drivers/staging/lustre/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, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include <asm/div64.h>
  38 #include <asm/page.h>
  39 #include "o2iblnd.h"
  40
  41 static lnd_t the_o2iblnd;
  42
  43 struct kib_data kiblnd_data;
  44
  45 static __u32 kiblnd_cksum(void *ptr, int nob)
  46 {
  47         char *c = ptr;
  48         __u32 sum = 0;
  49
  50         while (nob-- > 0)
  51                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  52
  53         /* ensure I don't return 0 (== no checksum) */
  54         return !sum ? 1 : sum;
  55 }
  56
  57 static char *kiblnd_msgtype2str(int type)
  58 {
  59         switch (type) {
  60         case IBLND_MSG_CONNREQ:
  61                 return "CONNREQ";
  62
  63         case IBLND_MSG_CONNACK:
  64                 return "CONNACK";
  65
  66         case IBLND_MSG_NOOP:
  67                 return "NOOP";
  68
  69         case IBLND_MSG_IMMEDIATE:
  70                 return "IMMEDIATE";
  71
  72         case IBLND_MSG_PUT_REQ:
  73                 return "PUT_REQ";
  74
  75         case IBLND_MSG_PUT_NAK:
  76                 return "PUT_NAK";
  77
  78         case IBLND_MSG_PUT_ACK:
  79                 return "PUT_ACK";
  80
  81         case IBLND_MSG_PUT_DONE:
  82                 return "PUT_DONE";
  83
  84         case IBLND_MSG_GET_REQ:
  85                 return "GET_REQ";
  86
  87         case IBLND_MSG_GET_DONE:
  88                 return "GET_DONE";
  89
  90         default:
  91                 return "???";
  92         }
  93 }
  94
  95 static int kiblnd_msgtype2size(int type)
  96 {
  97         const int hdr_size = offsetof(struct kib_msg, ibm_u);
  98
  99         switch (type) {
 100         case IBLND_MSG_CONNREQ:
 101         case IBLND_MSG_CONNACK:
 102                 return hdr_size + sizeof(struct kib_connparams);
 103
 104         case IBLND_MSG_NOOP:
 105                 return hdr_size;
 106
 107         case IBLND_MSG_IMMEDIATE:
 108                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 109
 110         case IBLND_MSG_PUT_REQ:
 111                 return hdr_size + sizeof(struct kib_putreq_msg);
 112
 113         case IBLND_MSG_PUT_ACK:
 114                 return hdr_size + sizeof(struct kib_putack_msg);
 115
 116         case IBLND_MSG_GET_REQ:
 117                 return hdr_size + sizeof(struct kib_get_msg);
 118
 119         case IBLND_MSG_PUT_NAK:
 120         case IBLND_MSG_PUT_DONE:
 121         case IBLND_MSG_GET_DONE:
 122                 return hdr_size + sizeof(struct kib_completion_msg);
 123         default:
 124                 return -1;
 125         }
 126 }
 127
 128 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
 129 {
 130         struct kib_rdma_desc *rd;
 131         int msg_size;
 132         int nob;
 133         int n;
 134         int i;
 135
 136         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 137                 msg->ibm_type == IBLND_MSG_PUT_ACK);
 138
 139         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 140                               &msg->ibm_u.get.ibgm_rd :
 141                               &msg->ibm_u.putack.ibpam_rd;
 142
 143         if (flip) {
 144                 __swab32s(&rd->rd_key);
 145                 __swab32s(&rd->rd_nfrags);
 146         }
 147
 148         n = rd->rd_nfrags;
 149
 150         nob = offsetof(struct kib_msg, ibm_u) +
 151               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 152
 153         if (msg->ibm_nob < nob) {
 154                 CERROR("Short %s: %d(%d)\n",
 155                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 156                 return 1;
 157         }
 158
 159         msg_size = kiblnd_rd_size(rd);
 160         if (msg_size <= 0 || msg_size > LNET_MAX_PAYLOAD) {
 161                 CERROR("Bad msg_size: %d, should be 0 < n <= %d\n",
 162                        msg_size, LNET_MAX_PAYLOAD);
 163                 return 1;
 164         }
 165
 166         if (!flip)
 167                 return 0;
 168
 169         for (i = 0; i < n; i++) {
 170                 __swab32s(&rd->rd_frags[i].rf_nob);
 171                 __swab64s(&rd->rd_frags[i].rf_addr);
 172         }
 173
 174         return 0;
 175 }
 176
 177 void kiblnd_pack_msg(lnet_ni_t *ni, struct kib_msg *msg, int version,
 178                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 179 {
 180         struct kib_net *net = ni->ni_data;
 181
 182         /*
 183          * CAVEAT EMPTOR! all message fields not set here should have been
 184          * initialised previously.
 185          */
 186         msg->ibm_magic    = IBLND_MSG_MAGIC;
 187         msg->ibm_version  = version;
 188         /*   ibm_type */
 189         msg->ibm_credits  = credits;
 190         /*   ibm_nob */
 191         msg->ibm_cksum    = 0;
 192         msg->ibm_srcnid   = ni->ni_nid;
 193         msg->ibm_srcstamp = net->ibn_incarnation;
 194         msg->ibm_dstnid   = dstnid;
 195         msg->ibm_dststamp = dststamp;
 196
 197         if (*kiblnd_tunables.kib_cksum) {
 198                 /* NB ibm_cksum zero while computing cksum */
 199                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 200         }
 201 }
 202
 203 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 204 {
 205         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 206         __u32 msg_cksum;
 207         __u16 version;
 208         int msg_nob;
 209         int flip;
 210
 211         /* 6 bytes are enough to have received magic + version */
 212         if (nob < 6) {
 213                 CERROR("Short message: %d\n", nob);
 214                 return -EPROTO;
 215         }
 216
 217         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 218                 flip = 0;
 219         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 220                 flip = 1;
 221         } else {
 222                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 223                 return -EPROTO;
 224         }
 225
 226         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 227         if (version != IBLND_MSG_VERSION &&
 228             version != IBLND_MSG_VERSION_1) {
 229                 CERROR("Bad version: %x\n", version);
 230                 return -EPROTO;
 231         }
 232
 233         if (nob < hdr_size) {
 234                 CERROR("Short message: %d\n", nob);
 235                 return -EPROTO;
 236         }
 237
 238         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 239         if (msg_nob > nob) {
 240                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 241                 return -EPROTO;
 242         }
 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 &&
 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 endianness */
 260                 msg->ibm_version = version;
 261                 CLASSERT(sizeof(msg->ibm_type) == 1);
 262                 CLASSERT(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 kiblnd_create_peer(lnet_ni_t *ni, struct kib_peer **peerp, lnet_nid_t nid)
 317 {
 318         struct kib_peer *peer;
 319         struct kib_net *net = ni->ni_data;
 320         int cpt = lnet_cpt_of_nid(nid);
 321         unsigned long flags;
 322
 323         LASSERT(net);
 324         LASSERT(nid != LNET_NID_ANY);
 325
 326         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 327         if (!peer) {
 328                 CERROR("Cannot allocate peer\n");
 329                 return -ENOMEM;
 330         }
 331
 332         peer->ibp_ni = ni;
 333         peer->ibp_nid = nid;
 334         peer->ibp_error = 0;
 335         peer->ibp_last_alive = 0;
 336         peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni);
 337         peer->ibp_queue_depth = ni->ni_peertxcredits;
 338         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 339
 340         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 341         INIT_LIST_HEAD(&peer->ibp_conns);
 342         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 343
 344         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 345
 346         /* always called with a ref on ni, which prevents ni being shutdown */
 347         LASSERT(!net->ibn_shutdown);
 348
 349         /* npeers only grows with the global lock held */
 350         atomic_inc(&net->ibn_npeers);
 351
 352         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 353
 354         *peerp = peer;
 355         return 0;
 356 }
 357
 358 void kiblnd_destroy_peer(struct kib_peer *peer)
 359 {
 360         struct kib_net *net = peer->ibp_ni->ni_data;
 361
 362         LASSERT(net);
 363         LASSERT(!atomic_read(&peer->ibp_refcount));
 364         LASSERT(!kiblnd_peer_active(peer));
 365         LASSERT(kiblnd_peer_idle(peer));
 366         LASSERT(list_empty(&peer->ibp_tx_queue));
 367
 368         LIBCFS_FREE(peer, sizeof(*peer));
 369
 370         /*
 371          * NB a peer's connections keep a reference on their peer until
 372          * they are destroyed, so we can be assured that _all_ state to do
 373          * with this peer has been cleaned up when its refcount drops to
 374          * zero.
 375          */
 376         atomic_dec(&net->ibn_npeers);
 377 }
 378
 379 struct kib_peer *kiblnd_find_peer_locked(lnet_nid_t nid)
 380 {
 381         /*
 382          * the caller is responsible for accounting the additional reference
 383          * that this creates
 384          */
 385         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 386         struct list_head *tmp;
 387         struct kib_peer *peer;
 388
 389         list_for_each(tmp, peer_list) {
 390                 peer = list_entry(tmp, struct kib_peer, ibp_list);
 391                 LASSERT(!kiblnd_peer_idle(peer));
 392
 393                 if (peer->ibp_nid != nid)
 394                         continue;
 395
 396                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 397                        peer, libcfs_nid2str(nid),
 398                        atomic_read(&peer->ibp_refcount),
 399                        peer->ibp_version);
 400                 return peer;
 401         }
 402         return NULL;
 403 }
 404
 405 void kiblnd_unlink_peer_locked(struct kib_peer *peer)
 406 {
 407         LASSERT(list_empty(&peer->ibp_conns));
 408
 409         LASSERT(kiblnd_peer_active(peer));
 410         list_del_init(&peer->ibp_list);
 411         /* lose peerlist's ref */
 412         kiblnd_peer_decref(peer);
 413 }
 414
 415 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 416                                 lnet_nid_t *nidp, int *count)
 417 {
 418         struct kib_peer *peer;
 419         struct list_head *ptmp;
 420         int i;
 421         unsigned long flags;
 422
 423         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 424
 425         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 426                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 427                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 428                         LASSERT(!kiblnd_peer_idle(peer));
 429
 430                         if (peer->ibp_ni != ni)
 431                                 continue;
 432
 433                         if (index-- > 0)
 434                                 continue;
 435
 436                         *nidp = peer->ibp_nid;
 437                         *count = atomic_read(&peer->ibp_refcount);
 438
 439                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 440                                                flags);
 441                         return 0;
 442                 }
 443         }
 444
 445         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 446         return -ENOENT;
 447 }
 448
 449 static void kiblnd_del_peer_locked(struct kib_peer *peer)
 450 {
 451         struct list_head *ctmp;
 452         struct list_head *cnxt;
 453         struct kib_conn *conn;
 454
 455         if (list_empty(&peer->ibp_conns)) {
 456                 kiblnd_unlink_peer_locked(peer);
 457         } else {
 458                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 459                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
 460
 461                         kiblnd_close_conn_locked(conn, 0);
 462                 }
 463                 /* NB closing peer's last conn unlinked it. */
 464         }
 465         /*
 466          * NB peer now unlinked; might even be freed if the peer table had the
 467          * last ref on it.
