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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 lnd_t the_o2iblnd = {
  44         .lnd_type       = O2IBLND,
  45         .lnd_startup    = kiblnd_startup,
  46         .lnd_shutdown   = kiblnd_shutdown,
  47         .lnd_ctl        = kiblnd_ctl,
  48         .lnd_query      = kiblnd_query,
  49         .lnd_send       = kiblnd_send,
  50         .lnd_recv       = kiblnd_recv,
  51 };
  52
  53 kib_data_t              kiblnd_data;
  54
  55 __u32
  56 kiblnd_cksum (void *ptr, int nob)
  57 {
  58         char  *c  = ptr;
  59         __u32  sum = 0;
  60
  61         while (nob-- > 0)
  62                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  63
  64         /* ensure I don't return 0 (== no checksum) */
  65         return (sum == 0) ? 1 : sum;
  66 }
  67
  68 static char *
  69 kiblnd_msgtype2str(int type)
  70 {
  71         switch (type) {
  72         case IBLND_MSG_CONNREQ:
  73                 return "CONNREQ";
  74
  75         case IBLND_MSG_CONNACK:
  76                 return "CONNACK";
  77
  78         case IBLND_MSG_NOOP:
  79                 return "NOOP";
  80
  81         case IBLND_MSG_IMMEDIATE:
  82                 return "IMMEDIATE";
  83
  84         case IBLND_MSG_PUT_REQ:
  85                 return "PUT_REQ";
  86
  87         case IBLND_MSG_PUT_NAK:
  88                 return "PUT_NAK";
  89
  90         case IBLND_MSG_PUT_ACK:
  91                 return "PUT_ACK";
  92
  93         case IBLND_MSG_PUT_DONE:
  94                 return "PUT_DONE";
  95
  96         case IBLND_MSG_GET_REQ:
  97                 return "GET_REQ";
  98
  99         case IBLND_MSG_GET_DONE:
 100                 return "GET_DONE";
 101
 102         default:
 103                 return "???";
 104         }
 105 }
 106
 107 static int
 108 kiblnd_msgtype2size(int type)
 109 {
 110         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 111
 112         switch (type) {
 113         case IBLND_MSG_CONNREQ:
 114         case IBLND_MSG_CONNACK:
 115                 return hdr_size + sizeof(kib_connparams_t);
 116
 117         case IBLND_MSG_NOOP:
 118                 return hdr_size;
 119
 120         case IBLND_MSG_IMMEDIATE:
 121                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 122
 123         case IBLND_MSG_PUT_REQ:
 124                 return hdr_size + sizeof(kib_putreq_msg_t);
 125
 126         case IBLND_MSG_PUT_ACK:
 127                 return hdr_size + sizeof(kib_putack_msg_t);
 128
 129         case IBLND_MSG_GET_REQ:
 130                 return hdr_size + sizeof(kib_get_msg_t);
 131
 132         case IBLND_MSG_PUT_NAK:
 133         case IBLND_MSG_PUT_DONE:
 134         case IBLND_MSG_GET_DONE:
 135                 return hdr_size + sizeof(kib_completion_msg_t);
 136         default:
 137                 return -1;
 138         }
 139 }
 140
 141 static int
 142 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 143 {
 144         kib_rdma_desc_t   *rd;
 145         int                nob;
 146         int                n;
 147         int                i;
 148
 149         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 150                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 151
 152         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 153                               &msg->ibm_u.get.ibgm_rd :
 154                               &msg->ibm_u.putack.ibpam_rd;
 155
 156         if (flip) {
 157                 __swab32s(&rd->rd_key);
 158                 __swab32s(&rd->rd_nfrags);
 159         }
 160
 161         n = rd->rd_nfrags;
 162
 163         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 164                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 165                        n, IBLND_MAX_RDMA_FRAGS);
 166                 return 1;
 167         }
 168
 169         nob = offsetof (kib_msg_t, ibm_u) +
 170               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 171
 172         if (msg->ibm_nob < nob) {
 173                 CERROR("Short %s: %d(%d)\n",
 174                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 175                 return 1;
 176         }
 177
 178         if (!flip)
 179                 return 0;
 180
 181         for (i = 0; i < n; i++) {
 182                 __swab32s(&rd->rd_frags[i].rf_nob);
 183                 __swab64s(&rd->rd_frags[i].rf_addr);
 184         }
 185
 186         return 0;
 187 }
 188
 189 void
 190 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 191                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 192 {
 193         kib_net_t *net = ni->ni_data;
 194
 195         /* CAVEAT EMPTOR! all message fields not set here should have been
 196          * initialised previously. */
 197         msg->ibm_magic    = IBLND_MSG_MAGIC;
 198         msg->ibm_version  = version;
 199         /*   ibm_type */
 200         msg->ibm_credits  = credits;
 201         /*   ibm_nob */
 202         msg->ibm_cksum    = 0;
 203         msg->ibm_srcnid   = ni->ni_nid;
 204         msg->ibm_srcstamp = net->ibn_incarnation;
 205         msg->ibm_dstnid   = dstnid;
 206         msg->ibm_dststamp = dststamp;
 207
 208         if (*kiblnd_tunables.kib_cksum) {
 209                 /* NB ibm_cksum zero while computing cksum */
 210                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 211         }
 212 }
 213
 214 int
 215 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 216 {
 217         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 218         __u32     msg_cksum;
 219         __u16     version;
 220         int       msg_nob;
 221         int       flip;
 222
 223         /* 6 bytes are enough to have received magic + version */
 224         if (nob < 6) {
 225                 CERROR("Short message: %d\n", nob);
 226                 return -EPROTO;
 227         }
 228
 229         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 230                 flip = 0;
 231         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 232                 flip = 1;
 233         } else {
 234                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 235                 return -EPROTO;
 236         }
 237
 238         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 239         if (version != IBLND_MSG_VERSION &&
 240             version != IBLND_MSG_VERSION_1) {
 241                 CERROR("Bad version: %x\n", version);
 242                 return -EPROTO;
 243         }
 244
 245         if (nob < hdr_size) {
 246                 CERROR("Short message: %d\n", nob);
 247                 return -EPROTO;
 248         }
 249
 250         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 251         if (msg_nob > nob) {
 252                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 253                 return -EPROTO;
 254         }
 255
 256         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 257          * gets flipped */
 258         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 259         msg->ibm_cksum = 0;
 260         if (msg_cksum != 0 &&
 261             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 262                 CERROR("Bad checksum\n");
 263                 return -EPROTO;
 264         }
 265
 266         msg->ibm_cksum = msg_cksum;
 267
 268         if (flip) {
 269                 /* leave magic unflipped as a clue to peer endianness */
 270                 msg->ibm_version = version;
 271                 CLASSERT (sizeof(msg->ibm_type) == 1);
 272                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 273                 msg->ibm_nob     = msg_nob;
 274                 __swab64s(&msg->ibm_srcnid);
 275                 __swab64s(&msg->ibm_srcstamp);
 276                 __swab64s(&msg->ibm_dstnid);
 277                 __swab64s(&msg->ibm_dststamp);
 278         }
 279
 280         if (msg->ibm_srcnid == LNET_NID_ANY) {
 281                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 282                 return -EPROTO;
 283         }
 284
 285         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 286                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 287                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 288                 return -EPROTO;
 289         }
 290
 291         switch (msg->ibm_type) {
 292         default:
 293                 CERROR("Unknown message type %x\n", msg->ibm_type);
 294                 return -EPROTO;
 295
 296         case IBLND_MSG_NOOP:
 297         case IBLND_MSG_IMMEDIATE:
 298         case IBLND_MSG_PUT_REQ:
 299                 break;
 300
 301         case IBLND_MSG_PUT_ACK:
 302         case IBLND_MSG_GET_REQ:
 303                 if (kiblnd_unpack_rd(msg, flip))
 304                         return -EPROTO;
 305                 break;
 306
 307         case IBLND_MSG_PUT_NAK:
 308         case IBLND_MSG_PUT_DONE:
 309         case IBLND_MSG_GET_DONE:
 310                 if (flip)
 311                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 312                 break;
 313
 314         case IBLND_MSG_CONNREQ:
 315         case IBLND_MSG_CONNACK:
 316                 if (flip) {
 317                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 318                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 319                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 320                 }
 321                 break;
 322         }
 323         return 0;
 324 }
 325
 326 int
 327 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 328 {
 329         kib_peer_t      *peer;
 330         kib_net_t       *net = ni->ni_data;
 331         int             cpt = lnet_cpt_of_nid(nid);
 332         unsigned long   flags;
 333
 334         LASSERT(net != NULL);
 335         LASSERT(nid != LNET_NID_ANY);
 336
 337         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 338         if (peer == NULL) {
 339                 CERROR("Cannot allocate peer\n");
 340                 return -ENOMEM;
 341         }
 342
 343         peer->ibp_ni = ni;
 344         peer->ibp_nid = nid;
 345         peer->ibp_error = 0;
 346         peer->ibp_last_alive = 0;
 347         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 348
 349         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 350         INIT_LIST_HEAD(&peer->ibp_conns);
 351         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 352
 353         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 354
 355         /* always called with a ref on ni, which prevents ni being shutdown */
 356         LASSERT(net->ibn_shutdown == 0);
 357
 358         /* npeers only grows with the global lock held */
 359         atomic_inc(&net->ibn_npeers);
 360
 361         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 362
 363         *peerp = peer;
 364         return 0;
 365 }
 366
 367 void
 368 kiblnd_destroy_peer (kib_peer_t *peer)
 369 {
 370         kib_net_t *net = peer->ibp_ni->ni_data;
 371
 372         LASSERT(net != NULL);
 373         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 374         LASSERT(!kiblnd_peer_active(peer));
 375         LASSERT(peer->ibp_connecting == 0);
 376         LASSERT(peer->ibp_accepting == 0);
 377         LASSERT(list_empty(&peer->ibp_conns));
 378         LASSERT(list_empty(&peer->ibp_tx_queue));
 379
 380         LIBCFS_FREE(peer, sizeof(*peer));
 381
 382         /* NB a peer's connections keep a reference on their peer until
 383          * they are destroyed, so we can be assured that _all_ state to do
 384          * with this peer has been cleaned up when its refcount drops to
 385          * zero. */
 386         atomic_dec(&net->ibn_npeers);
 387 }
 388
 389 kib_peer_t *
 390 kiblnd_find_peer_locked (lnet_nid_t nid)
 391 {
 392         /* the caller is responsible for accounting the additional reference
 393          * that this creates */
 394         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 395         struct list_head        *tmp;
 396         kib_peer_t              *peer;
 397
 398         list_for_each(tmp, peer_list) {
 399
 400                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 401
 402                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 403                         peer->ibp_accepting > 0 ||
 404                         !list_empty(&peer->ibp_conns)); /* active conn */
 405
 406                 if (peer->ibp_nid != nid)
 407                         continue;
 408
 409                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 410                        peer, libcfs_nid2str(nid),
 411                        atomic_read(&peer->ibp_refcount),
 412                        peer->ibp_version);
 413                 return peer;
 414         }
 415         return NULL;
 416 }
 417
 418 void
 419 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 420 {
 421         LASSERT(list_empty(&peer->ibp_conns));
 422
 423         LASSERT (kiblnd_peer_active(peer));
 424         list_del_init(&peer->ibp_list);
 425         /* lose peerlist's ref */
 426         kiblnd_peer_decref(peer);
 427 }
 428
 429 int
 430 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 431                       lnet_nid_t *nidp, int *count)
 432 {
 433         kib_peer_t              *peer;
 434         struct list_head        *ptmp;
 435         int                      i;
 436         unsigned long            flags;
 437
 438         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 439
 440         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 441
 442                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 443
 444                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 445                         LASSERT(peer->ibp_connecting > 0 ||
 446                                 peer->ibp_accepting > 0 ||
 447                                 !list_empty(&peer->ibp_conns));
 448
 449                         if (peer->ibp_ni != ni)
 450                                 continue;
 451
 452                         if (index-- > 0)
 453                                 continue;
 454
 455                         *nidp = peer->ibp_nid;
 456                         *count = atomic_read(&peer->ibp_refcount);
 457
 458                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 459                                                flags);
 460                         return 0;
 461                 }
 462         }
 463
 464         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 465         return -ENOENT;
 466 }
 467
 468 void
 469 kiblnd_del_peer_locked (kib_peer_t *peer)
 470 {
 471         struct list_head        *ctmp;
 472         struct list_head        *cnxt;
 473         kib_conn_t              *conn;
 474
 475         if (list_empty(&peer->ibp_conns)) {
 476                 kiblnd_unlink_peer_locked(peer);
 477         } else {
 478                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 479                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 480
 481                         kiblnd_close_conn_locked(conn, 0);
 482                 }
 483                 /* NB closing peer's last conn unlinked it. */
 484         }
 485         /* NB peer now unlinked; might even be freed if the peer table had the
