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, true);
 828  failed_1:
 829         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 830  failed_0:
 831         return NULL;
 832 }
 833
 834 void kiblnd_destroy_conn(struct kib_conn *conn, bool free_conn)
 835 {
 836         struct rdma_cm_id *cmid = conn->ibc_cmid;
 837         struct kib_peer *peer = conn->ibc_peer;
 838         int rc;
 839
 840         LASSERT(!in_interrupt());
 841         LASSERT(!atomic_read(&conn->ibc_refcount));
 842         LASSERT(list_empty(&conn->ibc_early_rxs));
 843         LASSERT(list_empty(&conn->ibc_tx_noops));
 844         LASSERT(list_empty(&conn->ibc_tx_queue));
 845         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 846         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 847         LASSERT(list_empty(&conn->ibc_active_txs));
 848         LASSERT(!conn->ibc_noops_posted);
 849         LASSERT(!conn->ibc_nsends_posted);
 850
 851         switch (conn->ibc_state) {
 852         default:
 853                 /* conn must be completely disengaged from the network */
 854                 LBUG();
 855
 856         case IBLND_CONN_DISCONNECTED:
 857                 /* connvars should have been freed already */
 858                 LASSERT(!conn->ibc_connvars);
 859                 break;
 860
 861         case IBLND_CONN_INIT:
 862                 break;
 863         }
 864
 865         /* conn->ibc_cmid might be destroyed by CM already */
 866         if (cmid && cmid->qp)
 867                 rdma_destroy_qp(cmid);
 868
 869         if (conn->ibc_cq) {
 870                 rc = ib_destroy_cq(conn->ibc_cq);
 871                 if (rc)
 872                         CWARN("Error destroying CQ: %d\n", rc);
 873         }
 874
 875         if (conn->ibc_rx_pages)
 876                 kiblnd_unmap_rx_descs(conn);
 877
 878         if (conn->ibc_rxs) {
 879                 LIBCFS_FREE(conn->ibc_rxs,
 880                             IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 881         }
 882
 883         if (conn->ibc_connvars)
 884                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 885
 886         if (conn->ibc_hdev)
 887                 kiblnd_hdev_decref(conn->ibc_hdev);
 888
 889         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 890         if (conn->ibc_state != IBLND_CONN_INIT) {
 891                 struct kib_net *net = peer->ibp_ni->ni_data;
 892
 893                 kiblnd_peer_decref(peer);
 894                 rdma_destroy_id(cmid);
 895                 atomic_dec(&net->ibn_nconns);
 896         }
 897
 898         LIBCFS_FREE(conn, sizeof(*conn));
 899 }
 900
 901 int kiblnd_close_peer_conns_locked(struct kib_peer *peer, int why)
 902 {
 903         struct kib_conn *conn;
 904         struct list_head *ctmp;
 905         struct list_head *cnxt;
 906         int count = 0;
 907
 908         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 909                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 910
 911                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 912                        libcfs_nid2str(peer->ibp_nid),
 913                        conn->ibc_version, why);
 914
 915                 kiblnd_close_conn_locked(conn, why);
 916                 count++;
 917         }
 918
 919         return count;
 920 }
 921
 922 int kiblnd_close_stale_conns_locked(struct kib_peer *peer,
 923                                     int version, __u64 incarnation)
 924 {
 925         struct kib_conn *conn;
 926         struct list_head *ctmp;
 927         struct list_head *cnxt;
 928         int count = 0;
 929
 930         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 931                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 932
 933                 if (conn->ibc_version     == version &&
 934                     conn->ibc_incarnation == incarnation)
 935                         continue;
 936
 937                 CDEBUG(D_NET,
 938                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 939                        libcfs_nid2str(peer->ibp_nid),
 940                        conn->ibc_version, conn->ibc_incarnation,
 941                        version, incarnation);
 942
 943                 kiblnd_close_conn_locked(conn, -ESTALE);
 944                 count++;
 945         }
 946
 947         return count;
 948 }
 949
 950 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 951 {
 952         struct kib_peer *peer;
 953         struct list_head *ptmp;
 954         struct list_head *pnxt;
 955         int lo;
 956         int hi;
 957         int i;
 958         unsigned long flags;
 959         int count = 0;
 960
 961         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 962
 963         if (nid != LNET_NID_ANY) {
 964                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 965                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 966         } else {
 967                 lo = 0;
 968                 hi = kiblnd_data.kib_peer_hash_size - 1;
 969         }
 970
 971         for (i = lo; i <= hi; i++) {
 972                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 973                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 974                         LASSERT(!kiblnd_peer_idle(peer));
 975
 976                         if (peer->ibp_ni != ni)
 977                                 continue;
 978
 979                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 980                                 continue;
 981
 982                         count += kiblnd_close_peer_conns_locked(peer, 0);
 983                 }
 984         }
 985
 986         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 987
 988         /* wildcards always succeed */
 989         if (nid == LNET_NID_ANY)
 990                 return 0;
 991
 992         return !count ? -ENOENT : 0;
 993 }
 994
 995 static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
 996 {
 997         struct libcfs_ioctl_data *data = arg;
 998         int rc = -EINVAL;
 999
1000         switch (cmd) {
1001         case IOC_LIBCFS_GET_PEER: {
1002                 lnet_nid_t nid = 0;
1003                 int count = 0;
1004
1005                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1006                                           &nid, &count);
1007                 data->ioc_nid   = nid;
1008                 data->ioc_count = count;
1009                 break;
1010         }
1011
1012         case IOC_LIBCFS_DEL_PEER: {
1013                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1014                 break;
1015         }
1016         case IOC_LIBCFS_GET_CONN: {
1017                 struct kib_conn *conn;
1018
1019                 rc = 0;
1020                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1021                 if (!conn) {
1022                         rc = -ENOENT;
1023                         break;
1024                 }
1025
1026                 LASSERT(conn->ibc_cmid);
1027                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1028                 if (!conn->ibc_cmid->route.path_rec)
1029                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1030                 else
1031                         data->ioc_u32[0] =
1032                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1033                 kiblnd_conn_decref(conn);
1034                 break;
1035         }
1036         case IOC_LIBCFS_CLOSE_CONNECTION: {
1037                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1038                 break;
1039         }
1040
1041         default:
1042                 break;
1043         }
1044
1045         return rc;
1046 }
1047
1048 static void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1049 {
1050         unsigned long last_alive = 0;
1051         unsigned long now = cfs_time_current();
1052         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1053         struct kib_peer *peer;
1054         unsigned long flags;
1055
1056         read_lock_irqsave(glock, flags);
1057
1058         peer = kiblnd_find_peer_locked(nid);
1059         if (peer)
1060                 last_alive = peer->ibp_last_alive;
1061
1062         read_unlock_irqrestore(glock, flags);
1063
1064         if (last_alive)
1065                 *when = last_alive;
1066
1067         /*
1068          * peer is not persistent in hash, trigger peer creation
1069          * and connection establishment with a NULL tx
1070          */
1071         if (!peer)
1072                 kiblnd_launch_tx(ni, NULL, nid);
1073
1074         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1075                libcfs_nid2str(nid), peer,
1076                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1077 }
1078
1079 static void kiblnd_free_pages(struct kib_pages *p)
1080 {
1081         int npages = p->ibp_npages;
1082         int i;
1083
1084         for (i = 0; i < npages; i++) {
1085                 if (p->ibp_pages[i])
1086                         __free_page(p->ibp_pages[i]);
1087         }
1088
1089         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1090 }
1091
1092 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1093 {
1094         struct kib_pages *p;
1095         int i;
1096
1097         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1098                          offsetof(struct kib_pages, ibp_pages[npages]));
1099         if (!p) {
1100                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1101                 return -ENOMEM;
1102         }
1103
1104         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1105         p->ibp_npages = npages;
1106
1107         for (i = 0; i < npages; i++) {
1108                 p->ibp_pages[i] = alloc_pages_node(
1109                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1110                                     GFP_NOFS, 0);
1111                 if (!p->ibp_pages[i]) {
1112                         CERROR("Can't allocate page %d of %d\n", i, npages);
1113                         kiblnd_free_pages(p);
1114                         return -ENOMEM;
1115                 }
1116         }
1117
1118         *pp = p;
1119         return 0;
1120 }
1121
1122 void kiblnd_unmap_rx_descs(struct kib_conn *conn)
1123 {
1124         struct kib_rx *rx;
1125         int i;
1126
1127         LASSERT(conn->ibc_rxs);