 468          */
 469 }
 470
 471 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 472 {
 473         LIST_HEAD(zombies);
 474         struct list_head *ptmp;
 475         struct list_head *pnxt;
 476         struct kib_peer *peer;
 477         int lo;
 478         int hi;
 479         int i;
 480         unsigned long flags;
 481         int rc = -ENOENT;
 482
 483         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 484
 485         if (nid != LNET_NID_ANY) {
 486                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 487                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 488         } else {
 489                 lo = 0;
 490                 hi = kiblnd_data.kib_peer_hash_size - 1;
 491         }
 492
 493         for (i = lo; i <= hi; i++) {
 494                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 495                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 496                         LASSERT(!kiblnd_peer_idle(peer));
 497
 498                         if (peer->ibp_ni != ni)
 499                                 continue;
 500
 501                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 502                                 continue;
 503
 504                         if (!list_empty(&peer->ibp_tx_queue)) {
 505                                 LASSERT(list_empty(&peer->ibp_conns));
 506
 507                                 list_splice_init(&peer->ibp_tx_queue,
 508                                                  &zombies);
 509                         }
 510
 511                         kiblnd_del_peer_locked(peer);
 512                         rc = 0;  /* matched something */
 513                 }
 514         }
 515
 516         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 517
 518         kiblnd_txlist_done(ni, &zombies, -EIO);
 519
 520         return rc;
 521 }
 522
 523 static struct kib_conn *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 524 {
 525         struct kib_peer *peer;
 526         struct list_head *ptmp;
 527         struct kib_conn *conn;
 528         struct list_head *ctmp;
 529         int i;
 530         unsigned long flags;
 531
 532         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 533
 534         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 535                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 536                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 537                         LASSERT(!kiblnd_peer_idle(peer));
 538
 539                         if (peer->ibp_ni != ni)
 540                                 continue;
 541
 542                         list_for_each(ctmp, &peer->ibp_conns) {
 543                                 if (index-- > 0)
 544                                         continue;
 545
 546                                 conn = list_entry(ctmp, struct kib_conn,
 547                                                   ibc_list);
 548                                 kiblnd_conn_addref(conn);
 549                                 read_unlock_irqrestore(
 550                                         &kiblnd_data.kib_global_lock,
 551                                         flags);
 552                                 return conn;
 553                         }
 554                 }
 555         }
 556
 557         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 558         return NULL;
 559 }
 560
 561 int kiblnd_translate_mtu(int value)
 562 {
 563         switch (value) {
 564         default:
 565                 return -1;
 566         case 0:
 567                 return 0;
 568         case 256:
 569                 return IB_MTU_256;
 570         case 512:
 571                 return IB_MTU_512;
 572         case 1024:
 573                 return IB_MTU_1024;
 574         case 2048:
 575                 return IB_MTU_2048;
 576         case 4096:
 577                 return IB_MTU_4096;
 578         }
 579 }
 580
 581 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 582 {
 583         int mtu;
 584
 585         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 586         if (!cmid->route.path_rec)
 587                 return;
 588
 589         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 590         LASSERT(mtu >= 0);
 591         if (mtu)
 592                 cmid->route.path_rec->mtu = mtu;
 593 }
 594
 595 static int kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 596 {
 597         cpumask_t *mask;
 598         int vectors;
 599         int off;
 600         int i;
 601         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 602
 603         vectors = conn->ibc_cmid->device->num_comp_vectors;
 604         if (vectors <= 1)
 605                 return 0;
 606
 607         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 608         if (!mask)
 609                 return 0;
 610
 611         /* hash NID to CPU id in this partition... */
 612         off = do_div(nid, cpumask_weight(mask));
 613         for_each_cpu(i, mask) {
 614                 if (!off--)
 615                         return i % vectors;
 616         }
 617
 618         LBUG();
 619         return 1;
 620 }
 621
 622 struct kib_conn *kiblnd_create_conn(struct kib_peer *peer, struct rdma_cm_id *cmid,
 623                                     int state, int version)
 624 {
 625         /*
 626          * CAVEAT EMPTOR:
 627          * If the new conn is created successfully it takes over the caller's
 628          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 629          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 630          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 631          * to destroy 'cmid' here since I'm called from the CM which still has
 632          * its ref on 'cmid').
 633          */
 634         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 635         struct kib_net *net = peer->ibp_ni->ni_data;
 636         struct kib_dev *dev;
 637         struct ib_qp_init_attr *init_qp_attr;
 638         struct kib_sched_info *sched;
 639         struct ib_cq_init_attr cq_attr = {};
 640         struct kib_conn *conn;
 641         struct ib_cq *cq;
 642         unsigned long flags;
 643         int cpt;
 644         int rc;
 645         int i;
 646
 647         LASSERT(net);
 648         LASSERT(!in_interrupt());
 649
 650         dev = net->ibn_dev;
 651
 652         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 653         sched = kiblnd_data.kib_scheds[cpt];
 654
 655         LASSERT(sched->ibs_nthreads > 0);
 656
 657         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 658                          sizeof(*init_qp_attr));
 659         if (!init_qp_attr) {
 660                 CERROR("Can't allocate qp_attr for %s\n",
 661                        libcfs_nid2str(peer->ibp_nid));
 662                 goto failed_0;
 663         }
 664
 665         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 666         if (!conn) {
 667                 CERROR("Can't allocate connection for %s\n",
 668                        libcfs_nid2str(peer->ibp_nid));
 669                 goto failed_1;
 670         }
 671
 672         conn->ibc_state = IBLND_CONN_INIT;
 673         conn->ibc_version = version;
 674         conn->ibc_peer = peer;            /* I take the caller's ref */
 675         cmid->context = conn;              /* for future CM callbacks */
 676         conn->ibc_cmid = cmid;
 677         conn->ibc_max_frags = peer->ibp_max_frags;
 678         conn->ibc_queue_depth = peer->ibp_queue_depth;
 679
 680         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 681         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 682         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 683         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 684         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 685         INIT_LIST_HEAD(&conn->ibc_active_txs);
 686         spin_lock_init(&conn->ibc_lock);
 687
 688         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 689                          sizeof(*conn->ibc_connvars));
 690         if (!conn->ibc_connvars) {
 691                 CERROR("Can't allocate in-progress connection state\n");
 692                 goto failed_2;
 693         }
 694
 695         write_lock_irqsave(glock, flags);
 696         if (dev->ibd_failover) {
 697                 write_unlock_irqrestore(glock, flags);
 698                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 699                 goto failed_2;
 700         }
 701
 702         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 703                 /* wakeup failover thread and teardown connection */
 704                 if (kiblnd_dev_can_failover(dev)) {
 705                         list_add_tail(&dev->ibd_fail_list,
 706                                       &kiblnd_data.kib_failed_devs);
 707                         wake_up(&kiblnd_data.kib_failover_waitq);
 708                 }
 709
 710                 write_unlock_irqrestore(glock, flags);
 711                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 712                        cmid->device->name, dev->ibd_ifname);
 713                 goto failed_2;
 714         }
 715
 716         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 717         conn->ibc_hdev = dev->ibd_hdev;
 718
 719         kiblnd_setup_mtu_locked(cmid);
 720
 721         write_unlock_irqrestore(glock, flags);
 722
 723         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 724                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 725         if (!conn->ibc_rxs) {
 726                 CERROR("Cannot allocate RX buffers\n");
 727                 goto failed_2;
 728         }
 729
 730         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 731                                 IBLND_RX_MSG_PAGES(conn));
 732         if (rc)
 733                 goto failed_2;
 734
 735         kiblnd_map_rx_descs(conn);
 736
 737         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 738         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 739         cq = ib_create_cq(cmid->device,
 740                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 741                           &cq_attr);
 742         if (IS_ERR(cq)) {
 743                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 744                        IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 745                 goto failed_2;
 746         }
 747
 748         conn->ibc_cq = cq;
 749
 750         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 751         if (rc) {
 752                 CERROR("Can't request completion notification: %d\n", rc);
 753                 goto failed_2;
 754         }
 755
 756         init_qp_attr->event_handler = kiblnd_qp_event;
 757         init_qp_attr->qp_context = conn;
 758         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 759         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 760         init_qp_attr->cap.max_send_sge = 1;
 761         init_qp_attr->cap.max_recv_sge = 1;
 762         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 763         init_qp_attr->qp_type = IB_QPT_RC;
 764         init_qp_attr->send_cq = cq;
 765         init_qp_attr->recv_cq = cq;
 766
 767         conn->ibc_sched = sched;
 768
 769         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 770         if (rc) {
 771                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 772                        rc, init_qp_attr->cap.max_send_wr,
 773                        init_qp_attr->cap.max_recv_wr);
 774                 goto failed_2;
 775         }
 776
 777         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 778
 779         /* 1 ref for caller and each rxmsg */
 780         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 781         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 782
 783         /* post receives */
 784         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 785                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 786                                     IBLND_POSTRX_NO_CREDIT);
 787                 if (rc) {
 788                         CERROR("Can't post rxmsg: %d\n", rc);
 789
 790                         /* Make posted receives complete */
 791                         kiblnd_abort_receives(conn);
 792
 793                         /*
 794                          * correct # of posted buffers
 795                          * NB locking needed now I'm racing with completion
 796                          */
 797                         spin_lock_irqsave(&sched->ibs_lock, flags);
 798                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 799                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 800
 801                         /*
 802                          * cmid will be destroyed by CM(ofed) after cm_callback
 803                          * returned, so we can't refer it anymore
 804                          * (by kiblnd_connd()->kiblnd_destroy_conn)
 805                          */
 806                         rdma_destroy_qp(conn->ibc_cmid);
 807                         conn->ibc_cmid = NULL;
 808
 809                         /* Drop my own and unused rxbuffer refcounts */
 810                         while (i++ <= IBLND_RX_MSGS(conn))
 811                                 kiblnd_conn_decref(conn);
 812
 813                         return NULL;
 814                 }
 815         }
 816
 817         /* Init successful! */
 818         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 819                 state == IBLND_CONN_PASSIVE_WAIT);
 820         conn->ibc_state = state;
 821
 822         /* 1 more conn */
 823         atomic_inc(&net->ibn_nconns);
 824         return conn;
 825
 826  failed_2:
 827         kiblnd_destroy_conn(conn);
 828         LIBCFS_FREE(conn, sizeof(*conn));
 829  failed_1:
 830         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 831  failed_0:
 832         return NULL;
 833 }
 834
 835 void kiblnd_destroy_conn(struct kib_conn *conn)
 836 {
 837         struct rdma_cm_id *cmid = conn->ibc_cmid;
