drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.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         memset(peer, 0, sizeof(*peer));  /* zero flags etc */
 344
 345         peer->ibp_ni = ni;
 346         peer->ibp_nid = nid;
 347         peer->ibp_error = 0;
 348         peer->ibp_last_alive = 0;
 349         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 350
 351         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 352         INIT_LIST_HEAD(&peer->ibp_conns);
 353         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 354
 355         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 356
 357         /* always called with a ref on ni, which prevents ni being shutdown */
 358         LASSERT (net->ibn_shutdown == 0);
 359
 360         /* npeers only grows with the global lock held */
 361         atomic_inc(&net->ibn_npeers);
 362
 363         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 364
 365         *peerp = peer;
 366         return 0;
 367 }
 368
 369 void
 370 kiblnd_destroy_peer (kib_peer_t *peer)
 371 {
 372         kib_net_t *net = peer->ibp_ni->ni_data;
 373
 374         LASSERT (net != NULL);
 375         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 376         LASSERT (!kiblnd_peer_active(peer));
 377         LASSERT (peer->ibp_connecting == 0);
 378         LASSERT (peer->ibp_accepting == 0);
 379         LASSERT (list_empty(&peer->ibp_conns));
 380         LASSERT (list_empty(&peer->ibp_tx_queue));
 381
 382         LIBCFS_FREE(peer, sizeof(*peer));
 383
 384         /* NB a peer's connections keep a reference on their peer until
 385          * they are destroyed, so we can be assured that _all_ state to do
 386          * with this peer has been cleaned up when its refcount drops to
 387          * zero. */
 388         atomic_dec(&net->ibn_npeers);
 389 }
 390
 391 kib_peer_t *
 392 kiblnd_find_peer_locked (lnet_nid_t nid)
 393 {
 394         /* the caller is responsible for accounting the additional reference
 395          * that this creates */
 396         struct list_head       *peer_list = kiblnd_nid2peerlist(nid);
 397         struct list_head       *tmp;
 398         kib_peer_t       *peer;
 399
 400         list_for_each (tmp, peer_list) {
 401
 402                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 403
 404                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 405                          peer->ibp_accepting > 0 ||
 406                          !list_empty(&peer->ibp_conns));  /* active conn */
 407
 408                 if (peer->ibp_nid != nid)
 409                         continue;
 410
 411                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 412                        peer, libcfs_nid2str(nid),
 413                        atomic_read(&peer->ibp_refcount),
 414                        peer->ibp_version);
 415                 return peer;
 416         }
 417         return NULL;
 418 }
 419
 420 void
 421 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 422 {
 423         LASSERT (list_empty(&peer->ibp_conns));
 424
 425         LASSERT (kiblnd_peer_active(peer));
 426         list_del_init(&peer->ibp_list);
 427         /* lose peerlist's ref */
 428         kiblnd_peer_decref(peer);
 429 }
 430
 431 int
 432 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 433                       lnet_nid_t *nidp, int *count)
 434 {
 435         kib_peer_t          *peer;
 436         struct list_head            *ptmp;
 437         int                 i;
 438         unsigned long     flags;
 439
 440         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 441
 442         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 443
 444                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 445
 446                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 447                         LASSERT (peer->ibp_connecting > 0 ||
 448                                  peer->ibp_accepting > 0 ||
 449                                  !list_empty(&peer->ibp_conns));
 450
 451                         if (peer->ibp_ni != ni)
 452                                 continue;
 453
 454                         if (index-- > 0)
 455                                 continue;
 456
 457                         *nidp = peer->ibp_nid;
 458                         *count = atomic_read(&peer->ibp_refcount);
 459
 460                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 461                                                flags);
 462                         return 0;
 463                 }
 464         }
 465
 466         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 467         return -ENOENT;
 468 }
 469
 470 void
 471 kiblnd_del_peer_locked (kib_peer_t *peer)
 472 {
 473         struct list_head           *ctmp;
 474         struct list_head           *cnxt;
 475         kib_conn_t         *conn;
 476
 477         if (list_empty(&peer->ibp_conns)) {
 478                 kiblnd_unlink_peer_locked(peer);
 479         } else {
 480                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 481                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 482
 483                         kiblnd_close_conn_locked(conn, 0);
 484                 }
 485                 /* NB closing peer's last conn unlinked it. */
 486         }
 487         /* NB peer now unlinked; might even be freed if the peer table had the
 488          * last ref on it. */
 489 }
 490
 491 int
 492 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 493 {
 494         LIST_HEAD        (zombies);
 495         struct list_head            *ptmp;
 496         struct list_head            *pnxt;
 497         kib_peer_t          *peer;
 498         int                 lo;
 499         int                 hi;
 500         int                 i;
 501         unsigned long     flags;
 502         int                 rc = -ENOENT;
 503
 504         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 505
 506         if (nid != LNET_NID_ANY) {
 507                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 508         } else {
 509                 lo = 0;
 510                 hi = kiblnd_data.kib_peer_hash_size - 1;
 511         }
 512
 513         for (i = lo; i <= hi; i++) {
 514                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 515                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 516                         LASSERT (peer->ibp_connecting > 0 ||
 517                                  peer->ibp_accepting > 0 ||
 518                                  !list_empty(&peer->ibp_conns));
 519
 520                         if (peer->ibp_ni != ni)
 521                                 continue;
 522
 523                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 524                                 continue;