1128         LASSERT(conn->ibc_hdev);
1129
1130         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1131                 rx = &conn->ibc_rxs[i];
1132
1133                 LASSERT(rx->rx_nob >= 0); /* not posted */
1134
1135                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1136                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1137                                                           rx->rx_msgaddr),
1138                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1139         }
1140
1141         kiblnd_free_pages(conn->ibc_rx_pages);
1142
1143         conn->ibc_rx_pages = NULL;
1144 }
1145
1146 void kiblnd_map_rx_descs(struct kib_conn *conn)
1147 {
1148         struct kib_rx *rx;
1149         struct page *pg;
1150         int pg_off;
1151         int ipg;
1152         int i;
1153
1154         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1155                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1156                 rx = &conn->ibc_rxs[i];
1157
1158                 rx->rx_conn = conn;
1159                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1160
1161                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1162                                                        rx->rx_msg,
1163                                                        IBLND_MSG_SIZE,
1164                                                        DMA_FROM_DEVICE);
1165                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1166                                                   rx->rx_msgaddr));
1167                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1168
1169                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1170                        i, rx->rx_msg, rx->rx_msgaddr,
1171                        (__u64)(page_to_phys(pg) + pg_off));
1172
1173                 pg_off += IBLND_MSG_SIZE;
1174                 LASSERT(pg_off <= PAGE_SIZE);
1175
1176                 if (pg_off == PAGE_SIZE) {
1177                         pg_off = 0;
1178                         ipg++;
1179                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1180                 }
1181         }
1182 }
1183
1184 static void kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1185 {
1186         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1187         struct kib_tx *tx;
1188         int i;
1189
1190         LASSERT(!tpo->tpo_pool.po_allocated);
1191
1192         if (!hdev)
1193                 return;
1194
1195         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1196                 tx = &tpo->tpo_tx_descs[i];
1197                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1198                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1199                                                           tx->tx_msgaddr),
1200                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1201         }
1202
1203         kiblnd_hdev_decref(hdev);
1204         tpo->tpo_hdev = NULL;
1205 }
1206
1207 static struct kib_hca_dev *kiblnd_current_hdev(struct kib_dev *dev)
1208 {
1209         struct kib_hca_dev *hdev;
1210         unsigned long flags;
1211         int i = 0;
1212
1213         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1214         while (dev->ibd_failover) {
1215                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1216                 if (!(i++ % 50))
1217                         CDEBUG(D_NET, "%s: Wait for failover\n",
1218                                dev->ibd_ifname);
1219                 set_current_state(TASK_INTERRUPTIBLE);
1220                 schedule_timeout(cfs_time_seconds(1) / 100);
1221
1222                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1223         }
1224
1225         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1226         hdev = dev->ibd_hdev;
1227
1228         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1229
1230         return hdev;
1231 }
1232
1233 static void kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1234 {
1235         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1236         struct kib_pool *pool = &tpo->tpo_pool;
1237         struct kib_net *net = pool->po_owner->ps_net;
1238         struct kib_dev *dev;
1239         struct page *page;
1240         struct kib_tx *tx;
1241         int page_offset;
1242         int ipage;
1243         int i;
1244
1245         LASSERT(net);
1246
1247         dev = net->ibn_dev;
1248
1249         /* pre-mapped messages are not bigger than 1 page */
1250         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1251
1252         /* No fancy arithmetic when we do the buffer calculations */
1253         CLASSERT(!(PAGE_SIZE % IBLND_MSG_SIZE));
1254
1255         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1256
1257         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1258                 page = txpgs->ibp_pages[ipage];
1259                 tx = &tpo->tpo_tx_descs[i];
1260
1261                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1262                                            page_offset);
1263
1264                 tx->tx_msgaddr = kiblnd_dma_map_single(
1265                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1266                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1267                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1268                                                   tx->tx_msgaddr));
1269                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1270
1271                 list_add(&tx->tx_list, &pool->po_free_list);
1272
1273                 page_offset += IBLND_MSG_SIZE;
1274                 LASSERT(page_offset <= PAGE_SIZE);
1275
1276                 if (page_offset == PAGE_SIZE) {
1277                         page_offset = 0;
1278                         ipage++;
1279                         LASSERT(ipage <= txpgs->ibp_npages);
1280                 }
1281         }
1282 }
1283
1284 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd,
1285                                     int negotiated_nfrags)
1286 {
1287         struct kib_net *net = ni->ni_data;
1288         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
1289         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1290         __u16 nfrags;
1291         int mod;
1292
1293         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1294         mod = tunables->lnd_map_on_demand;
1295         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1296
1297         LASSERT(hdev->ibh_mrs);
1298
1299         if (mod > 0 && nfrags <= rd->rd_nfrags)
1300                 return NULL;
1301
1302         return hdev->ibh_mrs;
1303 }
1304
1305 static void kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1306 {
1307         LASSERT(!fpo->fpo_map_count);
1308
1309         if (fpo->fpo_is_fmr) {
1310                 if (fpo->fmr.fpo_fmr_pool)
1311                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1312         } else {
1313                 struct kib_fast_reg_descriptor *frd, *tmp;
1314                 int i = 0;
1315
1316                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1317                                          frd_list) {
1318                         list_del(&frd->frd_list);
1319                         ib_dereg_mr(frd->frd_mr);
1320                         LIBCFS_FREE(frd, sizeof(*frd));
1321                         i++;
1322                 }
1323                 if (i < fpo->fast_reg.fpo_pool_size)
1324                         CERROR("FastReg pool still has %d regions registered\n",
1325                                fpo->fast_reg.fpo_pool_size - i);
1326         }
1327
1328         if (fpo->fpo_hdev)
1329                 kiblnd_hdev_decref(fpo->fpo_hdev);
1330
1331         LIBCFS_FREE(fpo, sizeof(*fpo));
1332 }
1333
1334 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1335 {
1336         struct kib_fmr_pool *fpo, *tmp;
1337
1338         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1339                 list_del(&fpo->fpo_list);
1340                 kiblnd_destroy_fmr_pool(fpo);
1341         }
1342 }
1343
1344 static int
1345 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1346                      int ncpts)
1347 {
1348         int size = tunables->lnd_fmr_pool_size / ncpts;
1349
1350         return max(IBLND_FMR_POOL, size);
1351 }
1352
1353 static int
1354 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1355                          int ncpts)
1356 {
1357         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1358
1359         return max(IBLND_FMR_POOL_FLUSH, size);
1360 }
1361
1362 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1363 {
1364         struct ib_fmr_pool_param param = {
1365                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1366                 .page_shift        = PAGE_SHIFT,
1367                 .access            = (IB_ACCESS_LOCAL_WRITE |
1368                                       IB_ACCESS_REMOTE_WRITE),
1369                 .pool_size         = fps->fps_pool_size,
1370                 .dirty_watermark   = fps->fps_flush_trigger,
1371                 .flush_function    = NULL,
1372                 .flush_arg         = NULL,
1373                 .cache             = !!fps->fps_cache };
1374         int rc = 0;
1375
1376         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1377                                                    &param);
1378         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1379                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1380                 if (rc != -ENOSYS)
1381                         CERROR("Failed to create FMR pool: %d\n", rc);
1382                 else
1383                         CERROR("FMRs are not supported\n");
1384         }
1385
1386         return rc;
1387 }
1388
1389 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1390 {
1391         struct kib_fast_reg_descriptor *frd, *tmp;
1392         int i, rc;
1393
1394         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1395         fpo->fast_reg.fpo_pool_size = 0;
1396         for (i = 0; i < fps->fps_pool_size; i++) {
1397                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1398                                  sizeof(*frd));
1399                 if (!frd) {