 486          * last ref on it. */
 487 }
 488
 489 int
 490 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 491 {
 492         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 493         struct list_head        *ptmp;
 494         struct list_head        *pnxt;
 495         kib_peer_t              *peer;
 496         int                     lo;
 497         int                     hi;
 498         int                     i;
 499         unsigned long           flags;
 500         int                     rc = -ENOENT;
 501
 502         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 503
 504         if (nid != LNET_NID_ANY) {
 505                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 506         } else {
 507                 lo = 0;
 508                 hi = kiblnd_data.kib_peer_hash_size - 1;
 509         }
 510
 511         for (i = lo; i <= hi; i++) {
 512                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 513                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 514                         LASSERT(peer->ibp_connecting > 0 ||
 515                                 peer->ibp_accepting > 0 ||
 516                                 !list_empty(&peer->ibp_conns));
 517
 518                         if (peer->ibp_ni != ni)
 519                                 continue;
 520
 521                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 522                                 continue;
 523
 524                         if (!list_empty(&peer->ibp_tx_queue)) {
 525                                 LASSERT(list_empty(&peer->ibp_conns));
 526
 527                                 list_splice_init(&peer->ibp_tx_queue,
 528                                                  &zombies);
 529                         }
 530
 531                         kiblnd_del_peer_locked(peer);
 532                         rc = 0;         /* matched something */
 533                 }
 534         }
 535
 536         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 537
 538         kiblnd_txlist_done(ni, &zombies, -EIO);
 539
 540         return rc;
 541 }
 542
 543 kib_conn_t *
 544 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 545 {
 546         kib_peer_t              *peer;
 547         struct list_head        *ptmp;
 548         kib_conn_t              *conn;
 549         struct list_head        *ctmp;
 550         int                     i;
 551         unsigned long           flags;
 552
 553         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 554
 555         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 556                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 557
 558                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 559                         LASSERT(peer->ibp_connecting > 0 ||
 560                                 peer->ibp_accepting > 0 ||
 561                                 !list_empty(&peer->ibp_conns));
 562
 563                         if (peer->ibp_ni != ni)
 564                                 continue;
 565
 566                         list_for_each(ctmp, &peer->ibp_conns) {
 567                                 if (index-- > 0)
 568                                         continue;
 569
 570                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 571                                 kiblnd_conn_addref(conn);
 572                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 573                                                        flags);
 574                                 return conn;
 575                         }
 576                 }
 577         }
 578
 579         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 580         return NULL;
 581 }
 582
 583 void
 584 kiblnd_debug_rx (kib_rx_t *rx)
 585 {
 586         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 587                rx, rx->rx_status, rx->rx_msg->ibm_type,
 588                rx->rx_msg->ibm_credits);
 589 }
 590
 591 void
 592 kiblnd_debug_tx (kib_tx_t *tx)
 593 {
 594         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 595                "cookie "LPX64" msg %s%s type %x cred %d\n",
 596                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 597                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 598                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 599                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 600                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 601 }
 602
 603 void
 604 kiblnd_debug_conn (kib_conn_t *conn)
 605 {
 606         struct list_head        *tmp;
 607         int                     i;
 608
 609         spin_lock(&conn->ibc_lock);
 610
 611         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 612                atomic_read(&conn->ibc_refcount), conn,
 613                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 614         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 615                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 616                conn->ibc_nsends_posted, conn->ibc_credits,
 617                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 618         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 619
 620         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 621         list_for_each(tmp, &conn->ibc_early_rxs)
 622                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 623
 624         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 625         list_for_each(tmp, &conn->ibc_tx_noops)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 629         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 630                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 631
 632         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 633         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 634                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 635
 636         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 637         list_for_each(tmp, &conn->ibc_tx_queue)
 638                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 639
 640         CDEBUG(D_CONSOLE, "   active_txs:\n");
 641         list_for_each(tmp, &conn->ibc_active_txs)
 642                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 643
 644         CDEBUG(D_CONSOLE, "   rxs:\n");
 645         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 646                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 647
 648         spin_unlock(&conn->ibc_lock);
 649 }
 650
 651 int
 652 kiblnd_translate_mtu(int value)
 653 {
 654         switch (value) {
 655         default:
 656                 return -1;
 657         case 0:
 658                 return 0;
 659         case 256:
 660                 return IB_MTU_256;
 661         case 512:
 662                 return IB_MTU_512;
 663         case 1024:
 664                 return IB_MTU_1024;
 665         case 2048:
 666                 return IB_MTU_2048;
 667         case 4096:
 668                 return IB_MTU_4096;
 669         }
 670 }
 671
 672 static void
 673 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 674 {
 675         int           mtu;
 676
 677         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 678         if (cmid->route.path_rec == NULL)
 679                 return;
 680
 681         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 682         LASSERT (mtu >= 0);
 683         if (mtu != 0)
 684                 cmid->route.path_rec->mtu = mtu;
 685 }
 686
 687 #ifdef HAVE_OFED_IB_COMP_VECTOR
 688 static int
 689 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 690 {
 691         cpumask_t       *mask;
 692         int             vectors;
 693         int             off;
 694         int             i;
 695
 696         vectors = conn->ibc_cmid->device->num_comp_vectors;
 697         if (vectors <= 1)
 698                 return 0;
 699
 700         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 701
 702         /* hash NID to CPU id in this partition... */
 703         off = conn->ibc_peer->ibp_nid % cpus_weight(*mask);
 704         for_each_cpu_mask(i, *mask) {
 705                 if (off-- == 0)
 706                         return i % vectors;
 707         }
 708
 709         LBUG();
 710         return 1;
 711 }
 712 #endif
 713
 714 kib_conn_t *
 715 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 716                    int state, int version)
 717 {
 718         /* CAVEAT EMPTOR:
 719          * If the new conn is created successfully it takes over the caller's
 720          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 721          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 722          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 723          * to destroy 'cmid' here since I'm called from the CM which still has
 724          * its ref on 'cmid'). */
 725         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 726         kib_net_t              *net = peer->ibp_ni->ni_data;
 727         kib_dev_t              *dev;
 728         struct ib_qp_init_attr *init_qp_attr;
 729         struct kib_sched_info   *sched;
 730         kib_conn_t              *conn;
 731         struct ib_cq            *cq;
 732         unsigned long           flags;
 733         int                     cpt;
 734         int                     rc;
 735         int                     i;
 736
 737         LASSERT(net != NULL);
 738         LASSERT(!in_interrupt());
 739
 740         dev = net->ibn_dev;
 741
 742         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 743         sched = kiblnd_data.kib_scheds[cpt];
 744
 745         LASSERT(sched->ibs_nthreads > 0);
 746
 747         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 748                          sizeof(*init_qp_attr));
 749         if (init_qp_attr == NULL) {
 750                 CERROR("Can't allocate qp_attr for %s\n",
 751                        libcfs_nid2str(peer->ibp_nid));
 752                 goto failed_0;
 753         }
 754
 755         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 756         if (conn == NULL) {
 757                 CERROR("Can't allocate connection for %s\n",
 758                        libcfs_nid2str(peer->ibp_nid));
 759                 goto failed_1;
 760         }
 761
 762         conn->ibc_state = IBLND_CONN_INIT;
 763         conn->ibc_version = version;
 764         conn->ibc_peer = peer;                  /* I take the caller's ref */
 765         cmid->context = conn;                   /* for future CM callbacks */
 766         conn->ibc_cmid = cmid;
 767
 768         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 769         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 770         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 771         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 772         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 773         INIT_LIST_HEAD(&conn->ibc_active_txs);
 774         spin_lock_init(&conn->ibc_lock);
 775
 776         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 777                          sizeof(*conn->ibc_connvars));
 778         if (conn->ibc_connvars == NULL) {
 779                 CERROR("Can't allocate in-progress connection state\n");
 780                 goto failed_2;
 781         }
 782
 783         write_lock_irqsave(glock, flags);
 784         if (dev->ibd_failover) {
 785                 write_unlock_irqrestore(glock, flags);
 786                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 787                 goto failed_2;
 788         }
 789
 790         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 791                 /* wakeup failover thread and teardown connection */
 792                 if (kiblnd_dev_can_failover(dev)) {
 793                         list_add_tail(&dev->ibd_fail_list,
 794                                       &kiblnd_data.kib_failed_devs);
 795                         wake_up(&kiblnd_data.kib_failover_waitq);
 796                 }
 797
 798                 write_unlock_irqrestore(glock, flags);
 799                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 800                        cmid->device->name, dev->ibd_ifname);
 801                 goto failed_2;
 802         }
 803
 804         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 805         conn->ibc_hdev = dev->ibd_hdev;
 806
 807         kiblnd_setup_mtu_locked(cmid);
 808
 809         write_unlock_irqrestore(glock, flags);
 810
 811         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 812                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 813         if (conn->ibc_rxs == NULL) {
 814                 CERROR("Cannot allocate RX buffers\n");
 815                 goto failed_2;
 816         }
 817
 818         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 819                                 IBLND_RX_MSG_PAGES(version));
 820         if (rc != 0)
 821                 goto failed_2;
 822
 823         kiblnd_map_rx_descs(conn);
 824
 825 #ifdef HAVE_OFED_IB_COMP_VECTOR
 826         cq = ib_create_cq(cmid->device,
 827                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 828                           IBLND_CQ_ENTRIES(version),
 829                           kiblnd_get_completion_vector(conn, cpt));
 830 #else
 831         cq = ib_create_cq(cmid->device,
 832                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 833                           IBLND_CQ_ENTRIES(version));
 834 #endif
 835         if (IS_ERR(cq)) {
 836                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 837                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 838                 goto failed_2;
 839         }
 840
 841         conn->ibc_cq = cq;
 842
 843         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 844         if (rc != 0) {
 845                 CERROR("Can't request completion notificiation: %d\n", rc);
 846                 goto failed_2;
 847         }
 848
 849         init_qp_attr->event_handler = kiblnd_qp_event;
 850         init_qp_attr->qp_context = conn;
 851         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 852         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 853         init_qp_attr->cap.max_send_sge = 1;
 854         init_qp_attr->cap.max_recv_sge = 1;
 855         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 856         init_qp_attr->qp_type = IB_QPT_RC;
 857         init_qp_attr->send_cq = cq;
 858         init_qp_attr->recv_cq = cq;
 859
 860         conn->ibc_sched = sched;
 861
 862         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 863         if (rc != 0) {
 864                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 865                        rc, init_qp_attr->cap.max_send_wr,
 866                        init_qp_attr->cap.max_recv_wr);
 867                 goto failed_2;
 868         }
 869
 870         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 871
 872         /* 1 ref for caller and each rxmsg */
 873         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 874         conn->ibc_nrx = IBLND_RX_MSGS(version);
 875
 876         /* post receives */