 525
 526                         if (!list_empty(&peer->ibp_tx_queue)) {
 527                                 LASSERT (list_empty(&peer->ibp_conns));
 528
 529                                 list_splice_init(&peer->ibp_tx_queue,
 530                                                      &zombies);
 531                         }
 532
 533                         kiblnd_del_peer_locked(peer);
 534                         rc = 0;  /* matched something */
 535                 }
 536         }
 537
 538         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 539
 540         kiblnd_txlist_done(ni, &zombies, -EIO);
 541
 542         return rc;
 543 }
 544
 545 kib_conn_t *
 546 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 547 {
 548         kib_peer_t          *peer;
 549         struct list_head            *ptmp;
 550         kib_conn_t          *conn;
 551         struct list_head            *ctmp;
 552         int                 i;
 553         unsigned long     flags;
 554
 555         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 556
 557         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 558                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 559
 560                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 561                         LASSERT (peer->ibp_connecting > 0 ||
 562                                  peer->ibp_accepting > 0 ||
 563                                  !list_empty(&peer->ibp_conns));
 564
 565                         if (peer->ibp_ni != ni)
 566                                 continue;
 567
 568                         list_for_each (ctmp, &peer->ibp_conns) {
 569                                 if (index-- > 0)
 570                                         continue;
 571
 572                                 conn = list_entry(ctmp, kib_conn_t,
 573                                                       ibc_list);
 574                                 kiblnd_conn_addref(conn);
 575                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 576                                                        flags);
 577                                 return conn;
 578                         }
 579                 }
 580         }
 581
 582         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 583         return NULL;
 584 }
 585
 586 void
 587 kiblnd_debug_rx (kib_rx_t *rx)
 588 {
 589         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 590                rx, rx->rx_status, rx->rx_msg->ibm_type,
 591                rx->rx_msg->ibm_credits);
 592 }
 593
 594 void
 595 kiblnd_debug_tx (kib_tx_t *tx)
 596 {
 597         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 598                "cookie "LPX64" msg %s%s type %x cred %d\n",
 599                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 600                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 601                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 602                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 603                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 604 }
 605
 606 void
 607 kiblnd_debug_conn (kib_conn_t *conn)
 608 {
 609         struct list_head        *tmp;
 610         int             i;
 611
 612         spin_lock(&conn->ibc_lock);
 613
 614         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 615                atomic_read(&conn->ibc_refcount), conn,
 616                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 617         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 618                conn->ibc_state, conn->ibc_noops_posted,
 619                conn->ibc_nsends_posted, conn->ibc_credits,
 620                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 621         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 622
 623         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 624         list_for_each(tmp, &conn->ibc_early_rxs)
 625                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 626
 627         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 628         list_for_each(tmp, &conn->ibc_tx_noops)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 632         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 633                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 634
 635         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 636         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 637                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 638
 639         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 640         list_for_each(tmp, &conn->ibc_tx_queue)
 641                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 642
 643         CDEBUG(D_CONSOLE, "   active_txs:\n");
 644         list_for_each(tmp, &conn->ibc_active_txs)
 645                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 646
 647         CDEBUG(D_CONSOLE, "   rxs:\n");
 648         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 649                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 650
 651         spin_unlock(&conn->ibc_lock);
 652 }
 653
 654 int
 655 kiblnd_translate_mtu(int value)
 656 {
 657         switch (value) {
 658         default:
 659                 return -1;
 660         case 0:
 661                 return 0;
 662         case 256:
 663                 return IB_MTU_256;
 664         case 512:
 665                 return IB_MTU_512;
 666         case 1024:
 667                 return IB_MTU_1024;
 668         case 2048:
 669                 return IB_MTU_2048;
 670         case 4096:
 671                 return IB_MTU_4096;
 672         }
 673 }
 674
 675 static void
 676 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 677 {
 678         int        mtu;
 679
 680         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 681         if (cmid->route.path_rec == NULL)
 682                 return;
 683
 684         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 685         LASSERT (mtu >= 0);
 686         if (mtu != 0)
 687                 cmid->route.path_rec->mtu = mtu;
 688 }
 689
 690 static int
 691 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 692 {
 693         cpumask_t       *mask;
 694         int             vectors;
 695         int             off;
 696         int             i;
 697
 698         vectors = conn->ibc_cmid->device->num_comp_vectors;
 699         if (vectors <= 1)
 700                 return 0;
 701
 702         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 703
 704         /* hash NID to CPU id in this partition... */
 705         off = conn->ibc_peer->ibp_nid % cpus_weight(*mask);
 706         for_each_cpu_mask(i, *mask) {
 707                 if (off-- == 0)
 708                         return i % vectors;