1400                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1401                         rc = -ENOMEM;
1402                         goto out;
1403                 }
1404
1405                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1406                                           IB_MR_TYPE_MEM_REG,
1407                                           LNET_MAX_PAYLOAD / PAGE_SIZE);
1408                 if (IS_ERR(frd->frd_mr)) {
1409                         rc = PTR_ERR(frd->frd_mr);
1410                         CERROR("Failed to allocate ib_alloc_mr: %d\n", rc);
1411                         frd->frd_mr = NULL;
1412                         goto out_middle;
1413                 }
1414
1415                 frd->frd_valid = true;
1416
1417                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1418                 fpo->fast_reg.fpo_pool_size++;
1419         }
1420
1421         return 0;
1422
1423 out_middle:
1424         if (frd->frd_mr)
1425                 ib_dereg_mr(frd->frd_mr);
1426         LIBCFS_FREE(frd, sizeof(*frd));
1427
1428 out:
1429         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1430                                  frd_list) {
1431                 list_del(&frd->frd_list);
1432                 ib_dereg_mr(frd->frd_mr);
1433                 LIBCFS_FREE(frd, sizeof(*frd));
1434         }
1435
1436         return rc;
1437 }
1438
1439 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1440                                   struct kib_fmr_pool **pp_fpo)
1441 {
1442         struct kib_dev *dev = fps->fps_net->ibn_dev;
1443         struct ib_device_attr *dev_attr;
1444         struct kib_fmr_pool *fpo;
1445         int rc;
1446
1447         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1448         if (!fpo)
1449                 return -ENOMEM;
1450
1451         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1452         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1453
1454         /* Check for FMR or FastReg support */
1455         fpo->fpo_is_fmr = 0;
1456         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1457             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1458             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1459             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1460                 LCONSOLE_INFO("Using FMR for registration\n");
1461                 fpo->fpo_is_fmr = 1;
1462         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1463                 LCONSOLE_INFO("Using FastReg for registration\n");
1464         } else {
1465                 rc = -ENOSYS;
1466                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1467                 goto out_fpo;
1468         }
1469
1470         if (fpo->fpo_is_fmr)
1471                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1472         else
1473                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1474         if (rc)
1475                 goto out_fpo;
1476
1477         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1478         fpo->fpo_owner = fps;
1479         *pp_fpo = fpo;
1480
1481         return 0;
1482
1483 out_fpo:
1484         kiblnd_hdev_decref(fpo->fpo_hdev);
1485         LIBCFS_FREE(fpo, sizeof(*fpo));
1486         return rc;
1487 }
1488
1489 static void kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps,
1490                                     struct list_head *zombies)
1491 {
1492         if (!fps->fps_net) /* intialized? */
1493                 return;
1494
1495         spin_lock(&fps->fps_lock);
1496
1497         while (!list_empty(&fps->fps_pool_list)) {
1498                 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1499                                                  struct kib_fmr_pool, fpo_list);
1500                 fpo->fpo_failed = 1;
1501                 list_del(&fpo->fpo_list);
1502                 if (!fpo->fpo_map_count)
1503                         list_add(&fpo->fpo_list, zombies);
1504                 else
1505                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1506         }
1507
1508         spin_unlock(&fps->fps_lock);
1509 }
1510
1511 static void kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1512 {
1513         if (fps->fps_net) { /* initialized? */
1514                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1515                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1516         }
1517 }
1518
1519 static int
1520 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1521                         struct kib_net *net,
1522                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1523 {
1524         struct kib_fmr_pool *fpo;
1525         int rc;
1526
1527         memset(fps, 0, sizeof(*fps));
1528
1529         fps->fps_net = net;
1530         fps->fps_cpt = cpt;
1531
1532         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1533         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1534         fps->fps_cache = tunables->lnd_fmr_cache;
1535
1536         spin_lock_init(&fps->fps_lock);
1537         INIT_LIST_HEAD(&fps->fps_pool_list);
1538         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1539
1540         rc = kiblnd_create_fmr_pool(fps, &fpo);
1541         if (!rc)
1542                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1543
1544         return rc;
1545 }
1546
1547 static int kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, unsigned long now)
1548 {
1549         if (fpo->fpo_map_count) /* still in use */
1550                 return 0;
1551         if (fpo->fpo_failed)
1552                 return 1;
1553         return cfs_time_aftereq(now, fpo->fpo_deadline);
1554 }
1555
1556 static int
1557 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1558 {
1559         __u64 *pages = tx->tx_pages;
1560         struct kib_hca_dev *hdev;
1561         int npages;
1562         int size;
1563         int i;
1564
1565         hdev = tx->tx_pool->tpo_hdev;
1566
1567         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1568                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1569                      size += hdev->ibh_page_size) {
1570                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1571                                            hdev->ibh_page_mask) + size;
1572                 }
1573         }
1574
1575         return npages;
1576 }
1577
1578 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1579 {
1580         LIST_HEAD(zombies);
1581         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1582         struct kib_fmr_poolset *fps;
1583         unsigned long now = cfs_time_current();
1584         struct kib_fmr_pool *tmp;
1585         int rc;
1586
1587         if (!fpo)
1588                 return;
1589
1590         fps = fpo->fpo_owner;
1591         if (fpo->fpo_is_fmr) {
1592                 if (fmr->fmr_pfmr) {
1593                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1594                         LASSERT(!rc);
1595                         fmr->fmr_pfmr = NULL;
1596                 }
1597
1598                 if (status) {
1599                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1600                         LASSERT(!rc);
1601                 }
1602         } else {
1603                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1604
1605                 if (frd) {
1606                         frd->frd_valid = false;
1607                         spin_lock(&fps->fps_lock);
1608                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1609                         spin_unlock(&fps->fps_lock);
1610                         fmr->fmr_frd = NULL;
1611                 }
1612         }
1613         fmr->fmr_pool = NULL;
1614
1615         spin_lock(&fps->fps_lock);
1616         fpo->fpo_map_count--;  /* decref the pool */
1617
1618         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1619                 /* the first pool is persistent */
1620                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1621                         continue;
1622
1623                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1624                         list_move(&fpo->fpo_list, &zombies);
1625                         fps->fps_version++;
1626                 }
1627         }
1628         spin_unlock(&fps->fps_lock);
1629
1630         if (!list_empty(&zombies))
1631                 kiblnd_destroy_fmr_pool_list(&zombies);
1632 }
1633
1634 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1635                         struct kib_rdma_desc *rd, __u32 nob, __u64 iov,
1636                         struct kib_fmr *fmr)
1637 {
1638         __u64 *pages = tx->tx_pages;
1639         bool is_rx = (rd != tx->tx_rd);
1640         bool tx_pages_mapped = 0;
1641         struct kib_fmr_pool *fpo;
1642         int npages = 0;
1643         __u64 version;
1644         int rc;
1645
1646  again:
1647         spin_lock(&fps->fps_lock);
1648         version = fps->fps_version;
1649         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1650                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1651                 fpo->fpo_map_count++;
1652
1653                 if (fpo->fpo_is_fmr) {
1654                         struct ib_pool_fmr *pfmr;
1655
1656                         spin_unlock(&fps->fps_lock);
1657
1658                         if (!tx_pages_mapped) {
1659                                 npages = kiblnd_map_tx_pages(tx, rd);
1660                                 tx_pages_mapped = 1;
1661                         }
1662
1663                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1664                                                     pages, npages, iov);
1665                         if (likely(!IS_ERR(pfmr))) {
1666                                 fmr->fmr_key = is_rx ? pfmr->fmr->rkey :
1667                                                        pfmr->fmr->lkey;
1668                                 fmr->fmr_frd = NULL;
1669                                 fmr->fmr_pfmr = pfmr;
1670                                 fmr->fmr_pool = fpo;
1671                                 return 0;
1672                         }
1673                         rc = PTR_ERR(pfmr);
1674                 } else {
1675                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1676                                 struct kib_fast_reg_descriptor *frd;