 838         struct kib_peer *peer = conn->ibc_peer;
 839         int rc;
 840
 841         LASSERT(!in_interrupt());
 842         LASSERT(!atomic_read(&conn->ibc_refcount));
 843         LASSERT(list_empty(&conn->ibc_early_rxs));
 844         LASSERT(list_empty(&conn->ibc_tx_noops));
 845         LASSERT(list_empty(&conn->ibc_tx_queue));
 846         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 847         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 848         LASSERT(list_empty(&conn->ibc_active_txs));
 849         LASSERT(!conn->ibc_noops_posted);
 850         LASSERT(!conn->ibc_nsends_posted);
 851
 852         switch (conn->ibc_state) {
 853         default:
 854                 /* conn must be completely disengaged from the network */
 855                 LBUG();
 856
 857         case IBLND_CONN_DISCONNECTED:
 858                 /* connvars should have been freed already */
 859                 LASSERT(!conn->ibc_connvars);
 860                 break;
 861
 862         case IBLND_CONN_INIT:
 863                 break;
 864         }
 865
 866         /* conn->ibc_cmid might be destroyed by CM already */
 867         if (cmid && cmid->qp)
 868                 rdma_destroy_qp(cmid);
 869
 870         if (conn->ibc_cq) {
 871                 rc = ib_destroy_cq(conn->ibc_cq);
 872                 if (rc)
 873                         CWARN("Error destroying CQ: %d\n", rc);
 874         }
 875
 876         if (conn->ibc_rx_pages)
 877                 kiblnd_unmap_rx_descs(conn);
 878
 879         if (conn->ibc_rxs) {
 880                 LIBCFS_FREE(conn->ibc_rxs,
 881                             IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 882         }
 883
 884         if (conn->ibc_connvars)
 885                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 886
 887         if (conn->ibc_hdev)
 888                 kiblnd_hdev_decref(conn->ibc_hdev);
 889
 890         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 891         if (conn->ibc_state != IBLND_CONN_INIT) {
 892                 struct kib_net *net = peer->ibp_ni->ni_data;
 893
 894                 kiblnd_peer_decref(peer);
 895                 rdma_destroy_id(cmid);
 896                 atomic_dec(&net->ibn_nconns);
 897         }
 898 }
 899
 900 int kiblnd_close_peer_conns_locked(struct kib_peer *peer, int why)
 901 {
 902         struct kib_conn *conn;
 903         struct list_head *ctmp;
 904         struct list_head *cnxt;
 905         int count = 0;
 906
 907         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 908                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 909
 910                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 911                        libcfs_nid2str(peer->ibp_nid),
 912                        conn->ibc_version, why);
 913
 914                 kiblnd_close_conn_locked(conn, why);
 915                 count++;
 916         }
 917
 918         return count;
 919 }
 920
 921 int kiblnd_close_stale_conns_locked(struct kib_peer *peer,
 922                                     int version, __u64 incarnation)
 923 {
 924         struct kib_conn *conn;
 925         struct list_head *ctmp;
 926         struct list_head *cnxt;
 927         int count = 0;
 928
 929         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 930                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 931
 932                 if (conn->ibc_version     == version &&
 933                     conn->ibc_incarnation == incarnation)
 934                         continue;
 935
 936                 CDEBUG(D_NET,
 937                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 938                        libcfs_nid2str(peer->ibp_nid),
 939                        conn->ibc_version, conn->ibc_incarnation,
 940                        version, incarnation);
 941
 942                 kiblnd_close_conn_locked(conn, -ESTALE);
 943                 count++;
 944         }
 945
 946         return count;
 947 }
 948
 949 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 950 {
 951         struct kib_peer *peer;
 952         struct list_head *ptmp;
 953         struct list_head *pnxt;
 954         int lo;
 955         int hi;
 956         int i;
 957         unsigned long flags;
 958         int count = 0;
 959
 960         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 961
 962         if (nid != LNET_NID_ANY) {
 963                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 964                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 965         } else {
 966                 lo = 0;
 967                 hi = kiblnd_data.kib_peer_hash_size - 1;
 968         }
 969
 970         for (i = lo; i <= hi; i++) {
 971                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 972                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 973                         LASSERT(!kiblnd_peer_idle(peer));
 974
 975                         if (peer->ibp_ni != ni)
 976                                 continue;
 977
 978                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 979                                 continue;
 980
 981                         count += kiblnd_close_peer_conns_locked(peer, 0);
 982                 }
 983         }
 984
 985         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 986
 987         /* wildcards always succeed */
 988         if (nid == LNET_NID_ANY)
 989                 return 0;
 990
 991         return !count ? -ENOENT : 0;
 992 }
 993
 994 static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
 995 {
 996         struct libcfs_ioctl_data *data = arg;
 997         int rc = -EINVAL;
 998
 999         switch (cmd) {
1000         case IOC_LIBCFS_GET_PEER: {
1001                 lnet_nid_t nid = 0;
1002                 int count = 0;
1003
1004                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1005                                           &nid, &count);
1006                 data->ioc_nid   = nid;
1007                 data->ioc_count = count;
1008                 break;
1009         }
1010
1011         case IOC_LIBCFS_DEL_PEER: {
1012                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1013                 break;
1014         }
1015         case IOC_LIBCFS_GET_CONN: {
1016                 struct kib_conn *conn;
1017
1018                 rc = 0;
1019                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1020                 if (!conn) {
1021                         rc = -ENOENT;
1022                         break;
1023                 }
1024
1025                 LASSERT(conn->ibc_cmid);
1026                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1027                 if (!conn->ibc_cmid->route.path_rec)
1028                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1029                 else
1030                         data->ioc_u32[0] =
1031                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1032                 kiblnd_conn_decref(conn);
1033                 break;
1034         }
1035         case IOC_LIBCFS_CLOSE_CONNECTION: {
1036                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1037                 break;
1038         }
1039
1040         default:
1041                 break;
1042         }
1043
1044         return rc;
1045 }
1046
1047 static void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1048 {
1049         unsigned long last_alive = 0;
1050         unsigned long now = cfs_time_current();
1051         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1052         struct kib_peer *peer;
1053         unsigned long flags;
1054
1055         read_lock_irqsave(glock, flags);
1056
1057         peer = kiblnd_find_peer_locked(nid);
1058         if (peer)
1059                 last_alive = peer->ibp_last_alive;
1060
1061         read_unlock_irqrestore(glock, flags);
1062
1063         if (last_alive)
1064                 *when = last_alive;
1065
1066         /*
1067          * peer is not persistent in hash, trigger peer creation
1068          * and connection establishment with a NULL tx
1069          */
1070         if (!peer)
1071                 kiblnd_launch_tx(ni, NULL, nid);
1072
1073         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1074                libcfs_nid2str(nid), peer,
1075                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1076 }
1077
1078 static void kiblnd_free_pages(struct kib_pages *p)
1079 {
1080         int npages = p->ibp_npages;
1081         int i;
1082
1083         for (i = 0; i < npages; i++) {
1084                 if (p->ibp_pages[i])
1085                         __free_page(p->ibp_pages[i]);
1086         }
1087
1088         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1089 }
1090
1091 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1092 {
1093         struct kib_pages *p;
1094         int i;
1095
1096         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1097                          offsetof(struct kib_pages, ibp_pages[npages]));
1098         if (!p) {
1099                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1100                 return -ENOMEM;
1101         }
1102
1103         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1104         p->ibp_npages = npages;
1105
1106         for (i = 0; i < npages; i++) {
1107                 p->ibp_pages[i] = alloc_pages_node(
1108                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1109                                     GFP_NOFS, 0);
1110                 if (!p->ibp_pages[i]) {
1111                         CERROR("Can't allocate page %d of %d\n", i, npages);
1112                         kiblnd_free_pages(p);
1113                         return -ENOMEM;
1114                 }
1115         }
1116
1117         *pp = p;
1118         return 0;
1119 }
1120
1121 void kiblnd_unmap_rx_descs(struct kib_conn *conn)
1122 {
1123         struct kib_rx *rx;
1124         int i;
1125
1126         LASSERT(conn->ibc_rxs);
1127         LASSERT(conn->ibc_hdev);
1128
1129         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1130                 rx = &conn->ibc_rxs[i];
1131
1132                 LASSERT(rx->rx_nob >= 0); /* not posted */
1133
1134                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1135                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1136                                                           rx->rx_msgaddr),
1137                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1138         }
1139
1140         kiblnd_free_pages(conn->ibc_rx_pages);
1141
1142         conn->ibc_rx_pages = NULL;
1143 }
1144
1145 void kiblnd_map_rx_descs(struct kib_conn *conn)
1146 {
1147         struct kib_rx *rx;
1148         struct page *pg;
1149         int pg_off;
1150         int ipg;
1151         int i;
1152
1153         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1154                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1155                 rx = &conn->ibc_rxs[i];
1156
1157                 rx->rx_conn = conn;
1158                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1159
1160                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1161                                                        rx->rx_msg,
1162                                                        IBLND_MSG_SIZE,
1163                                                        DMA_FROM_DEVICE);
1164                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1165                                                   rx->rx_msgaddr));
1166                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1167
1168                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1169                        i, rx->rx_msg, rx->rx_msgaddr,
1170                        (__u64)(page_to_phys(pg) + pg_off));
1171
1172                 pg_off += IBLND_MSG_SIZE;
1173                 LASSERT(pg_off <= PAGE_SIZE);
1174
1175                 if (pg_off == PAGE_SIZE) {
1176                         pg_off = 0;
1177                         ipg++;
1178                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1179                 }
1180         }
1181 }
1182
1183 static void kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1184 {
1185         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1186         struct kib_tx *tx;
1187         int i;
1188
1189         LASSERT(!tpo->tpo_pool.po_allocated);
1190
1191         if (!hdev)
1192                 return;
1193
1194         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1195                 tx = &tpo->tpo_tx_descs[i];
1196                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1197                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1198                                                           tx->tx_msgaddr),
1199                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1200         }
1201
1202         kiblnd_hdev_decref(hdev);
1203         tpo->tpo_hdev = NULL;
1204 }
1205
1206 static struct kib_hca_dev *kiblnd_current_hdev(struct kib_dev *dev)
1207 {
1208         struct kib_hca_dev *hdev;
1209         unsigned long flags;
1210         int i = 0;
1211
1212         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1213         while (dev->ibd_failover) {
1214                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1215                 if (!(i++ % 50))
1216                         CDEBUG(D_NET, "%s: Wait for failover\n",
1217                                dev->ibd_ifname);
1218                 set_current_state(TASK_INTERRUPTIBLE);
1219                 schedule_timeout(cfs_time_seconds(1) / 100);
1220
1221                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1222         }
1223
1224         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1225         hdev = dev->ibd_hdev;
1226
1227         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1228
1229         return hdev;
1230 }
1231
1232 static void kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1233 {
1234         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1235         struct kib_pool *pool = &tpo->tpo_pool;
