 6);
3632         BUILD_BUG_ON(IBLND_REJECT_MSG_QUEUE_SIZE != 7);
3633         BUILD_BUG_ON(IBLND_REJECT_INVALID_SRV_ID != 8);
3634
3635         BUILD_BUG_ON((int)sizeof(struct kib_connparams) != 8);
3636         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_queue_depth) != 0);
3637         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_queue_depth) != 2);
3638         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_frags) != 2);
3639         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_frags) != 2);
3640         BUILD_BUG_ON((int)offsetof(struct kib_connparams, ibcp_max_msg_size) != 4);
3641         BUILD_BUG_ON((int)sizeof(((struct kib_connparams *)0)->ibcp_max_msg_size) != 4);
3642
3643         BUILD_BUG_ON((int)sizeof(struct kib_immediate_msg) != 72);
3644         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_hdr) != 0);
3645         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_hdr) != 72);
3646         BUILD_BUG_ON((int)offsetof(struct kib_immediate_msg, ibim_payload) != 72);
3647         BUILD_BUG_ON((int)sizeof(((struct kib_immediate_msg *)0)->ibim_payload) != 0);
3648
3649         BUILD_BUG_ON((int)sizeof(struct kib_rdma_frag) != 12);
3650         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_nob) != 0);
3651         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_nob) != 4);
3652         BUILD_BUG_ON((int)offsetof(struct kib_rdma_frag, rf_addr) != 4);
3653         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_frag *)0)->rf_addr) != 8);
3654
3655         BUILD_BUG_ON((int)sizeof(struct kib_rdma_desc) != 8);
3656         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_key) != 0);
3657         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_key) != 4);
3658         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_nfrags) != 4);
3659         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_nfrags) != 4);
3660         BUILD_BUG_ON((int)offsetof(struct kib_rdma_desc, rd_frags) != 8);
3661         BUILD_BUG_ON((int)sizeof(((struct kib_rdma_desc *)0)->rd_frags) != 0);
3662
3663         BUILD_BUG_ON((int)sizeof(struct kib_putreq_msg) != 80);
3664         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_hdr) != 0);
3665         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_hdr) != 72);
3666         BUILD_BUG_ON((int)offsetof(struct kib_putreq_msg, ibprm_cookie) != 72);
3667         BUILD_BUG_ON((int)sizeof(((struct kib_putreq_msg *)0)->ibprm_cookie) != 8);
3668
3669         BUILD_BUG_ON((int)sizeof(struct kib_putack_msg) != 24);
3670         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_src_cookie) != 0);
3671         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_src_cookie) != 8);
3672         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_dst_cookie) != 8);
3673         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_dst_cookie) != 8);
3674         BUILD_BUG_ON((int)offsetof(struct kib_putack_msg, ibpam_rd) != 16);
3675         BUILD_BUG_ON((int)sizeof(((struct kib_putack_msg *)0)->ibpam_rd) != 8);
3676
3677         BUILD_BUG_ON((int)sizeof(struct kib_get_msg) != 88);
3678         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_hdr) != 0);
3679         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_hdr) != 72);
3680         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_cookie) != 72);
3681         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_cookie) != 8);
3682         BUILD_BUG_ON((int)offsetof(struct kib_get_msg, ibgm_rd) != 80);
3683         BUILD_BUG_ON((int)sizeof(((struct kib_get_msg *)0)->ibgm_rd) != 8);
3684
3685         BUILD_BUG_ON((int)sizeof(struct kib_completion_msg) != 12);
3686         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_cookie) != 0);
3687         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_cookie) != 8);
3688         BUILD_BUG_ON((int)offsetof(struct kib_completion_msg, ibcm_status) != 8);
3689         BUILD_BUG_ON((int)sizeof(((struct kib_completion_msg *)0)->ibcm_status) != 4);
3690
3691         /* Checks for struct kib_msg */
3692         //BUILD_BUG_ON((int)sizeof(struct kib_msg) != 12);
3693         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_magic) != 0);
3694         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_magic) != 4);
3695         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_version) != 4);
3696         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_version) != 2);
3697         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_type) != 6);
3698         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_type) != 1);
3699         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_credits) != 7);
3700         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_credits) != 1);
3701         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_nob) != 8);
3702         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_nob) != 4);
3703         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_cksum) != 12);
3704         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_cksum) != 4);
3705         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcnid) != 16);
3706         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcnid) != 8);
3707         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_srcstamp) != 24);
3708         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_srcstamp) != 8);
3709         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dstnid) != 32);
3710         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dstnid) != 8);
3711         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_dststamp) != 40);
3712         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_dststamp) != 8);
3713
3714         /* Connparams */
3715         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_queue_depth) != 48);
3716         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_queue_depth) != 2);
3717         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_frags) != 50);
3718         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_frags) != 2);
3719         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.connparams.ibcp_max_msg_size) != 52);
3720         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.connparams.ibcp_max_msg_size) != 4);
3721
3722         /* Immediate message */
3723         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_hdr) != 48);
3724         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_hdr) != 72);
3725         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.immediate.ibim_payload) != 120);
3726         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.immediate.ibim_payload) != 0);
3727
3728         /* PUT req message */
3729         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_hdr) != 48);
3730         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_hdr) != 72);
3731         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putreq.ibprm_cookie) != 120);
3732         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putreq.ibprm_cookie) != 8);
3733
3734         /* Put ACK */
3735         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_src_cookie) != 48);
3736         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_src_cookie) != 8);
3737         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_dst_cookie) != 56);
3738         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_dst_cookie) != 8);
3739         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.putack.ibpam_rd) != 64);
3740         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.putack.ibpam_rd) != 8);
3741
3742         /* GET message */
3743         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_hdr) != 48);
3744         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_hdr) != 72);
3745         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_cookie) != 120);
3746         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_cookie) != 8);
3747         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.get.ibgm_rd) != 128);
3748         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.get.ibgm_rd) != 8);
3749
3750         /* Completion message */
3751         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_cookie) != 48);
3752         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_cookie) != 8);
3753         BUILD_BUG_ON((int)offsetof(struct kib_msg, ibm_u.completion.ibcm_status) != 56);
3754         BUILD_BUG_ON((int)sizeof(((struct kib_msg *)0)->ibm_u.completion.ibcm_status) != 4);
3755
3756         /* Sanity checks */
3757         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3758         BUILD_BUG_ON(offsetof(struct kib_msg,
3759                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3760                      IBLND_MSG_SIZE);
3761         BUILD_BUG_ON(offsetof(struct kib_msg,
3762                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3763                      IBLND_MSG_SIZE);
3764 }
3765
3766 static void __exit ko2iblnd_exit(void)
3767 {
3768         lnet_unregister_lnd(&the_o2iblnd);
3769 }
3770
3771 static int __init ko2iblnd_init(void)
3772 {
3773         int rc;
3774
3775         ko2inlnd_assert_wire_constants();
3776
3777         rc = kiblnd_tunables_init();
3778         if (rc != 0)
3779                 return rc;
3780
3781         lnet_register_lnd(&the_o2iblnd);
3782
3783         return 0;
3784 }
3785
3786 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3787 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3788 MODULE_VERSION("2.8.0");
3789 MODULE_LICENSE("GPL");
3790
3791 module_init(ko2iblnd_init);
3792 module_exit(ko2iblnd_exit);