1677                                 struct ib_reg_wr *wr;
1678                                 struct ib_mr *mr;
1679                                 int n;
1680
1681                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1682                                                        struct kib_fast_reg_descriptor,
1683                                                        frd_list);
1684                                 list_del(&frd->frd_list);
1685                                 spin_unlock(&fps->fps_lock);
1686
1687                                 mr = frd->frd_mr;
1688
1689                                 if (!frd->frd_valid) {
1690                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1691                                         struct ib_send_wr *inv_wr;
1692
1693                                         inv_wr = &frd->frd_inv_wr;
1694                                         memset(inv_wr, 0, sizeof(*inv_wr));
1695                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1696                                         inv_wr->wr_id = IBLND_WID_MR;
1697                                         inv_wr->ex.invalidate_rkey = key;
1698
1699                                         /* Bump the key */
1700                                         key = ib_inc_rkey(key);
1701                                         ib_update_fast_reg_key(mr, key);
1702                                 }
1703
1704                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1705                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1706                                 if (unlikely(n != tx->tx_nfrags)) {
1707                                         CERROR("Failed to map mr %d/%d elements\n",
1708                                                n, tx->tx_nfrags);
1709                                         return n < 0 ? n : -EINVAL;
1710                                 }
1711
1712                                 mr->iova = iov;
1713
1714                                 /* Prepare FastReg WR */
1715                                 wr = &frd->frd_fastreg_wr;
1716                                 memset(wr, 0, sizeof(*wr));
1717                                 wr->wr.opcode = IB_WR_REG_MR;
1718                                 wr->wr.wr_id = IBLND_WID_MR;
1719                                 wr->wr.num_sge = 0;
1720                                 wr->wr.send_flags = 0;
1721                                 wr->mr = mr;
1722                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1723                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1724                                               IB_ACCESS_REMOTE_WRITE);
1725
1726                                 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1727                                 fmr->fmr_frd = frd;
1728                                 fmr->fmr_pfmr = NULL;
1729                                 fmr->fmr_pool = fpo;
1730                                 return 0;
1731                         }
1732                         spin_unlock(&fps->fps_lock);
1733                         rc = -EBUSY;
1734                 }
1735
1736                 spin_lock(&fps->fps_lock);
1737                 fpo->fpo_map_count--;
1738                 if (rc != -EAGAIN) {
1739                         spin_unlock(&fps->fps_lock);
1740                         return rc;
1741                 }
1742
1743                 /* EAGAIN and ... */
1744                 if (version != fps->fps_version) {
1745                         spin_unlock(&fps->fps_lock);
1746                         goto again;
1747                 }
1748         }
1749
1750         if (fps->fps_increasing) {
1751                 spin_unlock(&fps->fps_lock);
1752                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1753                 schedule();
1754                 goto again;
1755         }
1756
1757         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1758                 /* someone failed recently */
1759                 spin_unlock(&fps->fps_lock);
1760                 return -EAGAIN;
1761         }
1762
1763         fps->fps_increasing = 1;
1764         spin_unlock(&fps->fps_lock);
1765
1766         CDEBUG(D_NET, "Allocate new FMR pool\n");
1767         rc = kiblnd_create_fmr_pool(fps, &fpo);
1768         spin_lock(&fps->fps_lock);
1769         fps->fps_increasing = 0;
1770         if (!rc) {
1771                 fps->fps_version++;
1772                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1773         } else {
1774                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1775         }
1776         spin_unlock(&fps->fps_lock);
1777
1778         goto again;
1779 }
1780
1781 static void kiblnd_fini_pool(struct kib_pool *pool)
1782 {
1783         LASSERT(list_empty(&pool->po_free_list));
1784         LASSERT(!pool->po_allocated);
1785
1786         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1787 }
1788
1789 static void kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1790 {
1791         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1792
1793         memset(pool, 0, sizeof(*pool));
1794         INIT_LIST_HEAD(&pool->po_free_list);
1795         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1796         pool->po_owner    = ps;
1797         pool->po_size     = size;
1798 }
1799
1800 static void kiblnd_destroy_pool_list(struct list_head *head)
1801 {
1802         struct kib_pool *pool;
1803
1804         while (!list_empty(head)) {
1805                 pool = list_entry(head->next, struct kib_pool, po_list);
1806                 list_del(&pool->po_list);
1807
1808                 LASSERT(pool->po_owner);
1809                 pool->po_owner->ps_pool_destroy(pool);
1810         }
1811 }
1812
1813 static void kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
1814 {
1815         if (!ps->ps_net) /* intialized? */
1816                 return;
1817
1818         spin_lock(&ps->ps_lock);
1819         while (!list_empty(&ps->ps_pool_list)) {
1820                 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
1821                                             struct kib_pool, po_list);
1822                 po->po_failed = 1;
1823                 list_del(&po->po_list);
1824                 if (!po->po_allocated)
1825                         list_add(&po->po_list, zombies);
1826                 else
1827                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1828         }
1829         spin_unlock(&ps->ps_lock);
1830 }
1831
1832 static void kiblnd_fini_poolset(struct kib_poolset *ps)
1833 {
1834         if (ps->ps_net) { /* initialized? */
1835                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1836                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1837         }
1838 }
1839
1840 static int kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
1841                                struct kib_net *net, char *name, int size,
1842                                kib_ps_pool_create_t po_create,
1843                                kib_ps_pool_destroy_t po_destroy,
1844                                kib_ps_node_init_t nd_init,
1845                                kib_ps_node_fini_t nd_fini)
1846 {
1847         struct kib_pool *pool;
1848         int rc;
1849
1850         memset(ps, 0, sizeof(*ps));
1851
1852         ps->ps_cpt          = cpt;
1853         ps->ps_net          = net;
1854         ps->ps_pool_create  = po_create;
1855         ps->ps_pool_destroy = po_destroy;
1856         ps->ps_node_init    = nd_init;
1857         ps->ps_node_fini    = nd_fini;
1858         ps->ps_pool_size    = size;
1859         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1860             >= sizeof(ps->ps_name))
1861                 return -E2BIG;
1862         spin_lock_init(&ps->ps_lock);
1863         INIT_LIST_HEAD(&ps->ps_pool_list);
1864         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1865
1866         rc = ps->ps_pool_create(ps, size, &pool);
1867         if (!rc)
1868                 list_add(&pool->po_list, &ps->ps_pool_list);
1869         else
1870                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1871
1872         return rc;
1873 }
1874
1875 static int kiblnd_pool_is_idle(struct kib_pool *pool, unsigned long now)
1876 {
1877         if (pool->po_allocated) /* still in use */
1878                 return 0;
1879         if (pool->po_failed)
1880                 return 1;
1881         return cfs_time_aftereq(now, pool->po_deadline);
1882 }
1883
1884 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
1885 {
1886         LIST_HEAD(zombies);
1887         struct kib_poolset *ps = pool->po_owner;
1888         struct kib_pool *tmp;
1889         unsigned long now = cfs_time_current();
1890
1891         spin_lock(&ps->ps_lock);
1892
1893         if (ps->ps_node_fini)
1894                 ps->ps_node_fini(pool, node);
1895
1896         LASSERT(pool->po_allocated > 0);
1897         list_add(node, &pool->po_free_list);
1898         pool->po_allocated--;
1899
1900         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1901                 /* the first pool is persistent */
1902                 if (ps->ps_pool_list.next == &pool->po_list)
1903                         continue;
1904
1905                 if (kiblnd_pool_is_idle(pool, now))
1906                         list_move(&pool->po_list, &zombies);
1907         }
1908         spin_unlock(&ps->ps_lock);
1909
1910         if (!list_empty(&zombies))
1911                 kiblnd_destroy_pool_list(&zombies);
1912 }
1913
1914 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps)
1915 {
1916         struct list_head *node;
1917         struct kib_pool *pool;
1918         unsigned int interval = 1;
1919         unsigned long time_before;
1920         unsigned int trips = 0;
1921         int rc;
1922
1923  again:
1924         spin_lock(&ps->ps_lock);
1925         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1926                 if (list_empty(&pool->po_free_list))
1927                         continue;
1928
1929                 pool->po_allocated++;
1930                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1931                 node = pool->po_free_list.next;
1932                 list_del(node);
1933
1934                 if (ps->ps_node_init) {
1935                         /* still hold the lock */
1936                         ps->ps_node_init(pool, node);