1236         struct kib_net *net = pool->po_owner->ps_net;
1237         struct kib_dev *dev;
1238         struct page *page;
1239         struct kib_tx *tx;
1240         int page_offset;
1241         int ipage;
1242         int i;
1243
1244         LASSERT(net);
1245
1246         dev = net->ibn_dev;
1247
1248         /* pre-mapped messages are not bigger than 1 page */
1249         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1250
1251         /* No fancy arithmetic when we do the buffer calculations */
1252         CLASSERT(!(PAGE_SIZE % IBLND_MSG_SIZE));
1253
1254         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1255
1256         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1257                 page = txpgs->ibp_pages[ipage];
1258                 tx = &tpo->tpo_tx_descs[i];
1259
1260                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1261                                            page_offset);
1262
1263                 tx->tx_msgaddr = kiblnd_dma_map_single(
1264                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1265                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1266                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1267                                                   tx->tx_msgaddr));
1268                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1269
1270                 list_add(&tx->tx_list, &pool->po_free_list);
1271
1272                 page_offset += IBLND_MSG_SIZE;
1273                 LASSERT(page_offset <= PAGE_SIZE);
1274
1275                 if (page_offset == PAGE_SIZE) {
1276                         page_offset = 0;
1277                         ipage++;
1278                         LASSERT(ipage <= txpgs->ibp_npages);
1279                 }
1280         }
1281 }
1282
1283 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd,
1284                                     int negotiated_nfrags)
1285 {
1286         struct kib_net *net = ni->ni_data;
1287         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
1288         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1289         __u16 nfrags;
1290         int mod;
1291
1292         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1293         mod = tunables->lnd_map_on_demand;
1294         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1295
1296         LASSERT(hdev->ibh_mrs);
1297
1298         if (mod > 0 && nfrags <= rd->rd_nfrags)
1299                 return NULL;
1300
1301         return hdev->ibh_mrs;
1302 }
1303
1304 static void kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1305 {
1306         LASSERT(!fpo->fpo_map_count);
1307
1308         if (fpo->fpo_is_fmr) {
1309                 if (fpo->fmr.fpo_fmr_pool)
1310                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1311         } else {
1312                 struct kib_fast_reg_descriptor *frd, *tmp;
1313                 int i = 0;
1314
1315                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1316                                          frd_list) {
1317                         list_del(&frd->frd_list);
1318                         ib_dereg_mr(frd->frd_mr);
1319                         LIBCFS_FREE(frd, sizeof(*frd));
1320                         i++;
1321                 }
1322                 if (i < fpo->fast_reg.fpo_pool_size)
1323                         CERROR("FastReg pool still has %d regions registered\n",
1324                                fpo->fast_reg.fpo_pool_size - i);
1325         }
1326
1327         if (fpo->fpo_hdev)
1328                 kiblnd_hdev_decref(fpo->fpo_hdev);
1329
1330         LIBCFS_FREE(fpo, sizeof(*fpo));
1331 }
1332
1333 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1334 {
1335         struct kib_fmr_pool *fpo, *tmp;
1336
1337         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1338                 list_del(&fpo->fpo_list);
1339                 kiblnd_destroy_fmr_pool(fpo);
1340         }
1341 }
1342
1343 static int
1344 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1345                      int ncpts)
1346 {
1347         int size = tunables->lnd_fmr_pool_size / ncpts;
1348
1349         return max(IBLND_FMR_POOL, size);
1350 }
1351
1352 static int
1353 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1354                          int ncpts)
1355 {
1356         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1357
1358         return max(IBLND_FMR_POOL_FLUSH, size);
1359 }
1360
1361 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1362 {
1363         struct ib_fmr_pool_param param = {
1364                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1365                 .page_shift        = PAGE_SHIFT,
1366                 .access            = (IB_ACCESS_LOCAL_WRITE |
1367                                       IB_ACCESS_REMOTE_WRITE),
1368                 .pool_size         = fps->fps_pool_size,
1369                 .dirty_watermark   = fps->fps_flush_trigger,
1370                 .flush_function    = NULL,
1371                 .flush_arg         = NULL,
1372                 .cache             = !!fps->fps_cache };
1373         int rc = 0;
1374
1375         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1376                                                    &param);
1377         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1378                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1379                 if (rc != -ENOSYS)
1380                         CERROR("Failed to create FMR pool: %d\n", rc);
1381                 else
1382                         CERROR("FMRs are not supported\n");
1383         }
1384
1385         return rc;
1386 }
1387
1388 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1389 {
1390         struct kib_fast_reg_descriptor *frd, *tmp;
1391         int i, rc;
1392
1393         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1394         fpo->fast_reg.fpo_pool_size = 0;
1395         for (i = 0; i < fps->fps_pool_size; i++) {
1396                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1397                                  sizeof(*frd));
1398                 if (!frd) {
1399                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1400                         rc = -ENOMEM;
1401                         goto out;
1402                 }
1403
1404                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1405                                           IB_MR_TYPE_MEM_REG,
1406                                           LNET_MAX_PAYLOAD / PAGE_SIZE);
1407                 if (IS_ERR(frd->frd_mr)) {
1408                         rc = PTR_ERR(frd->frd_mr);
1409                         CERROR("Failed to allocate ib_alloc_mr: %d\n", rc);
1410                         frd->frd_mr = NULL;
1411                         goto out_middle;
1412                 }
1413
1414                 frd->frd_valid = true;
1415
1416                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1417                 fpo->fast_reg.fpo_pool_size++;
1418         }
1419
1420         return 0;
1421
1422 out_middle:
1423         if (frd->frd_mr)
1424                 ib_dereg_mr(frd->frd_mr);
1425         LIBCFS_FREE(frd, sizeof(*frd));
1426
1427 out:
1428         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1429                                  frd_list) {
1430                 list_del(&frd->frd_list);
1431                 ib_dereg_mr(frd->frd_mr);
1432                 LIBCFS_FREE(frd, sizeof(*frd));
1433         }
1434
1435         return rc;
1436 }
1437
1438 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1439                                   struct kib_fmr_pool **pp_fpo)
1440 {
1441         struct kib_dev *dev = fps->fps_net->ibn_dev;
1442         struct ib_device_attr *dev_attr;
1443         struct kib_fmr_pool *fpo;
1444         int rc;
1445
1446         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1447         if (!fpo)
1448                 return -ENOMEM;
1449
1450         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1451         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1452
1453         /* Check for FMR or FastReg support */
1454         fpo->fpo_is_fmr = 0;
1455         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1456             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1457             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1458             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1459                 LCONSOLE_INFO("Using FMR for registration\n");
1460                 fpo->fpo_is_fmr = 1;
1461         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1462                 LCONSOLE_INFO("Using FastReg for registration\n");
1463         } else {
1464                 rc = -ENOSYS;
1465                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1466                 goto out_fpo;
1467         }
1468
1469         if (fpo->fpo_is_fmr)
1470                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1471         else
1472                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1473         if (rc)
1474                 goto out_fpo;
1475
1476         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1477         fpo->fpo_owner = fps;
1478         *pp_fpo = fpo;
1479
1480         return 0;
1481
1482 out_fpo:
1483         kiblnd_hdev_decref(fpo->fpo_hdev);
1484         LIBCFS_FREE(fpo, sizeof(*fpo));
1485         return rc;
1486 }
1487
1488 static void kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps,
1489                                     struct list_head *zombies)
1490 {
1491         if (!fps->fps_net) /* intialized? */
1492                 return;
1493
1494         spin_lock(&fps->fps_lock);
1495
1496         while (!list_empty(&fps->fps_pool_list)) {
1497                 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1498                                                  struct kib_fmr_pool, fpo_list);
1499                 fpo->fpo_failed = 1;
1500                 list_del(&fpo->fpo_list);
1501                 if (!fpo->fpo_map_count)
1502                         list_add(&fpo->fpo_list, zombies);
1503                 else
1504                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1505         }
1506
1507         spin_unlock(&fps->fps_lock);
1508 }
1509
1510 static void kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1511 {
1512         if (fps->fps_net) { /* initialized? */
1513                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1514                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1515         }
1516 }
1517
1518 static int
1519 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1520                         struct kib_net *net,
1521                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1522 {
1523         struct kib_fmr_pool *fpo;
1524         int rc;
1525
1526         memset(fps, 0, sizeof(*fps));
1527
1528         fps->fps_net = net;
1529         fps->fps_cpt = cpt;
1530
1531         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1532         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1533         fps->fps_cache = tunables->lnd_fmr_cache;
1534
1535         spin_lock_init(&fps->fps_lock);
1536         INIT_LIST_HEAD(&fps->fps_pool_list);
1537         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1538
1539         rc = kiblnd_create_fmr_pool(fps, &fpo);
1540         if (!rc)
1541                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1542
1543         return rc;
1544 }
1545
1546 static int kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, unsigned long now)
1547 {
1548         if (fpo->fpo_map_count) /* still in use */
1549                 return 0;
1550         if (fpo->fpo_failed)
1551                 return 1;
1552         return cfs_time_aftereq(now, fpo->fpo_deadline);
1553 }
1554
1555 static int
1556 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1557 {
1558         __u64 *pages = tx->tx_pages;
1559         struct kib_hca_dev *hdev;
1560         int npages;
1561         int size;
1562         int i;
1563
1564         hdev = tx->tx_pool->tpo_hdev;
1565
1566         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1567                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1568                      size += hdev->ibh_page_size) {
1569                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1570                                            hdev->ibh_page_mask) + size;
1571                 }
1572         }
1573
1574         return npages;
1575 }
1576
1577 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1578 {
1579         LIST_HEAD(zombies);
1580         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1581         struct kib_fmr_poolset *fps;
1582         unsigned long now = cfs_time_current();
1583         struct kib_fmr_pool *tmp;
1584         int rc;
1585
1586         if (!fpo)
1587                 return;
1588
1589         fps = fpo->fpo_owner;
1590         if (fpo->fpo_is_fmr) {
1591                 if (fmr->fmr_pfmr) {
1592                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1593                         LASSERT(!rc);
1594                         fmr->fmr_pfmr = NULL;
1595                 }
1596
1597                 if (status) {
1598                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1599                         LASSERT(!rc);
1600                 }
1601         } else {
1602                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1603
1604                 if (frd) {
1605                         frd->frd_valid = false;
1606                         spin_lock(&fps->fps_lock);
1607                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1608                         spin_unlock(&fps->fps_lock);
1609                         fmr->fmr_frd = NULL;
1610                 }
1611         }
1612         fmr->fmr_pool = NULL;
1613
1614         spin_lock(&fps->fps_lock);
1615         fpo->fpo_map_count--;  /* decref the pool */
1616
1617         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1618                 /* the first pool is persistent */