 877         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 878                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 879                                     IBLND_POSTRX_NO_CREDIT);
 880                 if (rc != 0) {
 881                         CERROR("Can't post rxmsg: %d\n", rc);
 882
 883                         /* Make posted receives complete */
 884                         kiblnd_abort_receives(conn);
 885
 886                         /* correct # of posted buffers
 887                          * NB locking needed now I'm racing with completion */
 888                         spin_lock_irqsave(&sched->ibs_lock, flags);
 889                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 890                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 891
 892                         /* cmid will be destroyed by CM(ofed) after cm_callback
 893                          * returned, so we can't refer it anymore
 894                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 895                         rdma_destroy_qp(conn->ibc_cmid);
 896                         conn->ibc_cmid = NULL;
 897
 898                         /* Drop my own and unused rxbuffer refcounts */
 899                         while (i++ <= IBLND_RX_MSGS(version))
 900                                 kiblnd_conn_decref(conn);
 901
 902                         return NULL;
 903                 }
 904         }
 905
 906         /* Init successful! */
 907         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 908                  state == IBLND_CONN_PASSIVE_WAIT);
 909         conn->ibc_state = state;
 910
 911         /* 1 more conn */
 912         atomic_inc(&net->ibn_nconns);
 913         return conn;
 914
 915  failed_2:
 916         kiblnd_destroy_conn(conn);
 917  failed_1:
 918         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 919  failed_0:
 920         return NULL;
 921 }
 922
 923 void
 924 kiblnd_destroy_conn (kib_conn_t *conn)
 925 {
 926         struct rdma_cm_id *cmid = conn->ibc_cmid;
 927         kib_peer_t        *peer = conn->ibc_peer;
 928         int                rc;
 929
 930         LASSERT (!in_interrupt());
 931         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 932         LASSERT(list_empty(&conn->ibc_early_rxs));
 933         LASSERT(list_empty(&conn->ibc_tx_noops));
 934         LASSERT(list_empty(&conn->ibc_tx_queue));
 935         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 936         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 937         LASSERT(list_empty(&conn->ibc_active_txs));
 938         LASSERT (conn->ibc_noops_posted == 0);
 939         LASSERT (conn->ibc_nsends_posted == 0);
 940
 941         switch (conn->ibc_state) {
 942         default:
 943                 /* conn must be completely disengaged from the network */
 944                 LBUG();
 945
 946         case IBLND_CONN_DISCONNECTED:
 947                 /* connvars should have been freed already */
 948                 LASSERT (conn->ibc_connvars == NULL);
 949                 break;
 950
 951         case IBLND_CONN_INIT:
 952                 break;
 953         }
 954
 955         /* conn->ibc_cmid might be destroyed by CM already */
 956         if (cmid != NULL && cmid->qp != NULL)
 957                 rdma_destroy_qp(cmid);
 958
 959         if (conn->ibc_cq != NULL) {
 960                 rc = ib_destroy_cq(conn->ibc_cq);
 961                 if (rc != 0)
 962                         CWARN("Error destroying CQ: %d\n", rc);
 963         }
 964
 965         if (conn->ibc_rx_pages != NULL)
 966                 kiblnd_unmap_rx_descs(conn);
 967
 968         if (conn->ibc_rxs != NULL) {
 969                 LIBCFS_FREE(conn->ibc_rxs,
 970                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 971         }
 972
 973         if (conn->ibc_connvars != NULL)
 974                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 975
 976         if (conn->ibc_hdev != NULL)
 977                 kiblnd_hdev_decref(conn->ibc_hdev);
 978
 979         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 980         if (conn->ibc_state != IBLND_CONN_INIT) {
 981                 kib_net_t *net = peer->ibp_ni->ni_data;
 982
 983                 kiblnd_peer_decref(peer);
 984                 rdma_destroy_id(cmid);
 985                 atomic_dec(&net->ibn_nconns);
 986         }
 987
 988         LIBCFS_FREE(conn, sizeof(*conn));
 989 }
 990
 991 int
 992 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 993 {
 994         kib_conn_t              *conn;
 995         struct list_head        *ctmp;
 996         struct list_head        *cnxt;
 997         int                     count = 0;
 998
 999         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1000                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1001
1002                 CDEBUG(D_NET, "Closing conn -> %s, "
1003                               "version: %x, reason: %d\n",
1004                        libcfs_nid2str(peer->ibp_nid),
1005                        conn->ibc_version, why);
1006
1007                 kiblnd_close_conn_locked(conn, why);
1008                 count++;
1009         }
1010
1011         return count;
1012 }
1013
1014 int
1015 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1016                                 int version, __u64 incarnation)
1017 {
1018         kib_conn_t              *conn;
1019         struct list_head        *ctmp;
1020         struct list_head        *cnxt;
1021         int                     count = 0;
1022
1023         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1024                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1025
1026                 if (conn->ibc_version     == version &&
1027                     conn->ibc_incarnation == incarnation)
1028                         continue;
1029
1030                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1031                               "incarnation:"LPX64"(%x, "LPX64")\n",
1032                        libcfs_nid2str(peer->ibp_nid),
1033                        conn->ibc_version, conn->ibc_incarnation,
1034                        version, incarnation);
1035
1036                 kiblnd_close_conn_locked(conn, -ESTALE);
1037                 count++;
1038         }
1039
1040         return count;
1041 }
1042
1043 int
1044 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1045 {
1046         kib_peer_t              *peer;
1047         struct list_head        *ptmp;
1048         struct list_head        *pnxt;
1049         int                     lo;
1050         int                     hi;
1051         int                     i;
1052         unsigned long           flags;
1053         int                     count = 0;
1054
1055         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1056
1057         if (nid != LNET_NID_ANY)
1058                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1059         else {
1060                 lo = 0;
1061                 hi = kiblnd_data.kib_peer_hash_size - 1;
1062         }
1063
1064         for (i = lo; i <= hi; i++) {
1065                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1066
1067                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1068                         LASSERT(peer->ibp_connecting > 0 ||
1069                                 peer->ibp_accepting > 0 ||
1070                                 !list_empty(&peer->ibp_conns));
1071
1072                         if (peer->ibp_ni != ni)
1073                                 continue;
1074
1075                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1076                                 continue;
1077
1078                         count += kiblnd_close_peer_conns_locked(peer, 0);
1079                 }
1080         }
1081
1082         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1083
1084         /* wildcards always succeed */
1085         if (nid == LNET_NID_ANY)
1086                 return 0;
1087
1088         return (count == 0) ? -ENOENT : 0;
1089 }
1090
1091 int
1092 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1093 {
1094         struct libcfs_ioctl_data *data = arg;
1095         int                       rc = -EINVAL;
1096
1097         switch(cmd) {
1098         case IOC_LIBCFS_GET_PEER: {
1099                 lnet_nid_t   nid = 0;
1100                 int          count = 0;
1101
1102                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1103                                           &nid, &count);
1104                 data->ioc_nid    = nid;
1105                 data->ioc_count  = count;
1106                 break;
1107         }
1108
1109         case IOC_LIBCFS_DEL_PEER: {
1110                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1111                 break;
1112         }
1113         case IOC_LIBCFS_GET_CONN: {
1114                 kib_conn_t *conn;
1115
1116                 rc = 0;
1117                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1118                 if (conn == NULL) {
1119                         rc = -ENOENT;
1120                         break;
1121                 }
1122
1123                 LASSERT (conn->ibc_cmid != NULL);
1124                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1125                 if (conn->ibc_cmid->route.path_rec == NULL)
1126                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1127                 else
1128                         data->ioc_u32[0] =
1129                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1130                 kiblnd_conn_decref(conn);
1131                 break;
1132         }
1133         case IOC_LIBCFS_CLOSE_CONNECTION: {
1134                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1135                 break;
1136         }
1137
1138         default:
1139                 break;
1140         }
1141
1142         return rc;
1143 }
1144
1145 void
1146 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1147 {
1148         cfs_time_t      last_alive = 0;
1149         cfs_time_t      now = cfs_time_current();
1150         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1151         kib_peer_t      *peer;
1152         unsigned long   flags;
1153
1154         read_lock_irqsave(glock, flags);
1155
1156         peer = kiblnd_find_peer_locked(nid);
1157         if (peer != NULL) {
1158                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1159                         peer->ibp_accepting > 0 ||
1160                         !list_empty(&peer->ibp_conns));  /* active conn */
1161                 last_alive = peer->ibp_last_alive;
1162         }
1163
1164         read_unlock_irqrestore(glock, flags);
1165
1166         if (last_alive != 0)
1167                 *when = last_alive;
1168
1169         /* peer is not persistent in hash, trigger peer creation
1170          * and connection establishment with a NULL tx */
1171         if (peer == NULL)
1172                 kiblnd_launch_tx(ni, NULL, nid);
1173
1174         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1175                libcfs_nid2str(nid), peer,
1176                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1177         return;
1178 }
1179
1180 void
1181 kiblnd_free_pages(kib_pages_t *p)
1182 {
1183         int     npages = p->ibp_npages;
1184         int     i;
1185
1186         for (i = 0; i < npages; i++) {
1187                 if (p->ibp_pages[i] != NULL)
1188                         __free_page(p->ibp_pages[i]);
1189         }
1190
1191         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1192 }
1193
1194 int
1195 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1196 {
1197         kib_pages_t     *p;
1198         int             i;
1199
1200         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1201                          offsetof(kib_pages_t, ibp_pages[npages]));
1202         if (p == NULL) {
1203                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1204                 return -ENOMEM;
1205         }
1206
1207         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1208         p->ibp_npages = npages;
1209
1210         for (i = 0; i < npages; i++) {
1211                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1212                                                      GFP_NOFS);
1213                 if (p->ibp_pages[i] == NULL) {
1214                         CERROR("Can't allocate page %d of %d\n", i, npages);
1215                         kiblnd_free_pages(p);
1216                         return -ENOMEM;
1217                 }
1218         }
1219
1220         *pp = p;
1221         return 0;
1222 }
1223
1224 void
1225 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1226 {
1227         kib_rx_t *rx;
1228         int       i;
1229
1230         LASSERT (conn->ibc_rxs != NULL);
1231         LASSERT (conn->ibc_hdev != NULL);
1232
1233         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1234                 rx = &conn->ibc_rxs[i];
1235
1236                 LASSERT (rx->rx_nob >= 0); /* not posted */
1237
1238                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1239                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1240                                                           rx->rx_msgaddr),
1241                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1242         }
1243
1244         kiblnd_free_pages(conn->ibc_rx_pages);
1245
1246         conn->ibc_rx_pages = NULL;
1247 }
1248
1249 void
1250 kiblnd_map_rx_descs(kib_conn_t *conn)
1251 {
1252         kib_rx_t       *rx;
1253         struct page    *pg;
1254         int             pg_off;
1255         int             ipg;
1256         int             i;
1257
1258         for (pg_off = ipg = i = 0;
1259              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1260                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1261                 rx = &conn->ibc_rxs[i];
1262
1263                 rx->rx_conn = conn;
1264                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1265
1266                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1267                                                        rx->rx_msg, IBLND_MSG_SIZE,
1268                                                        DMA_FROM_DEVICE);
1269                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1270                                                    rx->rx_msgaddr));
1271                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1272
1273                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1274                        i, rx->rx_msg, rx->rx_msgaddr,
1275                        lnet_page2phys(pg) + pg_off);
1276
1277                 pg_off += IBLND_MSG_SIZE;
1278                 LASSERT (pg_off <= PAGE_SIZE);
1279
1280                 if (pg_off == PAGE_SIZE) {
1281                         pg_off = 0;
1282                         ipg++;
1283                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1284                 }
1285         }
1286 }
1287
1288 static void
1289 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1290 {
1291         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1292         kib_tx_t       *tx;
1293         int             i;