1937                 }
1938                 spin_unlock(&ps->ps_lock);
1939                 return node;
1940         }
1941
1942         /* no available tx pool and ... */
1943         if (ps->ps_increasing) {
1944                 /* another thread is allocating a new pool */
1945                 spin_unlock(&ps->ps_lock);
1946                 trips++;
1947                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting %d HZs for her to complete. trips = %d\n",
1948                        ps->ps_name, interval, trips);
1949
1950                 set_current_state(TASK_INTERRUPTIBLE);
1951                 schedule_timeout(interval);
1952                 if (interval < cfs_time_seconds(1))
1953                         interval *= 2;
1954
1955                 goto again;
1956         }
1957
1958         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1959                 /* someone failed recently */
1960                 spin_unlock(&ps->ps_lock);
1961                 return NULL;
1962         }
1963
1964         ps->ps_increasing = 1;
1965         spin_unlock(&ps->ps_lock);
1966
1967         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1968         time_before = cfs_time_current();
1969         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1970         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
1971                cfs_time_current() - time_before);
1972
1973         spin_lock(&ps->ps_lock);
1974         ps->ps_increasing = 0;
1975         if (!rc) {
1976                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1977         } else {
1978                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1979                 CERROR("Can't allocate new %s pool because out of memory\n",
1980                        ps->ps_name);
1981         }
1982         spin_unlock(&ps->ps_lock);
1983
1984         goto again;
1985 }
1986
1987 static void kiblnd_destroy_tx_pool(struct kib_pool *pool)
1988 {
1989         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool, tpo_pool);
1990         int i;
1991
1992         LASSERT(!pool->po_allocated);
1993
1994         if (tpo->tpo_tx_pages) {
1995                 kiblnd_unmap_tx_pool(tpo);
1996                 kiblnd_free_pages(tpo->tpo_tx_pages);
1997         }
1998
1999         if (!tpo->tpo_tx_descs)
2000                 goto out;
2001
2002         for (i = 0; i < pool->po_size; i++) {
2003                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2004
2005                 list_del(&tx->tx_list);
2006                 if (tx->tx_pages)
2007                         LIBCFS_FREE(tx->tx_pages,
2008                                     LNET_MAX_IOV *
2009                                     sizeof(*tx->tx_pages));
2010                 if (tx->tx_frags)
2011                         LIBCFS_FREE(tx->tx_frags,
2012                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2013                                      sizeof(*tx->tx_frags));
2014                 if (tx->tx_wrq)
2015                         LIBCFS_FREE(tx->tx_wrq,
2016                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2017                                     sizeof(*tx->tx_wrq));
2018                 if (tx->tx_sge)
2019                         LIBCFS_FREE(tx->tx_sge,
2020                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2021                                     sizeof(*tx->tx_sge));
2022                 if (tx->tx_rd)
2023                         LIBCFS_FREE(tx->tx_rd,
2024                                     offsetof(struct kib_rdma_desc,
2025                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2026         }
2027
2028         LIBCFS_FREE(tpo->tpo_tx_descs,
2029                     pool->po_size * sizeof(struct kib_tx));
2030 out:
2031         kiblnd_fini_pool(pool);
2032         LIBCFS_FREE(tpo, sizeof(*tpo));
2033 }
2034
2035 static int kiblnd_tx_pool_size(int ncpts)
2036 {
2037         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2038
2039         return max(IBLND_TX_POOL, ntx);
2040 }
2041
2042 static int kiblnd_create_tx_pool(struct kib_poolset *ps, int size,
2043                                  struct kib_pool **pp_po)
2044 {
2045         int i;
2046         int npg;
2047         struct kib_pool *pool;
2048         struct kib_tx_pool *tpo;
2049
2050         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2051         if (!tpo) {
2052                 CERROR("Failed to allocate TX pool\n");
2053                 return -ENOMEM;
2054         }
2055
2056         pool = &tpo->tpo_pool;
2057         kiblnd_init_pool(ps, pool, size);
2058         tpo->tpo_tx_descs = NULL;
2059         tpo->tpo_tx_pages = NULL;
2060
2061         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2062         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
2063                 CERROR("Can't allocate tx pages: %d\n", npg);
2064                 LIBCFS_FREE(tpo, sizeof(*tpo));
2065                 return -ENOMEM;
2066         }
2067
2068         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2069                          size * sizeof(struct kib_tx));
2070         if (!tpo->tpo_tx_descs) {
2071                 CERROR("Can't allocate %d tx descriptors\n", size);
2072                 ps->ps_pool_destroy(pool);
2073                 return -ENOMEM;
2074         }
2075
2076         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2077
2078         for (i = 0; i < size; i++) {
2079                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2080
2081                 tx->tx_pool = tpo;
2082                 if (ps->ps_net->ibn_fmr_ps) {
2083                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2084                                          lnet_cpt_table(), ps->ps_cpt,
2085                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2086                         if (!tx->tx_pages)
2087                                 break;
2088                 }
2089
2090                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2091                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2092                                  sizeof(*tx->tx_frags));
2093                 if (!tx->tx_frags)
2094                         break;
2095
2096                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2097
2098                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2099                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2100                                  sizeof(*tx->tx_wrq));
2101                 if (!tx->tx_wrq)
2102                         break;
2103
2104                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2105                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2106                                  sizeof(*tx->tx_sge));
2107                 if (!tx->tx_sge)
2108                         break;
2109
2110                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2111                                  offsetof(struct kib_rdma_desc,
2112                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2113                 if (!tx->tx_rd)
2114                         break;
2115         }
2116
2117         if (i == size) {
2118                 kiblnd_map_tx_pool(tpo);
2119                 *pp_po = pool;
2120                 return 0;
2121         }
2122
2123         ps->ps_pool_destroy(pool);
2124         return -ENOMEM;
2125 }
2126
2127 static void kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2128 {
2129         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2130                                                   struct kib_tx_poolset,
2131                                                   tps_poolset);
2132         struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2133
2134         tx->tx_cookie = tps->tps_next_tx_cookie++;
2135 }
2136
2137 static void kiblnd_net_fini_pools(struct kib_net *net)
2138 {
2139         int i;
2140
2141         cfs_cpt_for_each(i, lnet_cpt_table()) {
2142                 struct kib_tx_poolset *tps;
2143                 struct kib_fmr_poolset *fps;
2144
2145                 if (net->ibn_tx_ps) {
2146                         tps = net->ibn_tx_ps[i];
2147                         kiblnd_fini_poolset(&tps->tps_poolset);
2148                 }
2149
2150                 if (net->ibn_fmr_ps) {
2151                         fps = net->ibn_fmr_ps[i];
2152                         kiblnd_fini_fmr_poolset(fps);
2153                 }
2154         }
2155
2156         if (net->ibn_tx_ps) {
2157                 cfs_percpt_free(net->ibn_tx_ps);
2158                 net->ibn_tx_ps = NULL;
2159         }
2160
2161         if (net->ibn_fmr_ps) {
2162                 cfs_percpt_free(net->ibn_fmr_ps);
2163                 net->ibn_fmr_ps = NULL;
2164         }
2165 }
2166
2167 static int kiblnd_net_init_pools(struct kib_net *net, lnet_ni_t *ni, __u32 *cpts,
2168                                  int ncpts)
2169 {
2170         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2171         unsigned long flags;
2172         int cpt;
2173         int rc;
2174         int i;
2175
2176         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2177
2178         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2179         if (!tunables->lnd_map_on_demand) {
2180                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2181                 goto create_tx_pool;
2182         }
2183
2184         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2185
2186         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2187                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2188                        tunables->lnd_fmr_pool_size,
2189                        *kiblnd_tunables.kib_ntx / 4);
2190                 rc = -EINVAL;
2191                 goto failed;
2192         }
2193
2194         /*
2195          * TX pool must be created later than FMR, see LU-2268
2196          * for details
2197          */
2198         LASSERT(!net->ibn_tx_ps);
2199
2200         /*
2201          * premapping can fail if ibd_nmr > 1, so we always create
2202          * FMR pool and map-on-demand if premapping failed
2203          *
2204          * cfs_precpt_alloc is creating an array of struct kib_fmr_poolset
2205          * The number of struct kib_fmr_poolsets create is equal to the
2206          * number of CPTs that exist, i.e net->ibn_fmr_ps[cpt].