1619                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1620                         continue;
1621
1622                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1623                         list_move(&fpo->fpo_list, &zombies);
1624                         fps->fps_version++;
1625                 }
1626         }
1627         spin_unlock(&fps->fps_lock);
1628
1629         if (!list_empty(&zombies))
1630                 kiblnd_destroy_fmr_pool_list(&zombies);
1631 }
1632
1633 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1634                         struct kib_rdma_desc *rd, __u32 nob, __u64 iov,
1635                         struct kib_fmr *fmr)
1636 {
1637         __u64 *pages = tx->tx_pages;
1638         bool is_rx = (rd != tx->tx_rd);
1639         bool tx_pages_mapped = 0;
1640         struct kib_fmr_pool *fpo;
1641         int npages = 0;
1642         __u64 version;
1643         int rc;
1644
1645  again:
1646         spin_lock(&fps->fps_lock);
1647         version = fps->fps_version;
1648         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1649                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1650                 fpo->fpo_map_count++;
1651
1652                 if (fpo->fpo_is_fmr) {
1653                         struct ib_pool_fmr *pfmr;
1654
1655                         spin_unlock(&fps->fps_lock);
1656
1657                         if (!tx_pages_mapped) {
1658                                 npages = kiblnd_map_tx_pages(tx, rd);
1659                                 tx_pages_mapped = 1;
1660                         }
1661
1662                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1663                                                     pages, npages, iov);
1664                         if (likely(!IS_ERR(pfmr))) {
1665                                 fmr->fmr_key = is_rx ? pfmr->fmr->rkey :
1666                                                        pfmr->fmr->lkey;
1667                                 fmr->fmr_frd = NULL;
1668                                 fmr->fmr_pfmr = pfmr;
1669                                 fmr->fmr_pool = fpo;
1670                                 return 0;
1671                         }
1672                         rc = PTR_ERR(pfmr);
1673                 } else {
1674                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1675                                 struct kib_fast_reg_descriptor *frd;
1676                                 struct ib_reg_wr *wr;
1677                                 struct ib_mr *mr;
1678                                 int n;
1679
1680                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1681                                                        struct kib_fast_reg_descriptor,
1682                                                        frd_list);
1683                                 list_del(&frd->frd_list);
1684                                 spin_unlock(&fps->fps_lock);
1685
1686                                 mr = frd->frd_mr;
1687
1688                                 if (!frd->frd_valid) {
1689                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1690                                         struct ib_send_wr *inv_wr;
1691
1692                                         inv_wr = &frd->frd_inv_wr;
1693                                         memset(inv_wr, 0, sizeof(*inv_wr));
1694                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1695                                         inv_wr->wr_id = IBLND_WID_MR;
1696                                         inv_wr->ex.invalidate_rkey = key;
1697
1698                                         /* Bump the key */
1699                                         key = ib_inc_rkey(key);
1700                                         ib_update_fast_reg_key(mr, key);
1701                                 }
1702
1703                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1704                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1705                                 if (unlikely(n != tx->tx_nfrags)) {
1706                                         CERROR("Failed to map mr %d/%d elements\n",
1707                                                n, tx->tx_nfrags);
1708                                         return n < 0 ? n : -EINVAL;
1709                                 }
1710
1711                                 mr->iova = iov;
1712
1713                                 /* Prepare FastReg WR */
1714                                 wr = &frd->frd_fastreg_wr;
1715                                 memset(wr, 0, sizeof(*wr));
1716                                 wr->wr.opcode = IB_WR_REG_MR;
1717                                 wr->wr.wr_id = IBLND_WID_MR;
1718                                 wr->wr.num_sge = 0;
1719                                 wr->wr.send_flags = 0;
1720                                 wr->mr = mr;
1721                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1722                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1723                                               IB_ACCESS_REMOTE_WRITE);
1724
1725                                 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1726                                 fmr->fmr_frd = frd;
1727                                 fmr->fmr_pfmr = NULL;
1728                                 fmr->fmr_pool = fpo;
1729                                 return 0;
1730                         }
1731                         spin_unlock(&fps->fps_lock);
1732                         rc = -EBUSY;
1733                 }
1734
1735                 spin_lock(&fps->fps_lock);
1736                 fpo->fpo_map_count--;
1737                 if (rc != -EAGAIN) {
1738                         spin_unlock(&fps->fps_lock);
1739                         return rc;
1740                 }
1741
1742                 /* EAGAIN and ... */
1743                 if (version != fps->fps_version) {
1744                         spin_unlock(&fps->fps_lock);
1745                         goto again;
1746                 }
1747         }
1748
1749         if (fps->fps_increasing) {
1750                 spin_unlock(&fps->fps_lock);
1751                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1752                 schedule();
1753                 goto again;
1754         }
1755
1756         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1757                 /* someone failed recently */
1758                 spin_unlock(&fps->fps_lock);
1759                 return -EAGAIN;
1760         }
1761
1762         fps->fps_increasing = 1;
1763         spin_unlock(&fps->fps_lock);
1764
1765         CDEBUG(D_NET, "Allocate new FMR pool\n");
1766         rc = kiblnd_create_fmr_pool(fps, &fpo);
1767         spin_lock(&fps->fps_lock);
1768         fps->fps_increasing = 0;
1769         if (!rc) {
1770                 fps->fps_version++;
1771                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1772         } else {
1773                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1774         }
1775         spin_unlock(&fps->fps_lock);
1776
1777         goto again;
1778 }
1779
1780 static void kiblnd_fini_pool(struct kib_pool *pool)
1781 {
1782         LASSERT(list_empty(&pool->po_free_list));
1783         LASSERT(!pool->po_allocated);
1784
1785         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1786 }
1787
1788 static void kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1789 {
1790         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1791
1792         memset(pool, 0, sizeof(*pool));
1793         INIT_LIST_HEAD(&pool->po_free_list);
1794         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1795         pool->po_owner    = ps;
1796         pool->po_size     = size;
1797 }
1798
1799 static void kiblnd_destroy_pool_list(struct list_head *head)
1800 {
1801         struct kib_pool *pool;
1802
1803         while (!list_empty(head)) {
1804                 pool = list_entry(head->next, struct kib_pool, po_list);
1805                 list_del(&pool->po_list);
1806
1807                 LASSERT(pool->po_owner);
1808                 pool->po_owner->ps_pool_destroy(pool);
1809         }
1810 }
1811
1812 static void kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
1813 {
1814         if (!ps->ps_net) /* intialized? */
1815                 return;
1816
1817         spin_lock(&ps->ps_lock);
1818         while (!list_empty(&ps->ps_pool_list)) {
1819                 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
1820                                             struct kib_pool, po_list);
1821                 po->po_failed = 1;
1822                 list_del(&po->po_list);
1823                 if (!po->po_allocated)
1824                         list_add(&po->po_list, zombies);
1825                 else
1826                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1827         }
1828         spin_unlock(&ps->ps_lock);
1829 }
1830
1831 static void kiblnd_fini_poolset(struct kib_poolset *ps)
1832 {
1833         if (ps->ps_net) { /* initialized? */
1834                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1835                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1836         }
1837 }
1838
1839 static int kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
1840                                struct kib_net *net, char *name, int size,
1841                                kib_ps_pool_create_t po_create,
1842                                kib_ps_pool_destroy_t po_destroy,
1843                                kib_ps_node_init_t nd_init,
1844                                kib_ps_node_fini_t nd_fini)
1845 {
1846         struct kib_pool *pool;
1847         int rc;
1848
1849         memset(ps, 0, sizeof(*ps));
1850
1851         ps->ps_cpt          = cpt;
1852         ps->ps_net          = net;
1853         ps->ps_pool_create  = po_create;
1854         ps->ps_pool_destroy = po_destroy;
1855         ps->ps_node_init    = nd_init;
1856         ps->ps_node_fini    = nd_fini;
1857         ps->ps_pool_size    = size;
1858         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1859             >= sizeof(ps->ps_name))
1860                 return -E2BIG;
1861         spin_lock_init(&ps->ps_lock);
1862         INIT_LIST_HEAD(&ps->ps_pool_list);
1863         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1864
1865         rc = ps->ps_pool_create(ps, size, &pool);
1866         if (!rc)
1867                 list_add(&pool->po_list, &ps->ps_pool_list);
1868         else
1869                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1870
1871         return rc;
1872 }
1873
1874 static int kiblnd_pool_is_idle(struct kib_pool *pool, unsigned long now)
1875 {
1876         if (pool->po_allocated) /* still in use */
1877                 return 0;
1878         if (pool->po_failed)
1879                 return 1;
1880         return cfs_time_aftereq(now, pool->po_deadline);
1881 }
1882
1883 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
1884 {
1885         LIST_HEAD(zombies);
1886         struct kib_poolset *ps = pool->po_owner;
1887         struct kib_pool *tmp;
1888         unsigned long now = cfs_time_current();
1889
1890         spin_lock(&ps->ps_lock);
1891
1892         if (ps->ps_node_fini)
1893                 ps->ps_node_fini(pool, node);
1894
1895         LASSERT(pool->po_allocated > 0);
1896         list_add(node, &pool->po_free_list);
1897         pool->po_allocated--;
1898
1899         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1900                 /* the first pool is persistent */
1901                 if (ps->ps_pool_list.next == &pool->po_list)
1902                         continue;
1903
1904                 if (kiblnd_pool_is_idle(pool, now))
1905                         list_move(&pool->po_list, &zombies);
1906         }
1907         spin_unlock(&ps->ps_lock);
1908
1909         if (!list_empty(&zombies))
1910                 kiblnd_destroy_pool_list(&zombies);
1911 }
1912
1913 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps)
1914 {
1915         struct list_head *node;
1916         struct kib_pool *pool;
1917         unsigned int interval = 1;
1918         unsigned long time_before;
1919         unsigned int trips = 0;
1920         int rc;
1921
1922  again:
1923         spin_lock(&ps->ps_lock);
1924         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1925                 if (list_empty(&pool->po_free_list))
1926                         continue;
1927
1928                 pool->po_allocated++;
1929                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1930                 node = pool->po_free_list.next;
1931                 list_del(node);
1932
1933                 if (ps->ps_node_init) {
1934                         /* still hold the lock */
1935                         ps->ps_node_init(pool, node);
1936                 }
1937                 spin_unlock(&ps->ps_lock);
1938                 return node;
1939         }
1940
1941         /* no available tx pool and ... */
1942         if (ps->ps_increasing) {
1943                 /* another thread is allocating a new pool */
1944                 spin_unlock(&ps->ps_lock);
1945                 trips++;
1946                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting %d HZs for her to complete. trips = %d\n",
1947                        ps->ps_name, interval, trips);
1948
1949                 set_current_state(TASK_INTERRUPTIBLE);
1950                 schedule_timeout(interval);
1951                 if (interval < cfs_time_seconds(1))
1952                         interval *= 2;
1953
1954                 goto again;
1955         }
1956
1957         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1958                 /* someone failed recently */
1959                 spin_unlock(&ps->ps_lock);
1960                 return NULL;
1961         }
1962
1963         ps->ps_increasing = 1;
1964         spin_unlock(&ps->ps_lock);
1965
1966         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1967         time_before = cfs_time_current();
1968         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1969         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
1970                cfs_time_current() - time_before);