1294
1295         LASSERT (tpo->tpo_pool.po_allocated == 0);
1296
1297         if (hdev == NULL)
1298                 return;
1299
1300         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1301                 tx = &tpo->tpo_tx_descs[i];
1302                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1303                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1304                                                           tx->tx_msgaddr),
1305                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1306         }
1307
1308         kiblnd_hdev_decref(hdev);
1309         tpo->tpo_hdev = NULL;
1310 }
1311
1312 static kib_hca_dev_t *
1313 kiblnd_current_hdev(kib_dev_t *dev)
1314 {
1315         kib_hca_dev_t *hdev;
1316         unsigned long  flags;
1317         int            i = 0;
1318
1319         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1320         while (dev->ibd_failover) {
1321                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1322                 if (i++ % 50 == 0)
1323                         CDEBUG(D_NET, "%s: Wait for failover\n",
1324                                dev->ibd_ifname);
1325                 schedule_timeout(cfs_time_seconds(1) / 100);
1326
1327                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1328         }
1329
1330         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331         hdev = dev->ibd_hdev;
1332
1333         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1334
1335         return hdev;
1336 }
1337
1338 static void
1339 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1340 {
1341         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1342         kib_pool_t     *pool  = &tpo->tpo_pool;
1343         kib_net_t      *net   = pool->po_owner->ps_net;
1344         kib_dev_t      *dev;
1345         struct page    *page;
1346         kib_tx_t       *tx;
1347         int             page_offset;
1348         int             ipage;
1349         int             i;
1350
1351         LASSERT (net != NULL);
1352
1353         dev = net->ibn_dev;
1354
1355         /* pre-mapped messages are not bigger than 1 page */
1356         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1357
1358         /* No fancy arithmetic when we do the buffer calculations */
1359         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1360
1361         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1362
1363         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1364                 page = txpgs->ibp_pages[ipage];
1365                 tx = &tpo->tpo_tx_descs[i];
1366
1367                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1368                                            page_offset);
1369
1370                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1371                                                        tx->tx_msg,
1372                                                        IBLND_MSG_SIZE,
1373                                                        DMA_TO_DEVICE);
1374                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1375                                                   tx->tx_msgaddr));
1376                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1377
1378                 list_add(&tx->tx_list, &pool->po_free_list);
1379
1380                 page_offset += IBLND_MSG_SIZE;
1381                 LASSERT(page_offset <= PAGE_SIZE);
1382
1383                 if (page_offset == PAGE_SIZE) {
1384                         page_offset = 0;
1385                         ipage++;
1386                         LASSERT(ipage <= txpgs->ibp_npages);
1387                 }
1388         }
1389 }
1390
1391 struct ib_mr *
1392 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1393 {
1394         __u64   index;
1395
1396         LASSERT (hdev->ibh_mrs[0] != NULL);
1397
1398         if (hdev->ibh_nmrs == 1)
1399                 return hdev->ibh_mrs[0];
1400
1401         index = addr >> hdev->ibh_mr_shift;
1402
1403         if (index <  hdev->ibh_nmrs &&
1404             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1405                 return hdev->ibh_mrs[index];
1406
1407         return NULL;
1408 }
1409
1410 struct ib_mr *
1411 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1412 {
1413         struct ib_mr *prev_mr;
1414         struct ib_mr *mr;
1415         int           i;
1416
1417         LASSERT (hdev->ibh_mrs[0] != NULL);
1418
1419         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1420             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1421                 return NULL;
1422
1423         if (hdev->ibh_nmrs == 1)
1424                 return hdev->ibh_mrs[0];
1425
1426         for (i = 0, mr = prev_mr = NULL;
1427              i < rd->rd_nfrags; i++) {
1428                 mr = kiblnd_find_dma_mr(hdev,
1429                                         rd->rd_frags[i].rf_addr,
1430                                         rd->rd_frags[i].rf_nob);
1431                 if (prev_mr == NULL)
1432                         prev_mr = mr;
1433
1434                 if (mr == NULL || prev_mr != mr) {
1435                         /* Can't covered by one single MR */
1436                         mr = NULL;
1437                         break;
1438                 }
1439         }
1440
1441         return mr;
1442 }
1443
1444 void
1445 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1446 {
1447         LASSERT (pool->fpo_map_count == 0);
1448
1449         if (pool->fpo_fmr_pool != NULL)
1450                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1451
1452         if (pool->fpo_hdev != NULL)
1453                 kiblnd_hdev_decref(pool->fpo_hdev);
1454
1455         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1456 }
1457
1458 void
1459 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1460 {
1461         kib_fmr_pool_t *pool;
1462
1463         while (!list_empty(head)) {
1464                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1465                 list_del(&pool->fpo_list);
1466                 kiblnd_destroy_fmr_pool(pool);
1467         }
1468 }
1469
1470 static int kiblnd_fmr_pool_size(int ncpts)
1471 {
1472         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1473
1474         return max(IBLND_FMR_POOL, size);
1475 }
1476
1477 static int kiblnd_fmr_flush_trigger(int ncpts)
1478 {
1479         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1480
1481         return max(IBLND_FMR_POOL_FLUSH, size);
1482 }
1483
1484 int
1485 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1486 {
1487         /* FMR pool for RDMA */
1488         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1489         kib_fmr_pool_t          *fpo;
1490         struct ib_fmr_pool_param param = {
1491                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1492                 .page_shift        = PAGE_SHIFT,
1493                 .access            = (IB_ACCESS_LOCAL_WRITE |
1494                                       IB_ACCESS_REMOTE_WRITE),
1495                 .pool_size         = fps->fps_pool_size,
1496                 .dirty_watermark   = fps->fps_flush_trigger,
1497                 .flush_function    = NULL,
1498                 .flush_arg         = NULL,
1499                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1500         int rc;
1501
1502         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1503         if (fpo == NULL)
1504                 return -ENOMEM;
1505
1506         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1507
1508         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1509         if (IS_ERR(fpo->fpo_fmr_pool)) {
1510                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1511                 CERROR("Failed to create FMR pool: %d\n", rc);
1512
1513                 kiblnd_hdev_decref(fpo->fpo_hdev);
1514                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1515                 return rc;
1516         }
1517
1518         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1519         fpo->fpo_owner    = fps;
1520         *pp_fpo = fpo;
1521
1522         return 0;
1523 }
1524
1525 static void
1526 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1527 {
1528         if (fps->fps_net == NULL) /* intialized? */
1529                 return;
1530
1531         spin_lock(&fps->fps_lock);
1532
1533         while (!list_empty(&fps->fps_pool_list)) {
1534                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1535                                                  kib_fmr_pool_t, fpo_list);
1536                 fpo->fpo_failed = 1;
1537                 list_del(&fpo->fpo_list);
1538                 if (fpo->fpo_map_count == 0)
1539                         list_add(&fpo->fpo_list, zombies);
1540                 else
1541                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1542         }
1543
1544         spin_unlock(&fps->fps_lock);
1545 }
1546
1547 static void
1548 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1549 {
1550         if (fps->fps_net != NULL) { /* initialized? */
1551                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1552                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1553         }
1554 }
1555
1556 static int
1557 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1558                         int pool_size, int flush_trigger)
1559 {
1560         kib_fmr_pool_t *fpo;
1561         int             rc;
1562
1563         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1564
1565         fps->fps_net = net;
1566         fps->fps_cpt = cpt;
1567         fps->fps_pool_size = pool_size;
1568         fps->fps_flush_trigger = flush_trigger;
1569         spin_lock_init(&fps->fps_lock);
1570         INIT_LIST_HEAD(&fps->fps_pool_list);
1571         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1572
1573         rc = kiblnd_create_fmr_pool(fps, &fpo);
1574         if (rc == 0)
1575                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1576
1577         return rc;
1578 }
1579
1580 static int
1581 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1582 {
1583         if (fpo->fpo_map_count != 0) /* still in use */
1584                 return 0;
1585         if (fpo->fpo_failed)
1586                 return 1;
1587         return cfs_time_aftereq(now, fpo->fpo_deadline);
1588 }
1589
1590 void
1591 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1592 {
1593         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1594         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1595         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1596         cfs_time_t         now = cfs_time_current();
1597         kib_fmr_pool_t    *tmp;
1598         int                rc;
1599
1600         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1601         LASSERT(rc == 0);
1602
1603         if (status != 0) {
1604                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1605                 LASSERT(rc == 0);
1606         }
1607
1608         fmr->fmr_pool = NULL;
1609         fmr->fmr_pfmr = NULL;
1610
1611         spin_lock(&fps->fps_lock);
1612         fpo->fpo_map_count--;   /* decref the pool */
1613
1614         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1615                 /* the first pool is persistent */
1616                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1617                         continue;
1618
1619                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1620                         list_move(&fpo->fpo_list, &zombies);
1621                         fps->fps_version++;
1622                 }
1623         }
1624         spin_unlock(&fps->fps_lock);
1625
1626         if (!list_empty(&zombies))
1627                 kiblnd_destroy_fmr_pool_list(&zombies);
1628 }
1629
1630 int
1631 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1632                     __u64 iov, kib_fmr_t *fmr)
1633 {
1634         struct ib_pool_fmr *pfmr;
1635         kib_fmr_pool_t     *fpo;
1636         __u64               version;
1637         int                 rc;
1638
1639 again:
1640         spin_lock(&fps->fps_lock);
1641         version = fps->fps_version;
1642         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1643                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1644                 fpo->fpo_map_count++;
1645                 spin_unlock(&fps->fps_lock);
1646
1647                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1648                                             pages, npages, iov);
1649                 if (likely(!IS_ERR(pfmr))) {
1650                         fmr->fmr_pool = fpo;
1651                         fmr->fmr_pfmr = pfmr;
1652                         return 0;
1653                 }
1654
1655                 spin_lock(&fps->fps_lock);
1656                 fpo->fpo_map_count--;
1657                 if (PTR_ERR(pfmr) != -EAGAIN) {
1658                         spin_unlock(&fps->fps_lock);
1659                         return PTR_ERR(pfmr);
1660                 }
1661
1662                 /* EAGAIN and ... */
1663                 if (version != fps->fps_version) {
1664                         spin_unlock(&fps->fps_lock);
1665                         goto again;
1666                 }
1667         }
1668
1669         if (fps->fps_increasing) {
1670                 spin_unlock(&fps->fps_lock);
1671                 CDEBUG(D_NET, "Another thread is allocating new "
1672                        "FMR pool, waiting for her to complete\n");
1673                 schedule();
1674                 goto again;
1675
1676         }
1677
1678         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1679                 /* someone failed recently */
1680                 spin_unlock(&fps->fps_lock);
1681                 return -EAGAIN;
1682         }
1683
1684         fps->fps_increasing = 1;
1685         spin_unlock(&fps->fps_lock);
1686
1687         CDEBUG(D_NET, "Allocate new FMR pool\n");
1688         rc = kiblnd_create_fmr_pool(fps, &fpo);
1689         spin_lock(&fps->fps_lock);
1690         fps->fps_increasing = 0;
1691         if (rc == 0) {
1692                 fps->fps_version++;
1693                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1694         } else {
1695                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1696         }
1697         spin_unlock(&fps->fps_lock);
1698
1699         goto again;
1700 }
1701
1702 static void
1703 kiblnd_fini_pool(kib_pool_t *pool)
1704 {
1705         LASSERT(list_empty(&pool->po_free_list));
1706         LASSERT(pool->po_allocated == 0);
1707
1708         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1709 }
1710
1711 static void
1712 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1713 {
1714         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1715
1716         memset(pool, 0, sizeof(kib_pool_t));
1717         INIT_LIST_HEAD(&pool->po_free_list);
1718         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1719         pool->po_owner    = ps;
1720         pool->po_size     = size;
1721 }
1722
1723 void
1724 kiblnd_destroy_pool_list(struct list_head *head)
1725 {
1726         kib_pool_t *pool;
1727
1728         while (!list_empty(head)) {