2207          */
2208         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2209                                            sizeof(struct kib_fmr_poolset));
2210         if (!net->ibn_fmr_ps) {
2211                 CERROR("Failed to allocate FMR pool array\n");
2212                 rc = -ENOMEM;
2213                 goto failed;
2214         }
2215
2216         for (i = 0; i < ncpts; i++) {
2217                 cpt = !cpts ? i : cpts[i];
2218                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2219                                              net, tunables);
2220                 if (rc) {
2221                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2222                                cpt, rc);
2223                         goto failed;
2224                 }
2225         }
2226
2227         if (i > 0)
2228                 LASSERT(i == ncpts);
2229
2230  create_tx_pool:
2231         /*
2232          * cfs_precpt_alloc is creating an array of struct kib_tx_poolset
2233          * The number of struct kib_tx_poolsets create is equal to the
2234          * number of CPTs that exist, i.e net->ibn_tx_ps[cpt].
2235          */
2236         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2237                                           sizeof(struct kib_tx_poolset));
2238         if (!net->ibn_tx_ps) {
2239                 CERROR("Failed to allocate tx pool array\n");
2240                 rc = -ENOMEM;
2241                 goto failed;
2242         }
2243
2244         for (i = 0; i < ncpts; i++) {
2245                 cpt = !cpts ? i : cpts[i];
2246                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2247                                          cpt, net, "TX",
2248                                          kiblnd_tx_pool_size(ncpts),
2249                                          kiblnd_create_tx_pool,
2250                                          kiblnd_destroy_tx_pool,
2251                                          kiblnd_tx_init, NULL);
2252                 if (rc) {
2253                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2254                                cpt, rc);
2255                         goto failed;
2256                 }
2257         }
2258
2259         return 0;
2260  failed:
2261         kiblnd_net_fini_pools(net);
2262         LASSERT(rc);
2263         return rc;
2264 }
2265
2266 static int kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2267 {
2268         /*
2269          * It's safe to assume a HCA can handle a page size
2270          * matching that of the native system
2271          */
2272         hdev->ibh_page_shift = PAGE_SHIFT;
2273         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2274         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2275
2276         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2277         if (hdev->ibh_mr_size == ~0ULL) {
2278                 hdev->ibh_mr_shift = 64;
2279                 return 0;
2280         }
2281
2282         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2283         return -EINVAL;
2284 }
2285
2286 static void kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2287 {
2288         if (!hdev->ibh_mrs)
2289                 return;
2290
2291         ib_dereg_mr(hdev->ibh_mrs);
2292
2293         hdev->ibh_mrs = NULL;
2294 }
2295
2296 void kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2297 {
2298         kiblnd_hdev_cleanup_mrs(hdev);
2299
2300         if (hdev->ibh_pd)
2301                 ib_dealloc_pd(hdev->ibh_pd);
2302
2303         if (hdev->ibh_cmid)
2304                 rdma_destroy_id(hdev->ibh_cmid);
2305
2306         LIBCFS_FREE(hdev, sizeof(*hdev));
2307 }
2308
2309 static int kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2310 {
2311         struct ib_mr *mr;
2312         int rc;
2313         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2314
2315         rc = kiblnd_hdev_get_attr(hdev);
2316         if (rc)
2317                 return rc;
2318
2319         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2320         if (IS_ERR(mr)) {
2321                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2322                 kiblnd_hdev_cleanup_mrs(hdev);
2323                 return PTR_ERR(mr);
2324         }
2325
2326         hdev->ibh_mrs = mr;
2327
2328         return 0;
2329 }
2330
2331 /* DUMMY */
2332 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2333                                  struct rdma_cm_event *event)
2334 {
2335         return 0;
2336 }
2337
2338 static int kiblnd_dev_need_failover(struct kib_dev *dev)
2339 {
2340         struct rdma_cm_id *cmid;
2341         struct sockaddr_in srcaddr;
2342         struct sockaddr_in dstaddr;
2343         int rc;
2344
2345         if (!dev->ibd_hdev || /* initializing */
2346             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2347             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2348                 return 1;
2349
2350         /*
2351          * XXX: it's UGLY, but I don't have better way to find
2352          * ib-bonding HCA failover because:
2353          *
2354          * a. no reliable CM event for HCA failover...
2355          * b. no OFED API to get ib_device for current net_device...
2356          *
2357          * We have only two choices at this point:
2358          *
2359          * a. rdma_bind_addr(), it will conflict with listener cmid
2360          * b. rdma_resolve_addr() to zero addr
2361          */
2362         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2363                                      IB_QPT_RC);
2364         if (IS_ERR(cmid)) {
2365                 rc = PTR_ERR(cmid);
2366                 CERROR("Failed to create cmid for failover: %d\n", rc);
2367                 return rc;
2368         }
2369
2370         memset(&srcaddr, 0, sizeof(srcaddr));
2371         srcaddr.sin_family = AF_INET;
2372         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2373
2374         memset(&dstaddr, 0, sizeof(dstaddr));
2375         dstaddr.sin_family = AF_INET;
2376         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2377                                (struct sockaddr *)&dstaddr, 1);
2378         if (rc || !cmid->device) {
2379                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2380                        dev->ibd_ifname, &dev->ibd_ifip,
2381                        cmid->device, rc);
2382                 rdma_destroy_id(cmid);
2383                 return rc;
2384         }
2385
2386         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2387         rdma_destroy_id(cmid);
2388
2389         return rc;
2390 }
2391
2392 int kiblnd_dev_failover(struct kib_dev *dev)
2393 {
2394         LIST_HEAD(zombie_tpo);
2395         LIST_HEAD(zombie_ppo);
2396         LIST_HEAD(zombie_fpo);
2397         struct rdma_cm_id *cmid  = NULL;
2398         struct kib_hca_dev *hdev  = NULL;
2399         struct ib_pd *pd;
2400         struct kib_net *net;
2401         struct sockaddr_in addr;
2402         unsigned long flags;
2403         int rc = 0;
2404         int i;
2405
2406         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2407                 dev->ibd_can_failover || !dev->ibd_hdev);
2408
2409         rc = kiblnd_dev_need_failover(dev);
2410         if (rc <= 0)
2411                 goto out;
2412
2413         if (dev->ibd_hdev &&
2414             dev->ibd_hdev->ibh_cmid) {
2415                 /*
2416                  * XXX it's not good to close old listener at here,
2417                  * because we can fail to create new listener.
2418                  * But we have to close it now, otherwise rdma_bind_addr
2419                  * will return EADDRINUSE... How crap!