1971
1972         spin_lock(&ps->ps_lock);
1973         ps->ps_increasing = 0;
1974         if (!rc) {
1975                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1976         } else {
1977                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1978                 CERROR("Can't allocate new %s pool because out of memory\n",
1979                        ps->ps_name);
1980         }
1981         spin_unlock(&ps->ps_lock);
1982
1983         goto again;
1984 }
1985
1986 static void kiblnd_destroy_tx_pool(struct kib_pool *pool)
1987 {
1988         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool, tpo_pool);
1989         int i;
1990
1991         LASSERT(!pool->po_allocated);
1992
1993         if (tpo->tpo_tx_pages) {
1994                 kiblnd_unmap_tx_pool(tpo);
1995                 kiblnd_free_pages(tpo->tpo_tx_pages);
1996         }
1997
1998         if (!tpo->tpo_tx_descs)
1999                 goto out;
2000
2001         for (i = 0; i < pool->po_size; i++) {
2002                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2003
2004                 list_del(&tx->tx_list);
2005                 if (tx->tx_pages)
2006                         LIBCFS_FREE(tx->tx_pages,
2007                                     LNET_MAX_IOV *
2008                                     sizeof(*tx->tx_pages));
2009                 if (tx->tx_frags)
2010                         LIBCFS_FREE(tx->tx_frags,
2011                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2012                                      sizeof(*tx->tx_frags));
2013                 if (tx->tx_wrq)
2014                         LIBCFS_FREE(tx->tx_wrq,
2015                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2016                                     sizeof(*tx->tx_wrq));
2017                 if (tx->tx_sge)
2018                         LIBCFS_FREE(tx->tx_sge,
2019                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2020                                     sizeof(*tx->tx_sge));
2021                 if (tx->tx_rd)
2022                         LIBCFS_FREE(tx->tx_rd,
2023                                     offsetof(struct kib_rdma_desc,
2024                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2025         }
2026
2027         LIBCFS_FREE(tpo->tpo_tx_descs,
2028                     pool->po_size * sizeof(struct kib_tx));
2029 out:
2030         kiblnd_fini_pool(pool);
2031         LIBCFS_FREE(tpo, sizeof(*tpo));
2032 }
2033
2034 static int kiblnd_tx_pool_size(int ncpts)
2035 {
2036         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2037
2038         return max(IBLND_TX_POOL, ntx);
2039 }
2040
2041 static int kiblnd_create_tx_pool(struct kib_poolset *ps, int size,
2042                                  struct kib_pool **pp_po)
2043 {
2044         int i;
2045         int npg;
2046         struct kib_pool *pool;
2047         struct kib_tx_pool *tpo;
2048
2049         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2050         if (!tpo) {
2051                 CERROR("Failed to allocate TX pool\n");
2052                 return -ENOMEM;
2053         }
2054
2055         pool = &tpo->tpo_pool;
2056         kiblnd_init_pool(ps, pool, size);
2057         tpo->tpo_tx_descs = NULL;
2058         tpo->tpo_tx_pages = NULL;
2059
2060         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2061         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
2062                 CERROR("Can't allocate tx pages: %d\n", npg);
2063                 LIBCFS_FREE(tpo, sizeof(*tpo));
2064                 return -ENOMEM;
2065         }
2066
2067         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2068                          size * sizeof(struct kib_tx));
2069         if (!tpo->tpo_tx_descs) {
2070                 CERROR("Can't allocate %d tx descriptors\n", size);
2071                 ps->ps_pool_destroy(pool);
2072                 return -ENOMEM;
2073         }
2074
2075         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2076
2077         for (i = 0; i < size; i++) {
2078                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2079
2080                 tx->tx_pool = tpo;
2081                 if (ps->ps_net->ibn_fmr_ps) {
2082                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2083                                          lnet_cpt_table(), ps->ps_cpt,
2084                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2085                         if (!tx->tx_pages)
2086                                 break;
2087                 }
2088
2089                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2090                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2091                                  sizeof(*tx->tx_frags));
2092                 if (!tx->tx_frags)
2093                         break;
2094
2095                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2096
2097                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2098                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2099                                  sizeof(*tx->tx_wrq));
2100                 if (!tx->tx_wrq)
2101                         break;
2102
2103                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2104                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2105                                  sizeof(*tx->tx_sge));
2106                 if (!tx->tx_sge)
2107                         break;
2108
2109                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2110                                  offsetof(struct kib_rdma_desc,
2111                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2112                 if (!tx->tx_rd)
2113                         break;
2114         }
2115
2116         if (i == size) {
2117                 kiblnd_map_tx_pool(tpo);
2118                 *pp_po = pool;
2119                 return 0;
2120         }
2121
2122         ps->ps_pool_destroy(pool);
2123         return -ENOMEM;
2124 }
2125
2126 static void kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2127 {
2128         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2129                                                   struct kib_tx_poolset,
2130                                                   tps_poolset);
2131         struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2132
2133         tx->tx_cookie = tps->tps_next_tx_cookie++;
2134 }
2135
2136 static void kiblnd_net_fini_pools(struct kib_net *net)
2137 {
2138         int i;
2139
2140         cfs_cpt_for_each(i, lnet_cpt_table()) {
2141                 struct kib_tx_poolset *tps;
2142                 struct kib_fmr_poolset *fps;
2143
2144                 if (net->ibn_tx_ps) {
2145                         tps = net->ibn_tx_ps[i];
2146                         kiblnd_fini_poolset(&tps->tps_poolset);
2147                 }
2148
2149                 if (net->ibn_fmr_ps) {
2150                         fps = net->ibn_fmr_ps[i];
2151                         kiblnd_fini_fmr_poolset(fps);
2152                 }
2153         }
2154
2155         if (net->ibn_tx_ps) {
2156                 cfs_percpt_free(net->ibn_tx_ps);
2157                 net->ibn_tx_ps = NULL;
2158         }
2159
2160         if (net->ibn_fmr_ps) {
2161                 cfs_percpt_free(net->ibn_fmr_ps);
2162                 net->ibn_fmr_ps = NULL;
2163         }
2164 }
2165
2166 static int kiblnd_net_init_pools(struct kib_net *net, lnet_ni_t *ni, __u32 *cpts,
2167                                  int ncpts)
2168 {
2169         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2170         unsigned long flags;
2171         int cpt;
2172         int rc;
2173         int i;
2174
2175         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2176
2177         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2178         if (!tunables->lnd_map_on_demand) {
2179                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2180                 goto create_tx_pool;
2181         }
2182
2183         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2184
2185         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2186                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2187                        tunables->lnd_fmr_pool_size,
2188                        *kiblnd_tunables.kib_ntx / 4);
2189                 rc = -EINVAL;
2190                 goto failed;
2191         }
2192
2193         /*
2194          * TX pool must be created later than FMR, see LU-2268
2195          * for details
2196          */
2197         LASSERT(!net->ibn_tx_ps);
2198
2199         /*
2200          * premapping can fail if ibd_nmr > 1, so we always create
2201          * FMR pool and map-on-demand if premapping failed
2202          *
2203          * cfs_precpt_alloc is creating an array of struct kib_fmr_poolset
2204          * The number of struct kib_fmr_poolsets create is equal to the
2205          * number of CPTs that exist, i.e net->ibn_fmr_ps[cpt].
2206          */
2207         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2208                                            sizeof(struct kib_fmr_poolset));
2209         if (!net->ibn_fmr_ps) {
2210                 CERROR("Failed to allocate FMR pool array\n");
2211                 rc = -ENOMEM;
2212                 goto failed;
2213         }
2214
2215         for (i = 0; i < ncpts; i++) {
2216                 cpt = !cpts ? i : cpts[i];
2217                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2218                                              net, tunables);
2219                 if (rc) {
2220                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2221                                cpt, rc);
2222                         goto failed;
2223                 }
2224         }
2225
2226         if (i > 0)
2227                 LASSERT(i == ncpts);
2228
2229  create_tx_pool:
2230         /*
2231          * cfs_precpt_alloc is creating an array of struct kib_tx_poolset
2232          * The number of struct kib_tx_poolsets create is equal to the
2233          * number of CPTs that exist, i.e net->ibn_tx_ps[cpt].
2234          */
2235         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2236                                           sizeof(struct kib_tx_poolset));
2237         if (!net->ibn_tx_ps) {
2238                 CERROR("Failed to allocate tx pool array\n");
2239                 rc = -ENOMEM;
2240                 goto failed;
2241         }
2242
2243         for (i = 0; i < ncpts; i++) {
2244                 cpt = !cpts ? i : cpts[i];
2245                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2246                                          cpt, net, "TX",
2247                                          kiblnd_tx_pool_size(ncpts),
2248                                          kiblnd_create_tx_pool,
2249                                          kiblnd_destroy_tx_pool,
2250                                          kiblnd_tx_init, NULL);
2251                 if (rc) {
2252                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2253                                cpt, rc);
2254                         goto failed;
2255                 }
2256         }
2257
2258         return 0;
2259  failed:
2260         kiblnd_net_fini_pools(net);
2261         LASSERT(rc);
2262         return rc;
2263 }
2264
2265 static int kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2266 {
2267         /*
2268          * It's safe to assume a HCA can handle a page size
2269          * matching that of the native system
2270          */
2271         hdev->ibh_page_shift = PAGE_SHIFT;
2272         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2273         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2274
2275         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2276         if (hdev->ibh_mr_size == ~0ULL) {
2277                 hdev->ibh_mr_shift = 64;
2278                 return 0;
2279         }
2280
2281         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2282         return -EINVAL;
2283 }
2284
2285 static void kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2286 {
2287         if (!hdev->ibh_mrs)
2288                 return;
2289
2290         ib_dereg_mr(hdev->ibh_mrs);
2291
2292         hdev->ibh_mrs = NULL;
2293 }
2294
2295 void kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2296 {
2297         kiblnd_hdev_cleanup_mrs(hdev);
2298
2299         if (hdev->ibh_pd)
2300                 ib_dealloc_pd(hdev->ibh_pd);
2301
2302         if (hdev->ibh_cmid)
2303                 rdma_destroy_id(hdev->ibh_cmid);
2304
2305         LIBCFS_FREE(hdev, sizeof(*hdev));
2306 }
2307
2308 static int kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2309 {
2310         struct ib_mr *mr;
2311         int rc;
2312         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2313
2314         rc = kiblnd_hdev_get_attr(hdev);
2315         if (rc)
2316                 return rc;
2317
2318         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2319         if (IS_ERR(mr)) {
2320                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2321                 kiblnd_hdev_cleanup_mrs(hdev);
2322                 return PTR_ERR(mr);
2323         }
2324
2325         hdev->ibh_mrs = mr;
2326
2327         return 0;
2328 }
2329
2330 /* DUMMY */
2331 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2332                                  struct rdma_cm_event *event)
2333 {
2334         return 0;
2335 }
2336
2337 static int kiblnd_dev_need_failover(struct kib_dev *dev)
2338 {
2339         struct rdma_cm_id *cmid;
2340         struct sockaddr_in srcaddr;
2341         struct sockaddr_in dstaddr;
2342         int rc;
2343
2344         if (!dev->ibd_hdev || /* initializing */
2345             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2346             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2347                 return 1;
2348
2349         /*
2350          * XXX: it's UGLY, but I don't have better way to find
2351          * ib-bonding HCA failover because:
2352          *
2353          * a. no reliable CM event for HCA failover...