1729                 pool = list_entry(head->next, kib_pool_t, po_list);
1730                 list_del(&pool->po_list);
1731
1732                 LASSERT(pool->po_owner != NULL);
1733                 pool->po_owner->ps_pool_destroy(pool);
1734         }
1735 }
1736
1737 static void
1738 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1739 {
1740         if (ps->ps_net == NULL) /* intialized? */
1741                 return;
1742
1743         spin_lock(&ps->ps_lock);
1744         while (!list_empty(&ps->ps_pool_list)) {
1745                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1746                                             kib_pool_t, po_list);
1747                 po->po_failed = 1;
1748                 list_del(&po->po_list);
1749                 if (po->po_allocated == 0)
1750                         list_add(&po->po_list, zombies);
1751                 else
1752                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1753         }
1754         spin_unlock(&ps->ps_lock);
1755 }
1756
1757 static void
1758 kiblnd_fini_poolset(kib_poolset_t *ps)
1759 {
1760         if (ps->ps_net != NULL) { /* initialized? */
1761                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1762                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1763         }
1764 }
1765
1766 static int
1767 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1768                     kib_net_t *net, char *name, int size,
1769                     kib_ps_pool_create_t po_create,
1770                     kib_ps_pool_destroy_t po_destroy,
1771                     kib_ps_node_init_t nd_init,
1772                     kib_ps_node_fini_t nd_fini)
1773 {
1774         kib_pool_t      *pool;
1775         int             rc;
1776
1777         memset(ps, 0, sizeof(kib_poolset_t));
1778
1779         ps->ps_cpt          = cpt;
1780         ps->ps_net          = net;
1781         ps->ps_pool_create  = po_create;
1782         ps->ps_pool_destroy = po_destroy;
1783         ps->ps_node_init    = nd_init;
1784         ps->ps_node_fini    = nd_fini;
1785         ps->ps_pool_size    = size;
1786         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1787             >= sizeof(ps->ps_name))
1788                 return -E2BIG;
1789         spin_lock_init(&ps->ps_lock);
1790         INIT_LIST_HEAD(&ps->ps_pool_list);
1791         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1792
1793         rc = ps->ps_pool_create(ps, size, &pool);
1794         if (rc == 0)
1795                 list_add(&pool->po_list, &ps->ps_pool_list);
1796         else
1797                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1798
1799         return rc;
1800 }
1801
1802 static int
1803 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1804 {
1805         if (pool->po_allocated != 0) /* still in use */
1806                 return 0;
1807         if (pool->po_failed)
1808                 return 1;
1809         return cfs_time_aftereq(now, pool->po_deadline);
1810 }
1811
1812 void
1813 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1814 {
1815         struct list_head zombies = LIST_HEAD_INIT(zombies);
1816         kib_poolset_t   *ps = pool->po_owner;
1817         kib_pool_t      *tmp;
1818         cfs_time_t       now = cfs_time_current();
1819
1820         spin_lock(&ps->ps_lock);
1821
1822         if (ps->ps_node_fini != NULL)
1823                 ps->ps_node_fini(pool, node);
1824
1825         LASSERT(pool->po_allocated > 0);
1826         list_add(node, &pool->po_free_list);
1827         pool->po_allocated--;
1828
1829         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1830                 /* the first pool is persistent */
1831                 if (ps->ps_pool_list.next == &pool->po_list)
1832                         continue;
1833
1834                 if (kiblnd_pool_is_idle(pool, now))
1835                         list_move(&pool->po_list, &zombies);
1836         }
1837         spin_unlock(&ps->ps_lock);
1838
1839         if (!list_empty(&zombies))
1840                 kiblnd_destroy_pool_list(&zombies);
1841 }
1842
1843 struct list_head *
1844 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1845 {
1846         struct list_head        *node;
1847         kib_pool_t              *pool;
1848         int                     rc;
1849
1850 again:
1851         spin_lock(&ps->ps_lock);
1852         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1853                 if (list_empty(&pool->po_free_list))
1854                         continue;
1855
1856                 pool->po_allocated++;
1857                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1858                 node = pool->po_free_list.next;
1859                 list_del(node);
1860
1861                 if (ps->ps_node_init != NULL) {
1862                         /* still hold the lock */
1863                         ps->ps_node_init(pool, node);
1864                 }
1865                 spin_unlock(&ps->ps_lock);
1866                 return node;
1867         }
1868
1869         /* no available tx pool and ... */
1870         if (ps->ps_increasing) {
1871                 /* another thread is allocating a new pool */
1872                 spin_unlock(&ps->ps_lock);
1873                 CDEBUG(D_NET, "Another thread is allocating new "
1874                        "%s pool, waiting for her to complete\n",
1875                        ps->ps_name);
1876                 schedule();
1877                 goto again;
1878         }
1879
1880         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1881                 /* someone failed recently */
1882                 spin_unlock(&ps->ps_lock);
1883                 return NULL;
1884         }
1885
1886         ps->ps_increasing = 1;
1887         spin_unlock(&ps->ps_lock);
1888
1889         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1890
1891         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1892
1893         spin_lock(&ps->ps_lock);
1894         ps->ps_increasing = 0;
1895         if (rc == 0) {
1896                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1897         } else {
1898                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1899                 CERROR("Can't allocate new %s pool because out of memory\n",
1900                        ps->ps_name);
1901         }
1902         spin_unlock(&ps->ps_lock);
1903
1904         goto again;
1905 }
1906
1907 void
1908 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1909 {
1910         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1911         struct ib_mr        *mr  = pmr->pmr_mr;
1912
1913         pmr->pmr_mr = NULL;
1914         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1915         if (mr != NULL)
1916                 ib_dereg_mr(mr);
1917 }
1918
1919 int
1920 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1921                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1922 {
1923         kib_phys_mr_t    *pmr;
1924         struct list_head *node;
1925         int               rc;
1926         int               i;
1927
1928         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1929         if (node == NULL) {
1930                 CERROR("Failed to allocate PMR descriptor\n");
1931                 return -ENOMEM;
1932         }
1933
1934         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1935         if (pmr->pmr_pool->ppo_hdev != hdev) {
1936                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1937                 return -EAGAIN;
1938         }
1939
1940         for (i = 0; i < rd->rd_nfrags; i ++) {
1941                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1942                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1943         }
1944
1945         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1946                                      pmr->pmr_ipb, rd->rd_nfrags,
1947                                      IB_ACCESS_LOCAL_WRITE |
1948                                      IB_ACCESS_REMOTE_WRITE,
1949                                      iova);
1950         if (!IS_ERR(pmr->pmr_mr)) {
1951                 pmr->pmr_iova = *iova;
1952                 *pp_pmr = pmr;
1953                 return 0;
1954         }
1955
1956         rc = PTR_ERR(pmr->pmr_mr);
1957         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1958
1959         pmr->pmr_mr = NULL;
1960         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1961
1962         return rc;
1963 }
1964
1965 static void
1966 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1967 {
1968         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1969         kib_phys_mr_t  *pmr;
1970
1971         LASSERT (pool->po_allocated == 0);
1972
1973         while (!list_empty(&pool->po_free_list)) {
1974                 pmr = list_entry(pool->po_free_list.next,
1975                                      kib_phys_mr_t, pmr_list);
1976
1977                 LASSERT (pmr->pmr_mr == NULL);
1978                 list_del(&pmr->pmr_list);
1979
1980                 if (pmr->pmr_ipb != NULL) {
1981                         LIBCFS_FREE(pmr->pmr_ipb,
1982                                     IBLND_MAX_RDMA_FRAGS *
1983                                     sizeof(struct ib_phys_buf));
1984                 }
1985
1986                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1987         }
1988
1989         kiblnd_fini_pool(pool);
1990         if (ppo->ppo_hdev != NULL)
1991                 kiblnd_hdev_decref(ppo->ppo_hdev);
1992
1993         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1994 }
1995
1996 static inline int kiblnd_pmr_pool_size(int ncpts)
1997 {
1998         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1999
2000         return max(IBLND_PMR_POOL, size);
2001 }
2002
2003 static int
2004 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2005 {
2006         struct kib_pmr_pool     *ppo;
2007         struct kib_pool         *pool;
2008         kib_phys_mr_t           *pmr;
2009         int                     i;
2010
2011         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2012                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2013         if (ppo == NULL) {
2014                 CERROR("Failed to allocate PMR pool\n");
2015                 return -ENOMEM;
2016         }
2017
2018         pool = &ppo->ppo_pool;
2019         kiblnd_init_pool(ps, pool, size);
2020
2021         for (i = 0; i < size; i++) {
2022                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2023                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2024                 if (pmr == NULL)
2025                         break;
2026
2027                 pmr->pmr_pool = ppo;
2028                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2029                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2030                 if (pmr->pmr_ipb == NULL)
2031                         break;
2032
2033                 list_add(&pmr->pmr_list, &pool->po_free_list);
2034         }
2035
2036         if (i < size) {
2037                 ps->ps_pool_destroy(pool);
2038                 return -ENOMEM;
2039         }
2040
2041         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2042         *pp_po = pool;
2043         return 0;
2044 }
2045
2046 static void
2047 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2048 {
2049         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2050         int             i;
2051
2052         LASSERT (pool->po_allocated == 0);
2053
2054         if (tpo->tpo_tx_pages != NULL) {
2055                 kiblnd_unmap_tx_pool(tpo);
2056                 kiblnd_free_pages(tpo->tpo_tx_pages);
2057         }
2058
2059         if (tpo->tpo_tx_descs == NULL)
2060                 goto out;
2061
2062         for (i = 0; i < pool->po_size; i++) {
2063                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2064
2065                 list_del(&tx->tx_list);
2066                 if (tx->tx_pages != NULL)
2067                         LIBCFS_FREE(tx->tx_pages,
2068                                     LNET_MAX_IOV *
2069                                     sizeof(*tx->tx_pages));
2070                 if (tx->tx_frags != NULL)
2071                         LIBCFS_FREE(tx->tx_frags,
2072                                     IBLND_MAX_RDMA_FRAGS *
2073                                             sizeof(*tx->tx_frags));
2074                 if (tx->tx_wrq != NULL)
2075                         LIBCFS_FREE(tx->tx_wrq,
2076                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2077                                     sizeof(*tx->tx_wrq));
2078                 if (tx->tx_sge != NULL)
2079                         LIBCFS_FREE(tx->tx_sge,
2080                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2081                                     sizeof(*tx->tx_sge));
2082                 if (tx->tx_rd != NULL)
2083                         LIBCFS_FREE(tx->tx_rd,
2084                                     offsetof(kib_rdma_desc_t,
2085                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2086         }
2087
2088         LIBCFS_FREE(tpo->tpo_tx_descs,
2089                     pool->po_size * sizeof(kib_tx_t));
2090 out:
2091         kiblnd_fini_pool(pool);
2092         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2093 }
2094
2095 static int kiblnd_tx_pool_size(int ncpts)
2096 {
2097         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2098
2099         return max(IBLND_TX_POOL, ntx);
2100 }
2101
2102 static int
2103 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2104 {
2105         int            i;
2106         int            npg;
2107         kib_pool_t    *pool;
2108         kib_tx_pool_t *tpo;
2109
2110         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2111         if (tpo == NULL) {
2112                 CERROR("Failed to allocate TX pool\n");
2113                 return -ENOMEM;
2114         }
2115
2116         pool = &tpo->tpo_pool;
2117         kiblnd_init_pool(ps, pool, size);
2118         tpo->tpo_tx_descs = NULL;
2119         tpo->tpo_tx_pages = NULL;
2120
2121         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2122         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2123                 CERROR("Can't allocate tx pages: %d\n", npg);
2124                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2125                 return -ENOMEM;
2126         }
2127
2128         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2129                          size * sizeof(kib_tx_t));
2130         if (tpo->tpo_tx_descs == NULL) {
2131                 CERROR("Can't allocate %d tx descriptors\n", size);
2132                 ps->ps_pool_destroy(pool);
2133                 return -ENOMEM;
2134         }
2135
2136         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2137
2138         for (i = 0; i < size; i++) {
2139                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2140
2141                 tx->tx_pool = tpo;
2142                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2143                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2144                                          lnet_cpt_table(), ps->ps_cpt,
2145                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2146                         if (tx->tx_pages == NULL)