2420                  */
2421                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2422
2423                 cmid = dev->ibd_hdev->ibh_cmid;
2424                 /*
2425                  * make next schedule of kiblnd_dev_need_failover()
2426                  * return 1 for me
2427                  */
2428                 dev->ibd_hdev->ibh_cmid  = NULL;
2429                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2430
2431                 rdma_destroy_id(cmid);
2432         }
2433
2434         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2435                                      IB_QPT_RC);
2436         if (IS_ERR(cmid)) {
2437                 rc = PTR_ERR(cmid);
2438                 CERROR("Failed to create cmid for failover: %d\n", rc);
2439                 goto out;
2440         }
2441
2442         memset(&addr, 0, sizeof(addr));
2443         addr.sin_family      = AF_INET;
2444         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2445         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2446
2447         /* Bind to failover device or port */
2448         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2449         if (rc || !cmid->device) {
2450                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2451                        dev->ibd_ifname, &dev->ibd_ifip,
2452                        cmid->device, rc);
2453                 rdma_destroy_id(cmid);
2454                 goto out;
2455         }
2456
2457         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2458         if (!hdev) {
2459                 CERROR("Failed to allocate kib_hca_dev\n");
2460                 rdma_destroy_id(cmid);
2461                 rc = -ENOMEM;
2462                 goto out;
2463         }
2464
2465         atomic_set(&hdev->ibh_ref, 1);
2466         hdev->ibh_dev   = dev;
2467         hdev->ibh_cmid  = cmid;
2468         hdev->ibh_ibdev = cmid->device;
2469
2470         pd = ib_alloc_pd(cmid->device, 0);
2471         if (IS_ERR(pd)) {
2472                 rc = PTR_ERR(pd);
2473                 CERROR("Can't allocate PD: %d\n", rc);
2474                 goto out;
2475         }
2476
2477         hdev->ibh_pd = pd;
2478
2479         rc = rdma_listen(cmid, 0);
2480         if (rc) {
2481                 CERROR("Can't start new listener: %d\n", rc);
2482                 goto out;
2483         }
2484
2485         rc = kiblnd_hdev_setup_mrs(hdev);
2486         if (rc) {
2487                 CERROR("Can't setup device: %d\n", rc);
2488                 goto out;
2489         }
2490
2491         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2492
2493         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2494
2495         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2496                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2497                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2498                                             &zombie_tpo);
2499
2500                         if (net->ibn_fmr_ps)
2501                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2502                                                         &zombie_fpo);
2503                 }
2504         }
2505
2506         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2507  out:
2508         if (!list_empty(&zombie_tpo))
2509                 kiblnd_destroy_pool_list(&zombie_tpo);
2510         if (!list_empty(&zombie_ppo))
2511                 kiblnd_destroy_pool_list(&zombie_ppo);
2512         if (!list_empty(&zombie_fpo))
2513                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2514         if (hdev)
2515                 kiblnd_hdev_decref(hdev);
2516
2517         if (rc)
2518                 dev->ibd_failed_failover++;
2519         else
2520                 dev->ibd_failed_failover = 0;
2521
2522         return rc;
2523 }
2524
2525 void kiblnd_destroy_dev(struct kib_dev *dev)
2526 {
2527         LASSERT(!dev->ibd_nnets);
2528         LASSERT(list_empty(&dev->ibd_nets));
2529
2530         list_del(&dev->ibd_fail_list);
2531         list_del(&dev->ibd_list);
2532
2533         if (dev->ibd_hdev)
2534                 kiblnd_hdev_decref(dev->ibd_hdev);
2535
2536         LIBCFS_FREE(dev, sizeof(*dev));
2537 }
2538
2539 static struct kib_dev *kiblnd_create_dev(char *ifname)
2540 {
2541         struct net_device *netdev;
2542         struct kib_dev *dev;
2543         __u32 netmask;
2544         __u32 ip;
2545         int up;
2546         int rc;
2547
2548         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2549         if (rc) {
2550                 CERROR("Can't query IPoIB interface %s: %d\n",
2551                        ifname, rc);
2552                 return NULL;
2553         }
2554
2555         if (!up) {
2556                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2557                 return NULL;
2558         }
2559
2560         LIBCFS_ALLOC(dev, sizeof(*dev));
2561         if (!dev)
2562                 return NULL;
2563
2564         netdev = dev_get_by_name(&init_net, ifname);
2565         if (!netdev) {
2566                 dev->ibd_can_failover = 0;
2567         } else {
2568                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2569                 dev_put(netdev);
2570         }
2571
2572         INIT_LIST_HEAD(&dev->ibd_nets);
2573         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2574         INIT_LIST_HEAD(&dev->ibd_fail_list);
2575         dev->ibd_ifip = ip;
2576         strcpy(&dev->ibd_ifname[0], ifname);
2577
2578         /* initialize the device */
2579         rc = kiblnd_dev_failover(dev);
2580         if (rc) {
2581                 CERROR("Can't initialize device: %d\n", rc);
2582                 LIBCFS_FREE(dev, sizeof(*dev));
2583                 return NULL;
2584         }
2585
2586         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2587         return dev;
2588 }
2589
2590 static void kiblnd_base_shutdown(void)
2591 {
2592         struct kib_sched_info *sched;
2593         int i;
2594
2595         LASSERT(list_empty(&kiblnd_data.kib_devs));
2596
2597         switch (kiblnd_data.kib_init) {
2598         default:
2599                 LBUG();
2600
2601         case IBLND_INIT_ALL:
2602         case IBLND_INIT_DATA:
2603                 LASSERT(kiblnd_data.kib_peers);
2604                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2605                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2606                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2607                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2608                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2609                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2610
2611                 /* flag threads to terminate; wake and wait for them to die */
2612                 kiblnd_data.kib_shutdown = 1;
2613
2614                 /*
2615                  * NB: we really want to stop scheduler threads net by net
2616                  * instead of the whole module, this should be improved
2617                  * with dynamic configuration LNet
2618                  */
2619                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2620                         wake_up_all(&sched->ibs_waitq);
2621
2622                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2623                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2624
2625                 i = 2;
2626                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2627                         i++;
2628                         /* power of 2 ? */
2629                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2630                                "Waiting for %d threads to terminate\n",
2631                                atomic_read(&kiblnd_data.kib_nthreads));
2632                         set_current_state(TASK_UNINTERRUPTIBLE);
2633                         schedule_timeout(cfs_time_seconds(1));
2634                 }
2635
2636                 /* fall through */
2637
2638         case IBLND_INIT_NOTHING:
2639                 break;
2640         }
2641
2642         if (kiblnd_data.kib_peers) {
2643                 LIBCFS_FREE(kiblnd_data.kib_peers,
2644                             sizeof(struct list_head) *
2645                             kiblnd_data.kib_peer_hash_size);
2646         }
2647
2648         if (kiblnd_data.kib_scheds)
2649                 cfs_percpt_free(kiblnd_data.kib_scheds);
2650
2651         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2652         module_put(THIS_MODULE);
2653 }
2654
2655 static void kiblnd_shutdown(lnet_ni_t *ni)
2656 {
2657         struct kib_net *net = ni->ni_data;
2658         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2659         int i;
2660         unsigned long flags;
2661
2662         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2663
2664         if (!net)
2665                 goto out;
2666
2667         write_lock_irqsave(g_lock, flags);
2668         net->ibn_shutdown = 1;
2669         write_unlock_irqrestore(g_lock, flags);
2670
2671         switch (net->ibn_init) {
2672         default:
2673                 LBUG();
2674
2675         case IBLND_INIT_ALL:
2676                 /* nuke all existing peers within this net */
2677                 kiblnd_del_peer(ni, LNET_NID_ANY);
2678
2679                 /* Wait for all peer state to clean up */
2680                 i = 2;
2681                 while (atomic_read(&net->ibn_npeers)) {
2682                         i++;
2683                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2684                                "%s: waiting for %d peers to disconnect\n",
2685                                libcfs_nid2str(ni->ni_nid),
2686                                atomic_read(&net->ibn_npeers));
2687                         set_current_state(TASK_UNINTERRUPTIBLE);
2688                         schedule_timeout(cfs_time_seconds(1));
2689                 }
2690
2691                 kiblnd_net_fini_pools(net);
2692
2693                 write_lock_irqsave(g_lock, flags);
2694                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2695                 net->ibn_dev->ibd_nnets--;
2696                 list_del(&net->ibn_list);
2697                 write_unlock_irqrestore(g_lock, flags);
2698
2699                 /* fall through */
2700
2701         case IBLND_INIT_NOTHING:
2702                 LASSERT(!atomic_read(&net->ibn_nconns));
2703
2704                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2705                         kiblnd_destroy_dev(net->ibn_dev);
2706
2707                 break;
2708         }
2709
2710         net->ibn_init = IBLND_INIT_NOTHING;
2711         ni->ni_data = NULL;
2712
2713         LIBCFS_FREE(net, sizeof(*net));
2714
2715 out:
2716         if (list_empty(&kiblnd_data.kib_devs))
2717                 kiblnd_base_shutdown();
2718 }
2719
2720 static int kiblnd_base_startup(void)
2721 {
2722         struct kib_sched_info *sched;
2723         int rc;
2724         int i;
2725
2726         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2727
2728         try_module_get(THIS_MODULE);
2729         /* zero pointers, flags etc */
2730         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2731