2354          * b. no OFED API to get ib_device for current net_device...
2355          *
2356          * We have only two choices at this point:
2357          *
2358          * a. rdma_bind_addr(), it will conflict with listener cmid
2359          * b. rdma_resolve_addr() to zero addr
2360          */
2361         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2362                                      IB_QPT_RC);
2363         if (IS_ERR(cmid)) {
2364                 rc = PTR_ERR(cmid);
2365                 CERROR("Failed to create cmid for failover: %d\n", rc);
2366                 return rc;
2367         }
2368
2369         memset(&srcaddr, 0, sizeof(srcaddr));
2370         srcaddr.sin_family = AF_INET;
2371         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2372
2373         memset(&dstaddr, 0, sizeof(dstaddr));
2374         dstaddr.sin_family = AF_INET;
2375         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2376                                (struct sockaddr *)&dstaddr, 1);
2377         if (rc || !cmid->device) {
2378                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2379                        dev->ibd_ifname, &dev->ibd_ifip,
2380                        cmid->device, rc);
2381                 rdma_destroy_id(cmid);
2382                 return rc;
2383         }
2384
2385         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2386         rdma_destroy_id(cmid);
2387
2388         return rc;
2389 }
2390
2391 int kiblnd_dev_failover(struct kib_dev *dev)
2392 {
2393         LIST_HEAD(zombie_tpo);
2394         LIST_HEAD(zombie_ppo);
2395         LIST_HEAD(zombie_fpo);
2396         struct rdma_cm_id *cmid  = NULL;
2397         struct kib_hca_dev *hdev  = NULL;
2398         struct ib_pd *pd;
2399         struct kib_net *net;
2400         struct sockaddr_in addr;
2401         unsigned long flags;
2402         int rc = 0;
2403         int i;
2404
2405         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2406                 dev->ibd_can_failover || !dev->ibd_hdev);
2407
2408         rc = kiblnd_dev_need_failover(dev);
2409         if (rc <= 0)
2410                 goto out;
2411
2412         if (dev->ibd_hdev &&
2413             dev->ibd_hdev->ibh_cmid) {
2414                 /*
2415                  * XXX it's not good to close old listener at here,
2416                  * because we can fail to create new listener.
2417                  * But we have to close it now, otherwise rdma_bind_addr
2418                  * will return EADDRINUSE... How crap!
2419                  */
2420                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2421
2422                 cmid = dev->ibd_hdev->ibh_cmid;
2423                 /*
2424                  * make next schedule of kiblnd_dev_need_failover()
2425                  * return 1 for me
2426                  */
2427                 dev->ibd_hdev->ibh_cmid  = NULL;
2428                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2429
2430                 rdma_destroy_id(cmid);
2431         }
2432
2433         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2434                                      IB_QPT_RC);
2435         if (IS_ERR(cmid)) {
2436                 rc = PTR_ERR(cmid);
2437                 CERROR("Failed to create cmid for failover: %d\n", rc);
2438                 goto out;
2439         }
2440
2441         memset(&addr, 0, sizeof(addr));
2442         addr.sin_family      = AF_INET;
2443         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2444         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2445
2446         /* Bind to failover device or port */
2447         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2448         if (rc || !cmid->device) {
2449                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2450                        dev->ibd_ifname, &dev->ibd_ifip,
2451                        cmid->device, rc);
2452                 rdma_destroy_id(cmid);
2453                 goto out;
2454         }
2455
2456         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2457         if (!hdev) {
2458                 CERROR("Failed to allocate kib_hca_dev\n");
2459                 rdma_destroy_id(cmid);
2460                 rc = -ENOMEM;
2461                 goto out;
2462         }
2463
2464         atomic_set(&hdev->ibh_ref, 1);
2465         hdev->ibh_dev   = dev;
2466         hdev->ibh_cmid  = cmid;
2467         hdev->ibh_ibdev = cmid->device;
2468
2469         pd = ib_alloc_pd(cmid->device, 0);
2470         if (IS_ERR(pd)) {
2471                 rc = PTR_ERR(pd);
2472                 CERROR("Can't allocate PD: %d\n", rc);
2473                 goto out;
2474         }
2475
2476         hdev->ibh_pd = pd;
2477
2478         rc = rdma_listen(cmid, 0);
2479         if (rc) {
2480                 CERROR("Can't start new listener: %d\n", rc);
2481                 goto out;
2482         }
2483
2484         rc = kiblnd_hdev_setup_mrs(hdev);
2485         if (rc) {
2486                 CERROR("Can't setup device: %d\n", rc);
2487                 goto out;
2488         }
2489
2490         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2491
2492         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2493
2494         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2495                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2496                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2497                                             &zombie_tpo);
2498
2499                         if (net->ibn_fmr_ps)
2500                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2501                                                         &zombie_fpo);
2502                 }
2503         }
2504
2505         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2506  out:
2507         if (!list_empty(&zombie_tpo))
2508                 kiblnd_destroy_pool_list(&zombie_tpo);
2509         if (!list_empty(&zombie_ppo))
2510                 kiblnd_destroy_pool_list(&zombie_ppo);
2511         if (!list_empty(&zombie_fpo))
2512                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2513         if (hdev)
2514                 kiblnd_hdev_decref(hdev);
2515
2516         if (rc)
2517                 dev->ibd_failed_failover++;
2518         else
2519                 dev->ibd_failed_failover = 0;
2520
2521         return rc;
2522 }
2523
2524 void kiblnd_destroy_dev(struct kib_dev *dev)
2525 {
2526         LASSERT(!dev->ibd_nnets);
2527         LASSERT(list_empty(&dev->ibd_nets));
2528
2529         list_del(&dev->ibd_fail_list);
2530         list_del(&dev->ibd_list);
2531
2532         if (dev->ibd_hdev)
2533                 kiblnd_hdev_decref(dev->ibd_hdev);
2534
2535         LIBCFS_FREE(dev, sizeof(*dev));
2536 }
2537
2538 static struct kib_dev *kiblnd_create_dev(char *ifname)
2539 {
2540         struct net_device *netdev;
2541         struct kib_dev *dev;
2542         __u32 netmask;
2543         __u32 ip;
2544         int up;
2545         int rc;
2546
2547         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2548         if (rc) {
2549                 CERROR("Can't query IPoIB interface %s: %d\n",
2550                        ifname, rc);
2551                 return NULL;
2552         }
2553
2554         if (!up) {
2555                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2556                 return NULL;
2557         }
2558
2559         LIBCFS_ALLOC(dev, sizeof(*dev));
2560         if (!dev)
2561                 return NULL;
2562
2563         netdev = dev_get_by_name(&init_net, ifname);
2564         if (!netdev) {
2565                 dev->ibd_can_failover = 0;
2566         } else {
2567                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2568                 dev_put(netdev);
2569         }
2570
2571         INIT_LIST_HEAD(&dev->ibd_nets);
2572         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2573         INIT_LIST_HEAD(&dev->ibd_fail_list);
2574         dev->ibd_ifip = ip;
2575         strcpy(&dev->ibd_ifname[0], ifname);
2576
2577         /* initialize the device */
2578         rc = kiblnd_dev_failover(dev);
2579         if (rc) {
2580                 CERROR("Can't initialize device: %d\n", rc);
2581                 LIBCFS_FREE(dev, sizeof(*dev));
2582                 return NULL;
2583         }
2584
2585         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2586         return dev;
2587 }
2588
2589 static void kiblnd_base_shutdown(void)
2590 {
2591         struct kib_sched_info *sched;
2592         int i;
2593
2594         LASSERT(list_empty(&kiblnd_data.kib_devs));
2595
2596         switch (kiblnd_data.kib_init) {
2597         default:
2598                 LBUG();
2599
2600         case IBLND_INIT_ALL:
2601         case IBLND_INIT_DATA:
2602                 LASSERT(kiblnd_data.kib_peers);
2603                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2604                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2605                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2606                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2607                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2608                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2609
2610                 /* flag threads to terminate; wake and wait for them to die */
2611                 kiblnd_data.kib_shutdown = 1;
2612
2613                 /*
2614                  * NB: we really want to stop scheduler threads net by net
2615                  * instead of the whole module, this should be improved
2616                  * with dynamic configuration LNet
2617                  */
2618                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2619                         wake_up_all(&sched->ibs_waitq);
2620
2621                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2622                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2623
2624                 i = 2;
2625                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2626                         i++;
2627                         /* power of 2 ? */
2628                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2629                                "Waiting for %d threads to terminate\n",
2630                                atomic_read(&kiblnd_data.kib_nthreads));
2631                         set_current_state(TASK_UNINTERRUPTIBLE);
2632                         schedule_timeout(cfs_time_seconds(1));
2633                 }
2634
2635                 /* fall through */
2636
2637         case IBLND_INIT_NOTHING:
2638                 break;
2639         }
2640
2641         if (kiblnd_data.kib_peers) {
2642                 LIBCFS_FREE(kiblnd_data.kib_peers,
2643                             sizeof(struct list_head) *
2644                             kiblnd_data.kib_peer_hash_size);
2645         }
2646
2647         if (kiblnd_data.kib_scheds)
2648                 cfs_percpt_free(kiblnd_data.kib_scheds);
2649
2650         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2651         module_put(THIS_MODULE);
2652 }
2653
2654 static void kiblnd_shutdown(lnet_ni_t *ni)
2655 {
2656         struct kib_net *net = ni->ni_data;
2657         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2658         int i;
2659         unsigned long flags;
2660
2661         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2662
2663         if (!net)
2664                 goto out;
2665
2666         write_lock_irqsave(g_lock, flags);
2667         net->ibn_shutdown = 1;
2668         write_unlock_irqrestore(g_lock, flags);
2669
2670         switch (net->ibn_init) {
2671         default:
2672                 LBUG();
2673
2674         case IBLND_INIT_ALL:
2675                 /* nuke all existing peers within this net */
2676                 kiblnd_del_peer(ni, LNET_NID_ANY);
2677
2678                 /* Wait for all peer state to clean up */
2679                 i = 2;
2680                 while (atomic_read(&net->ibn_npeers)) {
2681                         i++;
2682                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2683                                "%s: waiting for %d peers to disconnect\n",
2684                                libcfs_nid2str(ni->ni_nid),
2685                                atomic_read(&net->ibn_npeers));
2686                         set_current_state(TASK_UNINTERRUPTIBLE);
2687                         schedule_timeout(cfs_time_seconds(1));
2688                 }
2689