2147                                 break;
2148                 }
2149
2150                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2151                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2152                 if (tx->tx_frags == NULL)
2153                         break;
2154
2155                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2156
2157                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2158                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2159                                  sizeof(*tx->tx_wrq));
2160                 if (tx->tx_wrq == NULL)
2161                         break;
2162
2163                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2164                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2165                                  sizeof(*tx->tx_sge));
2166                 if (tx->tx_sge == NULL)
2167                         break;
2168
2169                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2170                                  offsetof(kib_rdma_desc_t,
2171                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2172                 if (tx->tx_rd == NULL)
2173                         break;
2174         }
2175
2176         if (i == size) {
2177                 kiblnd_map_tx_pool(tpo);
2178                 *pp_po = pool;
2179                 return 0;
2180         }
2181
2182         ps->ps_pool_destroy(pool);
2183         return -ENOMEM;
2184 }
2185
2186 static void
2187 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2188 {
2189         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2190                                              tps_poolset);
2191         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2192
2193         tx->tx_cookie = tps->tps_next_tx_cookie++;
2194 }
2195
2196 void
2197 kiblnd_net_fini_pools(kib_net_t *net)
2198 {
2199         int     i;
2200
2201         cfs_cpt_for_each(i, lnet_cpt_table()) {
2202                 kib_tx_poolset_t        *tps;
2203                 kib_fmr_poolset_t       *fps;
2204                 kib_pmr_poolset_t       *pps;
2205
2206                 if (net->ibn_tx_ps != NULL) {
2207                         tps = net->ibn_tx_ps[i];
2208                         kiblnd_fini_poolset(&tps->tps_poolset);
2209                 }
2210
2211                 if (net->ibn_fmr_ps != NULL) {
2212                         fps = net->ibn_fmr_ps[i];
2213                         kiblnd_fini_fmr_poolset(fps);
2214                 }
2215
2216                 if (net->ibn_pmr_ps != NULL) {
2217                         pps = net->ibn_pmr_ps[i];
2218                         kiblnd_fini_poolset(&pps->pps_poolset);
2219                 }
2220         }
2221
2222         if (net->ibn_tx_ps != NULL) {
2223                 cfs_percpt_free(net->ibn_tx_ps);
2224                 net->ibn_tx_ps = NULL;
2225         }
2226
2227         if (net->ibn_fmr_ps != NULL) {
2228                 cfs_percpt_free(net->ibn_fmr_ps);
2229                 net->ibn_fmr_ps = NULL;
2230         }
2231
2232         if (net->ibn_pmr_ps != NULL) {
2233                 cfs_percpt_free(net->ibn_pmr_ps);
2234                 net->ibn_pmr_ps = NULL;
2235         }
2236 }
2237
2238 int
2239 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2240 {
2241         unsigned long   flags;
2242         int             cpt;
2243         int             rc;
2244         int             i;
2245
2246         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2247         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2248             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2249                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2250                                            flags);
2251                 goto create_tx_pool;
2252         }
2253
2254         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2255
2256         if (*kiblnd_tunables.kib_fmr_pool_size <
2257             *kiblnd_tunables.kib_ntx / 4) {
2258                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2259                        *kiblnd_tunables.kib_fmr_pool_size,
2260                        *kiblnd_tunables.kib_ntx / 4);
2261                 rc = -EINVAL;
2262                 goto failed;
2263         }
2264
2265         /* TX pool must be created later than FMR/PMR, see LU-2268
2266          * for details */
2267         LASSERT(net->ibn_tx_ps == NULL);
2268
2269         /* premapping can fail if ibd_nmr > 1, so we always create
2270          * FMR/PMR pool and map-on-demand if premapping failed */
2271
2272         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2273                                            sizeof(kib_fmr_poolset_t));
2274         if (net->ibn_fmr_ps == NULL) {
2275                 CERROR("Failed to allocate FMR pool array\n");
2276                 rc = -ENOMEM;
2277                 goto failed;
2278         }
2279
2280         for (i = 0; i < ncpts; i++) {
2281                 cpt = (cpts == NULL) ? i : cpts[i];
2282                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2283                                              kiblnd_fmr_pool_size(ncpts),
2284                                              kiblnd_fmr_flush_trigger(ncpts));
2285                 if (rc == -ENOSYS && i == 0) /* no FMR */
2286                         break; /* create PMR pool */
2287
2288                 if (rc != 0) { /* a real error */
2289                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2290                                cpt, rc);
2291                         goto failed;
2292                 }
2293         }
2294
2295         if (i > 0) {
2296                 LASSERT(i == ncpts);
2297                 goto create_tx_pool;
2298         }
2299
2300         cfs_percpt_free(net->ibn_fmr_ps);
2301         net->ibn_fmr_ps = NULL;
2302
2303         CWARN("Device does not support FMR, failing back to PMR\n");
2304
2305         if (*kiblnd_tunables.kib_pmr_pool_size <
2306             *kiblnd_tunables.kib_ntx / 4) {
2307                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2308                        *kiblnd_tunables.kib_pmr_pool_size,
2309                        *kiblnd_tunables.kib_ntx / 4);
2310                 rc = -EINVAL;
2311                 goto failed;
2312         }
2313
2314         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2315                                            sizeof(kib_pmr_poolset_t));
2316         if (net->ibn_pmr_ps == NULL) {
2317                 CERROR("Failed to allocate PMR pool array\n");
2318                 rc = -ENOMEM;
2319                 goto failed;
2320         }
2321
2322         for (i = 0; i < ncpts; i++) {
2323                 cpt = (cpts == NULL) ? i : cpts[i];
2324                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2325                                          cpt, net, "PMR",
2326                                          kiblnd_pmr_pool_size(ncpts),
2327                                          kiblnd_create_pmr_pool,
2328                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2329                 if (rc != 0) {
2330                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2331                                cpt, rc);
2332                         goto failed;
2333                 }
2334         }
2335
2336  create_tx_pool:
2337         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2338                                           sizeof(kib_tx_poolset_t));
2339         if (net->ibn_tx_ps == NULL) {
2340                 CERROR("Failed to allocate tx pool array\n");
2341                 rc = -ENOMEM;
2342                 goto failed;
2343         }
2344
2345         for (i = 0; i < ncpts; i++) {
2346                 cpt = (cpts == NULL) ? i : cpts[i];
2347                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2348                                          cpt, net, "TX",
2349                                          kiblnd_tx_pool_size(ncpts),
2350                                          kiblnd_create_tx_pool,
2351                                          kiblnd_destroy_tx_pool,
2352                                          kiblnd_tx_init, NULL);
2353                 if (rc != 0) {
2354                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2355                                cpt, rc);
2356                         goto failed;
2357                 }
2358         }
2359
2360         return 0;
2361  failed:
2362         kiblnd_net_fini_pools(net);
2363         LASSERT(rc != 0);
2364         return rc;
2365 }
2366
2367 static int
2368 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2369 {
2370         struct ib_device_attr *attr;
2371         int                    rc;
2372
2373         /* It's safe to assume a HCA can handle a page size
2374          * matching that of the native system */
2375         hdev->ibh_page_shift = PAGE_SHIFT;
2376         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2377         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2378
2379         LIBCFS_ALLOC(attr, sizeof(*attr));
2380         if (attr == NULL) {
2381                 CERROR("Out of memory\n");
2382                 return -ENOMEM;
2383         }
2384
2385         rc = ib_query_device(hdev->ibh_ibdev, attr);
2386         if (rc == 0)
2387                 hdev->ibh_mr_size = attr->max_mr_size;
2388
2389         LIBCFS_FREE(attr, sizeof(*attr));
2390
2391         if (rc != 0) {
2392                 CERROR("Failed to query IB device: %d\n", rc);
2393                 return rc;
2394         }
2395
2396         if (hdev->ibh_mr_size == ~0ULL) {
2397                 hdev->ibh_mr_shift = 64;
2398                 return 0;
2399         }
2400
2401         for (hdev->ibh_mr_shift = 0;
2402              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2403                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2404                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2405                         return 0;
2406         }
2407
2408         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2409         return -EINVAL;
2410 }
2411
2412 void
2413 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2414 {
2415         int     i;
2416
2417         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2418                 return;
2419
2420         for (i = 0; i < hdev->ibh_nmrs; i++) {
2421                 if (hdev->ibh_mrs[i] == NULL)
2422                         break;
2423
2424                 ib_dereg_mr(hdev->ibh_mrs[i]);
2425         }
2426
2427         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2428         hdev->ibh_mrs  = NULL;
2429         hdev->ibh_nmrs = 0;
2430 }
2431
2432 void
2433 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2434 {
2435         kiblnd_hdev_cleanup_mrs(hdev);
2436
2437         if (hdev->ibh_pd != NULL)
2438                 ib_dealloc_pd(hdev->ibh_pd);
2439
2440         if (hdev->ibh_cmid != NULL)
2441                 rdma_destroy_id(hdev->ibh_cmid);
2442
2443         LIBCFS_FREE(hdev, sizeof(*hdev));
2444 }
2445
2446 int
2447 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2448 {
2449         struct ib_mr *mr;
2450         int           i;
2451         int           rc;
2452         __u64         mm_size;
2453         __u64         mr_size;
2454         int           acflags = IB_ACCESS_LOCAL_WRITE |
2455                                 IB_ACCESS_REMOTE_WRITE;
2456
2457         rc = kiblnd_hdev_get_attr(hdev);
2458         if (rc != 0)
2459                 return rc;
2460
2461         if (hdev->ibh_mr_shift == 64) {
2462                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2463                 if (hdev->ibh_mrs == NULL) {
2464                         CERROR("Failed to allocate MRs table\n");
2465                         return -ENOMEM;
2466                 }
2467
2468                 hdev->ibh_mrs[0] = NULL;
2469                 hdev->ibh_nmrs   = 1;
2470
2471                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2472                 if (IS_ERR(mr)) {
2473                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2474                         kiblnd_hdev_cleanup_mrs(hdev);
2475                         return PTR_ERR(mr);
2476                 }
2477
2478                 hdev->ibh_mrs[0] = mr;
2479
2480                 goto out;
2481         }
2482
2483         mr_size = (1ULL << hdev->ibh_mr_shift);
2484         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2485
2486         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2487
2488         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2489                 /* it's 4T..., assume we will re-code at that time */
2490                 CERROR("Can't support memory size: x"LPX64
2491                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2492                 return -EINVAL;
2493         }
2494
2495         /* create an array of MRs to cover all memory */
2496         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2497         if (hdev->ibh_mrs == NULL) {
2498                 CERROR("Failed to allocate MRs' table\n");
2499                 return -ENOMEM;
2500         }
2501
2502         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2503
2504         for (i = 0; i < hdev->ibh_nmrs; i++) {
2505                 struct ib_phys_buf ipb;
2506                 __u64              iova;
2507
2508                 ipb.size = hdev->ibh_mr_size;
2509                 ipb.addr = i * mr_size;
2510                 iova     = ipb.addr;
2511
2512                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2513                 if (IS_ERR(mr)) {
2514                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2515                                " size "LPX64" : %ld\n",
2516                                ipb.addr, ipb.size, PTR_ERR(mr));
2517                         kiblnd_hdev_cleanup_mrs(hdev);
2518                         return PTR_ERR(mr);
2519                 }
2520
2521                 LASSERT (iova == ipb.addr);
2522
2523                 hdev->ibh_mrs[i] = mr;
2524         }
2525
2526 out:
2527         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2528                 LCONSOLE_INFO("Register global MR array, MR size: "
2529                               LPX64", array size: %d\n",
2530                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2531         return 0;
2532 }
2533
2534 static int
2535 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2536 {       /* DUMMY */
2537         return 0;
2538 }
2539
2540 static int
2541 kiblnd_dev_need_failover(kib_dev_t *dev)
2542 {
2543         struct rdma_cm_id  *cmid;
2544         struct sockaddr_in  srcaddr;
2545         struct sockaddr_in  dstaddr;
2546         int                 rc;
2547
2548         if (dev->ibd_hdev == NULL || /* initializing */
2549             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2550             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2551                 return 1;
2552
2553         /* XXX: it's UGLY, but I don't have better way to find
2554          * ib-bonding HCA failover because:
2555          *
2556          * a. no reliable CM event for HCA failover...