2732         rwlock_init(&kiblnd_data.kib_global_lock);
2733
2734         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2735         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2736
2737         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2738         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2739                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2740         if (!kiblnd_data.kib_peers)
2741                 goto failed;
2742         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2743                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2744
2745         spin_lock_init(&kiblnd_data.kib_connd_lock);
2746         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2747         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2748         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2749         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2750
2751         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2752         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2753
2754         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2755                                                   sizeof(*sched));
2756         if (!kiblnd_data.kib_scheds)
2757                 goto failed;
2758
2759         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2760                 int nthrs;
2761
2762                 spin_lock_init(&sched->ibs_lock);
2763                 INIT_LIST_HEAD(&sched->ibs_conns);
2764                 init_waitqueue_head(&sched->ibs_waitq);
2765
2766                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2767                 if (*kiblnd_tunables.kib_nscheds > 0) {
2768                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2769                 } else {
2770                         /*
2771                          * max to half of CPUs, another half is reserved for
2772                          * upper layer modules
2773                          */
2774                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2775                 }
2776
2777                 sched->ibs_nthreads_max = nthrs;
2778                 sched->ibs_cpt = i;
2779         }
2780
2781         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2782
2783         /* lists/ptrs/locks initialised */
2784         kiblnd_data.kib_init = IBLND_INIT_DATA;
2785         /*****************************************************/
2786
2787         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2788         if (rc) {
2789                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2790                 goto failed;
2791         }
2792
2793         if (*kiblnd_tunables.kib_dev_failover)
2794                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2795                                          "kiblnd_failover");
2796
2797         if (rc) {
2798                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2799                 goto failed;
2800         }
2801
2802         /* flag everything initialised */
2803         kiblnd_data.kib_init = IBLND_INIT_ALL;
2804         /*****************************************************/
2805
2806         return 0;
2807
2808  failed:
2809         kiblnd_base_shutdown();
2810         return -ENETDOWN;
2811 }
2812
2813 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2814 {
2815         int rc = 0;
2816         int nthrs;
2817         int i;
2818
2819         if (!sched->ibs_nthreads) {
2820                 if (*kiblnd_tunables.kib_nscheds > 0) {
2821                         nthrs = sched->ibs_nthreads_max;
2822                 } else {
2823                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2824                                                sched->ibs_cpt);
2825                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2826                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2827                 }
2828         } else {
2829                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2830                 /* increase one thread if there is new interface */
2831                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2832         }
2833
2834         for (i = 0; i < nthrs; i++) {
2835                 long id;
2836                 char name[20];
2837
2838                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2839                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2840                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2841                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2842                 if (!rc)
2843                         continue;
2844
2845                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2846                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2847                 break;
2848         }
2849
2850         sched->ibs_nthreads += i;
2851         return rc;
2852 }
2853
2854 static int kiblnd_dev_start_threads(struct kib_dev *dev, int newdev, __u32 *cpts,
2855                                     int ncpts)
2856 {
2857         int cpt;
2858         int rc;
2859         int i;
2860
2861         for (i = 0; i < ncpts; i++) {
2862                 struct kib_sched_info *sched;
2863
2864                 cpt = !cpts ? i : cpts[i];
2865                 sched = kiblnd_data.kib_scheds[cpt];
2866
2867                 if (!newdev && sched->ibs_nthreads > 0)
2868                         continue;
2869
2870                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2871                 if (rc) {
2872                         CERROR("Failed to start scheduler threads for %s\n",
2873                                dev->ibd_ifname);
2874                         return rc;
2875                 }
2876         }
2877         return 0;
2878 }
2879
2880 static struct kib_dev *kiblnd_dev_search(char *ifname)
2881 {
2882         struct kib_dev *alias = NULL;
2883         struct kib_dev *dev;
2884         char *colon;
2885         char *colon2;
2886
2887         colon = strchr(ifname, ':');
2888         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2889                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2890                         return dev;
2891
2892                 if (alias)
2893                         continue;
2894
2895                 colon2 = strchr(dev->ibd_ifname, ':');
2896                 if (colon)
2897                         *colon = 0;
2898                 if (colon2)
2899                         *colon2 = 0;
2900
2901                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2902                         alias = dev;
2903
2904                 if (colon)
2905                         *colon = ':';
2906                 if (colon2)
2907                         *colon2 = ':';
2908         }
2909         return alias;
2910 }
2911
2912 static int kiblnd_startup(lnet_ni_t *ni)
2913 {
2914         char *ifname;
2915         struct kib_dev *ibdev = NULL;
2916         struct kib_net *net;
2917         struct timespec64 tv;
2918         unsigned long flags;
2919         int rc;
2920         int newdev;
2921
2922         LASSERT(ni->ni_lnd == &the_o2iblnd);
2923
2924         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2925                 rc = kiblnd_base_startup();
2926                 if (rc)
2927                         return rc;
2928         }
2929
2930         LIBCFS_ALLOC(net, sizeof(*net));
2931         ni->ni_data = net;
2932         if (!net)
2933                 goto net_failed;
2934
2935         ktime_get_real_ts64(&tv);
2936         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2937                                tv.tv_nsec / NSEC_PER_USEC;
2938
2939         rc = kiblnd_tunables_setup(ni);
2940         if (rc)
2941                 goto net_failed;
2942
2943         if (ni->ni_interfaces[0]) {
2944                 /* Use the IPoIB interface specified in 'networks=' */
2945
2946                 CLASSERT(LNET_MAX_INTERFACES > 1);
2947                 if (ni->ni_interfaces[1]) {
2948                         CERROR("Multiple interfaces not supported\n");
2949                         goto failed;
2950                 }
2951
2952                 ifname = ni->ni_interfaces[0];
2953         } else {
2954                 ifname = *kiblnd_tunables.kib_default_ipif;
2955         }
2956
2957         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2958                 CERROR("IPoIB interface name too long: %s\n", ifname);
2959                 goto failed;
2960         }
2961
2962         ibdev = kiblnd_dev_search(ifname);
2963
2964         newdev = !ibdev;
2965         /* hmm...create kib_dev even for alias */
2966         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2967                 ibdev = kiblnd_create_dev(ifname);
2968
2969         if (!ibdev)
2970                 goto failed;
2971
2972         net->ibn_dev = ibdev;
2973         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2974
2975         rc = kiblnd_dev_start_threads(ibdev, newdev,
2976                                       ni->ni_cpts, ni->ni_ncpts);
2977         if (rc)
2978                 goto failed;
2979
2980         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
2981         if (rc) {
2982                 CERROR("Failed to initialize NI pools: %d\n", rc);
2983                 goto failed;
2984         }
2985
2986         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2987         ibdev->ibd_nnets++;
2988         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2989         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2990
2991         net->ibn_init = IBLND_INIT_ALL;
2992
2993         return 0;
2994
2995 failed:
2996         if (!net->ibn_dev && ibdev)
2997                 kiblnd_destroy_dev(ibdev);
2998
2999 net_failed:
3000         kiblnd_shutdown(ni);
3001
3002         CDEBUG(D_NET, "kiblnd_startup failed\n");
3003         return -ENETDOWN;
3004 }
3005
3006 static lnd_t the_o2iblnd = {
3007         .lnd_type       = O2IBLND,
3008         .lnd_startup    = kiblnd_startup,
3009         .lnd_shutdown   = kiblnd_shutdown,
3010         .lnd_ctl        = kiblnd_ctl,
3011         .lnd_query      = kiblnd_query,
3012         .lnd_send       = kiblnd_send,
3013         .lnd_recv       = kiblnd_recv,
3014 };
3015
3016 static void __exit ko2iblnd_exit(void)
3017 {
3018         lnet_unregister_lnd(&the_o2iblnd);
3019 }
3020
3021 static int __init ko2iblnd_init(void)
3022 {
3023         CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3024         CLASSERT(offsetof(struct kib_msg,
3025                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3026                           <= IBLND_MSG_SIZE);
3027         CLASSERT(offsetof(struct kib_msg,
3028                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3029                           <= IBLND_MSG_SIZE);
3030
3031         kiblnd_tunables_init();
3032
3033         lnet_register_lnd(&the_o2iblnd);
3034
3035         return 0;
3036 }
3037
3038 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3039 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3040 MODULE_VERSION("2.7.0");
3041 MODULE_LICENSE("GPL");
3042
3043 module_init(ko2iblnd_init);
3044 module_exit(ko2iblnd_exit);