2690                 kiblnd_net_fini_pools(net);
2691
2692                 write_lock_irqsave(g_lock, flags);
2693                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2694                 net->ibn_dev->ibd_nnets--;
2695                 list_del(&net->ibn_list);
2696                 write_unlock_irqrestore(g_lock, flags);
2697
2698                 /* fall through */
2699
2700         case IBLND_INIT_NOTHING:
2701                 LASSERT(!atomic_read(&net->ibn_nconns));
2702
2703                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2704                         kiblnd_destroy_dev(net->ibn_dev);
2705
2706                 break;
2707         }
2708
2709         net->ibn_init = IBLND_INIT_NOTHING;
2710         ni->ni_data = NULL;
2711
2712         LIBCFS_FREE(net, sizeof(*net));
2713
2714 out:
2715         if (list_empty(&kiblnd_data.kib_devs))
2716                 kiblnd_base_shutdown();
2717 }
2718
2719 static int kiblnd_base_startup(void)
2720 {
2721         struct kib_sched_info *sched;
2722         int rc;
2723         int i;
2724
2725         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2726
2727         try_module_get(THIS_MODULE);
2728         /* zero pointers, flags etc */
2729         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2730
2731         rwlock_init(&kiblnd_data.kib_global_lock);
2732
2733         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2734         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2735
2736         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2737         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2738                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2739         if (!kiblnd_data.kib_peers)
2740                 goto failed;
2741         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2742                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2743
2744         spin_lock_init(&kiblnd_data.kib_connd_lock);
2745         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2746         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2747         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2748         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2749
2750         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2751         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2752
2753         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2754                                                   sizeof(*sched));
2755         if (!kiblnd_data.kib_scheds)
2756                 goto failed;
2757
2758         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2759                 int nthrs;
2760
2761                 spin_lock_init(&sched->ibs_lock);
2762                 INIT_LIST_HEAD(&sched->ibs_conns);
2763                 init_waitqueue_head(&sched->ibs_waitq);
2764
2765                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2766                 if (*kiblnd_tunables.kib_nscheds > 0) {
2767                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2768                 } else {
2769                         /*
2770                          * max to half of CPUs, another half is reserved for
2771                          * upper layer modules
2772                          */
2773                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2774                 }
2775
2776                 sched->ibs_nthreads_max = nthrs;
2777                 sched->ibs_cpt = i;
2778         }
2779
2780         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2781
2782         /* lists/ptrs/locks initialised */
2783         kiblnd_data.kib_init = IBLND_INIT_DATA;
2784         /*****************************************************/
2785
2786         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2787         if (rc) {
2788                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2789                 goto failed;
2790         }
2791
2792         if (*kiblnd_tunables.kib_dev_failover)
2793                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2794                                          "kiblnd_failover");
2795
2796         if (rc) {
2797                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2798                 goto failed;
2799         }
2800
2801         /* flag everything initialised */
2802         kiblnd_data.kib_init = IBLND_INIT_ALL;
2803         /*****************************************************/
2804
2805         return 0;
2806
2807  failed:
2808         kiblnd_base_shutdown();
2809         return -ENETDOWN;
2810 }
2811
2812 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2813 {
2814         int rc = 0;
2815         int nthrs;
2816         int i;
2817
2818         if (!sched->ibs_nthreads) {
2819                 if (*kiblnd_tunables.kib_nscheds > 0) {
2820                         nthrs = sched->ibs_nthreads_max;
2821                 } else {
2822                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2823                                                sched->ibs_cpt);
2824                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2825                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2826                 }
2827         } else {
2828                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2829                 /* increase one thread if there is new interface */
2830                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2831         }
2832
2833         for (i = 0; i < nthrs; i++) {
2834                 long id;
2835                 char name[20];
2836
2837                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2838                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2839                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2840                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2841                 if (!rc)
2842                         continue;
2843
2844                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2845                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2846                 break;
2847         }
2848
2849         sched->ibs_nthreads += i;
2850         return rc;
2851 }
2852
2853 static int kiblnd_dev_start_threads(struct kib_dev *dev, int newdev, __u32 *cpts,
2854                                     int ncpts)
2855 {
2856         int cpt;
2857         int rc;
2858         int i;
2859
2860         for (i = 0; i < ncpts; i++) {
2861                 struct kib_sched_info *sched;
2862
2863                 cpt = !cpts ? i : cpts[i];
2864                 sched = kiblnd_data.kib_scheds[cpt];
2865
2866                 if (!newdev && sched->ibs_nthreads > 0)
2867                         continue;
2868
2869                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2870                 if (rc) {
2871                         CERROR("Failed to start scheduler threads for %s\n",
2872                                dev->ibd_ifname);
2873                         return rc;
2874                 }
2875         }
2876         return 0;
2877 }
2878
2879 static struct kib_dev *kiblnd_dev_search(char *ifname)
2880 {
2881         struct kib_dev *alias = NULL;
2882         struct kib_dev *dev;
2883         char *colon;
2884         char *colon2;
2885
2886         colon = strchr(ifname, ':');
2887         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2888                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2889                         return dev;
2890
2891                 if (alias)
2892                         continue;
2893
2894                 colon2 = strchr(dev->ibd_ifname, ':');
2895                 if (colon)
2896                         *colon = 0;
2897                 if (colon2)
2898                         *colon2 = 0;
2899
2900                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2901                         alias = dev;
2902
2903                 if (colon)
2904                         *colon = ':';
2905                 if (colon2)
2906                         *colon2 = ':';
2907         }
2908         return alias;
2909 }
2910
2911 static int kiblnd_startup(lnet_ni_t *ni)
2912 {
2913         char *ifname;
2914         struct kib_dev *ibdev = NULL;
2915         struct kib_net *net;
2916         struct timespec64 tv;
2917         unsigned long flags;
2918         int rc;
2919         int newdev;
2920
2921         LASSERT(ni->ni_lnd == &the_o2iblnd);
2922
2923         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2924                 rc = kiblnd_base_startup();
2925                 if (rc)
2926                         return rc;
2927         }
2928
2929         LIBCFS_ALLOC(net, sizeof(*net));
2930         ni->ni_data = net;
2931         if (!net)
2932                 goto net_failed;
2933
2934         ktime_get_real_ts64(&tv);
2935         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2936                                tv.tv_nsec / NSEC_PER_USEC;
2937
2938         rc = kiblnd_tunables_setup(ni);
2939         if (rc)
2940                 goto net_failed;
2941
2942         if (ni->ni_interfaces[0]) {
2943                 /* Use the IPoIB interface specified in 'networks=' */
2944
2945                 CLASSERT(LNET_MAX_INTERFACES > 1);
2946                 if (ni->ni_interfaces[1]) {
2947                         CERROR("Multiple interfaces not supported\n");
2948                         goto failed;
2949                 }
2950
2951                 ifname = ni->ni_interfaces[0];
2952         } else {
2953                 ifname = *kiblnd_tunables.kib_default_ipif;
2954         }
2955
2956         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2957                 CERROR("IPoIB interface name too long: %s\n", ifname);
2958                 goto failed;
2959         }
2960
2961         ibdev = kiblnd_dev_search(ifname);
2962
2963         newdev = !ibdev;
2964         /* hmm...create kib_dev even for alias */
2965         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2966                 ibdev = kiblnd_create_dev(ifname);
2967
2968         if (!ibdev)
2969                 goto failed;
2970
2971         net->ibn_dev = ibdev;
2972         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2973
2974         rc = kiblnd_dev_start_threads(ibdev, newdev,
2975                                       ni->ni_cpts, ni->ni_ncpts);
2976         if (rc)
2977                 goto failed;
2978
2979         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
2980         if (rc) {
2981                 CERROR("Failed to initialize NI pools: %d\n", rc);
2982                 goto failed;
2983         }
2984
2985         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2986         ibdev->ibd_nnets++;
2987         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2988         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2989
2990         net->ibn_init = IBLND_INIT_ALL;
2991
2992         return 0;
2993
2994 failed:
2995         if (!net->ibn_dev && ibdev)
2996                 kiblnd_destroy_dev(ibdev);
2997
2998 net_failed:
2999         kiblnd_shutdown(ni);
3000
3001         CDEBUG(D_NET, "kiblnd_startup failed\n");
3002         return -ENETDOWN;
3003 }
3004
3005 static lnd_t the_o2iblnd = {
3006         .lnd_type       = O2IBLND,
3007         .lnd_startup    = kiblnd_startup,
3008         .lnd_shutdown   = kiblnd_shutdown,
3009         .lnd_ctl        = kiblnd_ctl,
3010         .lnd_query      = kiblnd_query,
3011         .lnd_send       = kiblnd_send,
3012         .lnd_recv       = kiblnd_recv,
3013 };
3014
3015 static void __exit ko2iblnd_exit(void)
3016 {
3017         lnet_unregister_lnd(&the_o2iblnd);
3018 }
3019
3020 static int __init ko2iblnd_init(void)
3021 {
3022         CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3023         CLASSERT(offsetof(struct kib_msg,
3024                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3025                           <= IBLND_MSG_SIZE);
3026         CLASSERT(offsetof(struct kib_msg,
3027                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3028                           <= IBLND_MSG_SIZE);
3029
3030         kiblnd_tunables_init();
3031
3032         lnet_register_lnd(&the_o2iblnd);
3033
3034         return 0;
3035 }
3036
3037 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3038 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3039 MODULE_VERSION("2.7.0");
3040 MODULE_LICENSE("GPL");
3041
3042 module_init(ko2iblnd_init);
3043 module_exit(ko2iblnd_exit);