2557          * b. no OFED API to get ib_device for current net_device...
2558          *
2559          * We have only two choices at this point:
2560          *
2561          * a. rdma_bind_addr(), it will conflict with listener cmid
2562          * b. rdma_resolve_addr() to zero addr */
2563         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2564                                      IB_QPT_RC);
2565         if (IS_ERR(cmid)) {
2566                 rc = PTR_ERR(cmid);
2567                 CERROR("Failed to create cmid for failover: %d\n", rc);
2568                 return rc;
2569         }
2570
2571         memset(&srcaddr, 0, sizeof(srcaddr));
2572         srcaddr.sin_family      = AF_INET;
2573         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2574
2575         memset(&dstaddr, 0, sizeof(dstaddr));
2576         dstaddr.sin_family = AF_INET;
2577         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2578                                (struct sockaddr *)&dstaddr, 1);
2579         if (rc != 0 || cmid->device == NULL) {
2580                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2581                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2582                        cmid->device, rc);
2583                 rdma_destroy_id(cmid);
2584                 return rc;
2585         }
2586
2587         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2588                 /* don't need device failover */
2589                 rdma_destroy_id(cmid);
2590                 return 0;
2591         }
2592
2593         return 1;
2594 }
2595
2596 int
2597 kiblnd_dev_failover(kib_dev_t *dev)
2598 {
2599         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2600         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2601         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2602         struct rdma_cm_id  *cmid  = NULL;
2603         kib_hca_dev_t      *hdev  = NULL;
2604         kib_hca_dev_t      *old;
2605         struct ib_pd       *pd;
2606         kib_net_t          *net;
2607         struct sockaddr_in  addr;
2608         unsigned long       flags;
2609         int                 rc = 0;
2610         int                 i;
2611
2612         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2613                  dev->ibd_can_failover ||
2614                  dev->ibd_hdev == NULL);
2615
2616         rc = kiblnd_dev_need_failover(dev);
2617         if (rc <= 0)
2618                 goto out;
2619
2620         if (dev->ibd_hdev != NULL &&
2621             dev->ibd_hdev->ibh_cmid != NULL) {
2622                 /* XXX it's not good to close old listener at here,
2623                  * because we can fail to create new listener.
2624                  * But we have to close it now, otherwise rdma_bind_addr
2625                  * will return EADDRINUSE... How crap! */
2626                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2627
2628                 cmid = dev->ibd_hdev->ibh_cmid;
2629                 /* make next schedule of kiblnd_dev_need_failover()
2630                  * return 1 for me */
2631                 dev->ibd_hdev->ibh_cmid  = NULL;
2632                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2633
2634                 rdma_destroy_id(cmid);
2635         }
2636
2637         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2638                                      IB_QPT_RC);
2639         if (IS_ERR(cmid)) {
2640                 rc = PTR_ERR(cmid);
2641                 CERROR("Failed to create cmid for failover: %d\n", rc);
2642                 goto out;
2643         }
2644
2645         memset(&addr, 0, sizeof(addr));
2646         addr.sin_family      = AF_INET;
2647         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2648         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2649
2650         /* Bind to failover device or port */
2651         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2652         if (rc != 0 || cmid->device == NULL) {
2653                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2654                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2655                        cmid->device, rc);
2656                 rdma_destroy_id(cmid);
2657                 goto out;
2658         }
2659
2660         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2661         if (hdev == NULL) {
2662                 CERROR("Failed to allocate kib_hca_dev\n");
2663                 rdma_destroy_id(cmid);
2664                 rc = -ENOMEM;
2665                 goto out;
2666         }
2667
2668         atomic_set(&hdev->ibh_ref, 1);
2669         hdev->ibh_dev   = dev;
2670         hdev->ibh_cmid  = cmid;
2671         hdev->ibh_ibdev = cmid->device;
2672
2673         pd = ib_alloc_pd(cmid->device);
2674         if (IS_ERR(pd)) {
2675                 rc = PTR_ERR(pd);
2676                 CERROR("Can't allocate PD: %d\n", rc);
2677                 goto out;
2678         }
2679
2680         hdev->ibh_pd = pd;
2681
2682         rc = rdma_listen(cmid, 0);
2683         if (rc != 0) {
2684                 CERROR("Can't start new listener: %d\n", rc);
2685                 goto out;
2686         }
2687
2688         rc = kiblnd_hdev_setup_mrs(hdev);
2689         if (rc != 0) {
2690                 CERROR("Can't setup device: %d\n", rc);
2691                 goto out;
2692         }
2693
2694         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2695
2696         old = dev->ibd_hdev;
2697         dev->ibd_hdev = hdev;   /* take over the refcount */
2698         hdev = old;
2699
2700         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2701                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2702                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2703                                             &zombie_tpo);
2704
2705                         if (net->ibn_fmr_ps != NULL) {
2706                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2707                                                         &zombie_fpo);
2708
2709                         } else if (net->ibn_pmr_ps != NULL) {
2710                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2711                                                     pps_poolset, &zombie_ppo);
2712                         }
2713                 }
2714         }
2715
2716         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2717  out:
2718         if (!list_empty(&zombie_tpo))
2719                 kiblnd_destroy_pool_list(&zombie_tpo);
2720         if (!list_empty(&zombie_ppo))
2721                 kiblnd_destroy_pool_list(&zombie_ppo);
2722         if (!list_empty(&zombie_fpo))
2723                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2724         if (hdev != NULL)
2725                 kiblnd_hdev_decref(hdev);
2726
2727         if (rc != 0)
2728                 dev->ibd_failed_failover++;
2729         else
2730                 dev->ibd_failed_failover = 0;
2731
2732         return rc;
2733 }
2734
2735 void
2736 kiblnd_destroy_dev (kib_dev_t *dev)
2737 {
2738         LASSERT (dev->ibd_nnets == 0);
2739         LASSERT(list_empty(&dev->ibd_nets));
2740
2741         list_del(&dev->ibd_fail_list);
2742         list_del(&dev->ibd_list);
2743
2744         if (dev->ibd_hdev != NULL)
2745                 kiblnd_hdev_decref(dev->ibd_hdev);
2746
2747         LIBCFS_FREE(dev, sizeof(*dev));
2748 }
2749
2750 kib_dev_t *
2751 kiblnd_create_dev(char *ifname)
2752 {
2753         struct net_device *netdev;
2754         kib_dev_t         *dev;
2755         __u32              netmask;
2756         __u32              ip;
2757         int                up;
2758         int                rc;
2759
2760         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2761         if (rc != 0) {
2762                 CERROR("Can't query IPoIB interface %s: %d\n",
2763                        ifname, rc);
2764                 return NULL;
2765         }
2766
2767         if (!up) {
2768                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2769                 return NULL;
2770         }
2771
2772         LIBCFS_ALLOC(dev, sizeof(*dev));
2773         if (dev == NULL)
2774                 return NULL;
2775
2776         memset(dev, 0, sizeof(*dev));
2777 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2778         netdev = dev_get_by_name(&init_net, ifname);
2779 #else
2780         netdev = dev_get_by_name(ifname);
2781 #endif
2782         if (netdev == NULL) {
2783                 dev->ibd_can_failover = 0;
2784         } else {
2785                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2786                 dev_put(netdev);
2787         }
2788
2789         INIT_LIST_HEAD(&dev->ibd_nets);
2790         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2791         INIT_LIST_HEAD(&dev->ibd_fail_list);
2792         dev->ibd_ifip = ip;
2793         strcpy(&dev->ibd_ifname[0], ifname);
2794
2795         /* initialize the device */
2796         rc = kiblnd_dev_failover(dev);
2797         if (rc != 0) {
2798                 CERROR("Can't initialize device: %d\n", rc);
2799                 LIBCFS_FREE(dev, sizeof(*dev));
2800                 return NULL;
2801         }
2802
2803         list_add_tail(&dev->ibd_list,
2804                           &kiblnd_data.kib_devs);
2805         return dev;
2806 }
2807
2808 void
2809 kiblnd_base_shutdown(void)
2810 {
2811         struct kib_sched_info   *sched;
2812         int                     i;
2813
2814         LASSERT(list_empty(&kiblnd_data.kib_devs));
2815
2816         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2817                atomic_read(&libcfs_kmemory));
2818
2819         switch (kiblnd_data.kib_init) {
2820         default:
2821                 LBUG();
2822
2823         case IBLND_INIT_ALL:
2824         case IBLND_INIT_DATA:
2825                 LASSERT (kiblnd_data.kib_peers != NULL);
2826                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2827                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2828                 }
2829                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2830                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2831
2832                 /* flag threads to terminate; wake and wait for them to die */
2833                 kiblnd_data.kib_shutdown = 1;
2834
2835                 /* NB: we really want to stop scheduler threads net by net
2836                  * instead of the whole module, this should be improved
2837                  * with dynamic configuration LNet */
2838                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2839                         wake_up_all(&sched->ibs_waitq);
2840
2841                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2842                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2843
2844                 i = 2;
2845                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2846                         i++;
2847                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2848                                "Waiting for %d threads to terminate\n",
2849                                atomic_read(&kiblnd_data.kib_nthreads));
2850                         cfs_pause(cfs_time_seconds(1));
2851                 }
2852
2853                 /* fall through */
2854
2855         case IBLND_INIT_NOTHING:
2856                 break;
2857         }
2858
2859         if (kiblnd_data.kib_peers != NULL) {
2860                 LIBCFS_FREE(kiblnd_data.kib_peers,
2861                             sizeof(struct list_head) *
2862                             kiblnd_data.kib_peer_hash_size);
2863         }
2864
2865         if (kiblnd_data.kib_scheds != NULL)
2866                 cfs_percpt_free(kiblnd_data.kib_scheds);
2867
2868         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2869                atomic_read(&libcfs_kmemory));
2870
2871         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2872         module_put(THIS_MODULE);
2873 }
2874
2875 void
2876 kiblnd_shutdown (lnet_ni_t *ni)
2877 {
2878         kib_net_t        *net = ni->ni_data;
2879         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2880         int               i;
2881         unsigned long     flags;
2882
2883         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2884
2885         if (net == NULL)
2886                 goto out;
2887
2888         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2889                atomic_read(&libcfs_kmemory));
2890
2891         write_lock_irqsave(g_lock, flags);
2892         net->ibn_shutdown = 1;
2893         write_unlock_irqrestore(g_lock, flags);
2894
2895         switch (net->ibn_init) {
2896         default:
2897                 LBUG();
2898
2899         case IBLND_INIT_ALL:
2900                 /* nuke all existing peers within this net */
2901                 kiblnd_del_peer(ni, LNET_NID_ANY);
2902
2903                 /* Wait for all peer state to clean up */
2904                 i = 2;
2905                 while (atomic_read(&net->ibn_npeers) != 0) {
2906                         i++;
2907                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2908                                "%s: waiting for %d peers to disconnect\n",
2909                                libcfs_nid2str(ni->ni_nid),
2910                                atomic_read(&net->ibn_npeers));
2911                         cfs_pause(cfs_time_seconds(1));
2912                 }
2913
2914                 kiblnd_net_fini_pools(net);
2915
2916                 write_lock_irqsave(g_lock, flags);
2917                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2918                 net->ibn_dev->ibd_nnets--;
2919                 list_del(&net->ibn_list);
2920                 write_unlock_irqrestore(g_lock, flags);
2921
2922                 /* fall through */
2923
2924         case IBLND_INIT_NOTHING:
2925                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2926
2927                 if (net->ibn_dev != NULL &&
2928                     net->ibn_dev->ibd_nnets == 0)
2929                         kiblnd_destroy_dev(net->ibn_dev);
2930
2931                 break;
2932         }
2933
2934         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2935                atomic_read(&libcfs_kmemory));
2936
2937         net->ibn_init = IBLND_INIT_NOTHING;
2938         ni->ni_data = NULL;
2939
2940         LIBCFS_FREE(net, sizeof(*net));
2941
2942 out:
2943         if (list_empty(&kiblnd_data.kib_devs))
2944                 kiblnd_base_shutdown();
2945         return;
2946 }
2947
2948 int
2949 kiblnd_base_startup(void)
2950 {
2951         struct kib_sched_info   *sched;
2952         int                     rc;
2953         int                     i;
2954
2955         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2956
2957         try_module_get(THIS_MODULE);
2958         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2959
2960         rwlock_init(&kiblnd_data.kib_global_lock);
2961
2962         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2963         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2964
2965         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2966         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2967                      sizeof(struct list_head) *
2968                      kiblnd_data.kib_peer_hash_size);
2969         if (kiblnd_data.kib_peers == NULL)
2970                 goto failed;
2971
2972         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2973                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2974
2975         spin_lock_init(&kiblnd_data.kib_connd_lock);
2976         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2977         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2978         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2979         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2980
2981         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2982                                                   sizeof(*sched));
2983         if (kiblnd_data.kib_scheds == NULL)
2984                 goto failed;
2985
2986         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2987                 int     nthrs;
2988
2989                 spin_lock_init(&sched->ibs_lock);
2990                 INIT_LIST_HEAD(&sched->ibs_conns);
2991                 init_waitqueue_head(&sched->ibs_waitq);
2992
2993                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2994                 if (*kiblnd_tunables.kib_nscheds > 0) {
2995                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2996                 } else {
2997                         /* max to half of CPUs, another half is reserved for
2998                          * upper layer modules */
2999                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3000                 }
3001
3002                 sched->ibs_nthreads_max = nthrs;
3003                 sched->ibs_cpt = i;
3004         }
3005
3006         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3007
3008         /* lists/ptrs/locks initialised */
3009         kiblnd_data.kib_init = IBLND_INIT_DATA;
3010         /*****************************************************/
3011
3012         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3013         if (rc != 0) {
3014                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3015                 goto failed;
3016         }
3017
3018         if (*kiblnd_tunables.kib_dev_failover != 0)
3019                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3020                                          "kiblnd_failover");
3021
3022         if (rc != 0) {
3023                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3024                 goto failed;
3025         }
3026
3027         /* flag everything initialised */
3028         kiblnd_data.kib_init = IBLND_INIT_ALL;
3029         /*****************************************************/
3030
3031         return 0;
3032
3033  failed:
3034         kiblnd_base_shutdown();
3035         return -ENETDOWN;
3036 }
3037
3038 int
3039 kiblnd_start_schedulers(struct kib_sched_info *sched)
3040 {
3041         int     rc = 0;
3042         int     nthrs;
3043         int     i;
3044
3045         if (sched->ibs_nthreads == 0) {
3046                 if (*kiblnd_tunables.kib_nscheds > 0) {
3047                         nthrs = sched->ibs_nthreads_max;
3048                 } else {
3049                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3050                                                sched->ibs_cpt);
3051                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3052                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3053                 }
3054         } else {
3055                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3056                 /* increase one thread if there is new interface */
3057                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3058         }
3059
3060         for (i = 0; i < nthrs; i++) {
3061                 long    id;
3062                 char    name[20];
3063                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3064                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3065                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3066                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3067                 if (rc == 0)
3068                         continue;
3069
3070                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3071                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3072                 break;
3073         }
3074
3075         sched->ibs_nthreads += i;
3076         return rc;
3077 }
3078
3079 int
3080 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3081 {
3082         int     cpt;
3083         int     rc;
3084         int     i;
3085
3086         for (i = 0; i < ncpts; i++) {
3087                 struct kib_sched_info *sched;
3088
3089                 cpt = (cpts == NULL) ? i : cpts[i];
3090                 sched = kiblnd_data.kib_scheds[cpt];
3091
3092                 if (!newdev && sched->ibs_nthreads > 0)
3093                         continue;
3094
3095                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3096                 if (rc != 0) {
3097                         CERROR("Failed to start scheduler threads for %s\n",
3098                                dev->ibd_ifname);
3099                         return rc;
3100                 }
3101         }
3102         return 0;
3103 }
3104
3105 kib_dev_t *
3106 kiblnd_dev_search(char *ifname)
3107 {
3108         kib_dev_t       *alias = NULL;
3109         kib_dev_t       *dev;
3110         char            *colon;
3111         char            *colon2;
3112
3113         colon = strchr(ifname, ':');
3114         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3115                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3116                         return dev;
3117
3118                 if (alias != NULL)
3119                         continue;
3120
3121                 colon2 = strchr(dev->ibd_ifname, ':');
3122                 if (colon != NULL)
3123                         *colon = 0;
3124                 if (colon2 != NULL)
3125                         *colon2 = 0;
3126
3127                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3128                         alias = dev;
3129
3130                 if (colon != NULL)
3131                         *colon = ':';
3132                 if (colon2 != NULL)
3133                         *colon2 = ':';
3134         }
3135         return alias;
3136 }
3137
3138 int
3139 kiblnd_startup (lnet_ni_t *ni)
3140 {
3141         char                     *ifname;
3142         kib_dev_t                *ibdev = NULL;
3143         kib_net_t                *net;
3144         struct timeval            tv;
3145         unsigned long             flags;
3146         int                       rc;
3147         int                       newdev;
3148
3149         LASSERT (ni->ni_lnd == &the_o2iblnd);
3150
3151         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3152                 rc = kiblnd_base_startup();
3153                 if (rc != 0)
3154                         return rc;
3155         }
3156
3157         LIBCFS_ALLOC(net, sizeof(*net));
3158         ni->ni_data = net;
3159         if (net == NULL)
3160                 goto failed;
3161
3162         memset(net, 0, sizeof(*net));
3163
3164         do_gettimeofday(&tv);
3165         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3166
3167         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3168         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3169         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3170         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3171
3172         if (ni->ni_interfaces[0] != NULL) {
3173                 /* Use the IPoIB interface specified in 'networks=' */
3174
3175                 CLASSERT (LNET_MAX_INTERFACES > 1);
3176                 if (ni->ni_interfaces[1] != NULL) {
3177                         CERROR("Multiple interfaces not supported\n");
3178                         goto failed;
3179                 }
3180
3181                 ifname = ni->ni_interfaces[0];
3182         } else {
3183                 ifname = *kiblnd_tunables.kib_default_ipif;
3184         }
3185
3186         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3187                 CERROR("IPoIB interface name too long: %s\n", ifname);
3188                 goto failed;
3189         }
3190
3191         ibdev = kiblnd_dev_search(ifname);
3192
3193         newdev = ibdev == NULL;
3194         /* hmm...create kib_dev even for alias */
3195         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3196                 ibdev = kiblnd_create_dev(ifname);
3197
3198         if (ibdev == NULL)
3199                 goto failed;
3200
3201         net->ibn_dev = ibdev;
3202         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3203
3204         rc = kiblnd_dev_start_threads(ibdev, newdev,
3205                                       ni->ni_cpts, ni->ni_ncpts);
3206         if (rc != 0)
3207                 goto failed;
3208
3209         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3210         if (rc != 0) {
3211                 CERROR("Failed to initialize NI pools: %d\n", rc);
3212                 goto failed;
3213         }
3214
3215         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3216         ibdev->ibd_nnets++;
3217         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3218         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3219
3220         net->ibn_init = IBLND_INIT_ALL;
3221
3222         return 0;
3223
3224 failed:
3225         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3226                 kiblnd_destroy_dev(ibdev);
3227
3228         kiblnd_shutdown(ni);
3229
3230         CDEBUG(D_NET, "kiblnd_startup failed\n");
3231         return -ENETDOWN;
3232 }
3233
3234 void __exit
3235 kiblnd_module_fini (void)
3236 {
3237         lnet_unregister_lnd(&the_o2iblnd);
3238         kiblnd_tunables_fini();
3239 }
3240
3241 int __init
3242 kiblnd_module_init (void)
3243 {
3244         int    rc;
3245
3246         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3247         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3248                   <= IBLND_MSG_SIZE);
3249         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3250                   <= IBLND_MSG_SIZE);
3251
3252         rc = kiblnd_tunables_init();
3253         if (rc != 0)
3254                 return rc;
3255
3256         lnet_register_lnd(&the_o2iblnd);
3257
3258         return 0;
3259 }
3260
3261 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3262 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3263 MODULE_LICENSE("GPL");
3264
3265 module_init(kiblnd_module_init);
3266 module_exit(kiblnd_module_fini);