 709         }
 710
 711         LBUG();
 712         return 1;
 713 }
 714
 715 kib_conn_t *
 716 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 717                    int state, int version)
 718 {
 719         /* CAVEAT EMPTOR:
 720          * If the new conn is created successfully it takes over the caller's
 721          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 722          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 723          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 724          * to destroy 'cmid' here since I'm called from the CM which still has
 725          * its ref on 'cmid'). */
 726         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
 727         kib_net_t             *net = peer->ibp_ni->ni_data;
 728         kib_dev_t             *dev;
 729         struct ib_qp_init_attr *init_qp_attr;
 730         struct kib_sched_info   *sched;
 731         kib_conn_t              *conn;
 732         struct ib_cq            *cq;
 733         unsigned long           flags;
 734         int                     cpt;
 735         int                     rc;
 736         int                     i;
 737
 738         LASSERT(net != NULL);
 739         LASSERT(!in_interrupt());
 740
 741         dev = net->ibn_dev;
 742
 743         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 744         sched = kiblnd_data.kib_scheds[cpt];
 745
 746         LASSERT(sched->ibs_nthreads > 0);
 747
 748         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 749                          sizeof(*init_qp_attr));
 750         if (init_qp_attr == NULL) {
 751                 CERROR("Can't allocate qp_attr for %s\n",
 752                        libcfs_nid2str(peer->ibp_nid));
 753                 goto failed_0;
 754         }
 755
 756         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 757         if (conn == NULL) {
 758                 CERROR("Can't allocate connection for %s\n",
 759                        libcfs_nid2str(peer->ibp_nid));
 760                 goto failed_1;
 761         }
 762
 763         conn->ibc_state = IBLND_CONN_INIT;
 764         conn->ibc_version = version;
 765         conn->ibc_peer = peer;            /* I take the caller's ref */
 766         cmid->context = conn;              /* for future CM callbacks */
 767         conn->ibc_cmid = cmid;
 768
 769         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 770         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 771         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 772         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 773         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 774         INIT_LIST_HEAD(&conn->ibc_active_txs);
 775         spin_lock_init(&conn->ibc_lock);
 776
 777         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 778                          sizeof(*conn->ibc_connvars));
 779         if (conn->ibc_connvars == NULL) {
 780                 CERROR("Can't allocate in-progress connection state\n");
 781                 goto failed_2;
 782         }
 783
 784         write_lock_irqsave(glock, flags);
 785         if (dev->ibd_failover) {
 786                 write_unlock_irqrestore(glock, flags);
 787                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 788                 goto failed_2;
 789         }
 790
 791         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 792                 /* wakeup failover thread and teardown connection */
 793                 if (kiblnd_dev_can_failover(dev)) {
 794                         list_add_tail(&dev->ibd_fail_list,
 795                                       &kiblnd_data.kib_failed_devs);
 796                         wake_up(&kiblnd_data.kib_failover_waitq);
 797                 }
 798
 799                 write_unlock_irqrestore(glock, flags);
 800                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 801                        cmid->device->name, dev->ibd_ifname);
 802                 goto failed_2;
 803         }
 804
 805         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 806         conn->ibc_hdev = dev->ibd_hdev;
 807
 808         kiblnd_setup_mtu_locked(cmid);
 809
 810         write_unlock_irqrestore(glock, flags);
 811
 812         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 813                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 814         if (conn->ibc_rxs == NULL) {
 815                 CERROR("Cannot allocate RX buffers\n");
 816                 goto failed_2;
 817         }
 818
 819         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 820                                 IBLND_RX_MSG_PAGES(version));
 821         if (rc != 0)
 822                 goto failed_2;
 823
 824         kiblnd_map_rx_descs(conn);
 825
 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         if (IS_ERR(cq)) {
 831                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 832                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 833                 goto failed_2;
 834         }
 835
 836         conn->ibc_cq = cq;
 837
 838         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 839         if (rc != 0) {
 840                 CERROR("Can't request completion notificiation: %d\n", rc);
 841                 goto failed_2;
 842         }
 843
 844         init_qp_attr->event_handler = kiblnd_qp_event;
 845         init_qp_attr->qp_context = conn;
 846         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 847         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 848         init_qp_attr->cap.max_send_sge = 1;
 849         init_qp_attr->cap.max_recv_sge = 1;
 850         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 851         init_qp_attr->qp_type = IB_QPT_RC;
 852         init_qp_attr->send_cq = cq;
 853         init_qp_attr->recv_cq = cq;
 854
 855         conn->ibc_sched = sched;
 856
 857         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 858         if (rc != 0) {
 859                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 860                        rc, init_qp_attr->cap.max_send_wr,
 861                        init_qp_attr->cap.max_recv_wr);
 862                 goto failed_2;
 863         }
 864
 865         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 866
 867         /* 1 ref for caller and each rxmsg */
 868         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 869         conn->ibc_nrx = IBLND_RX_MSGS(version);
 870
 871         /* post receives */
 872         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 873                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 874                                     IBLND_POSTRX_NO_CREDIT);
 875                 if (rc != 0) {
 876                         CERROR("Can't post rxmsg: %d\n", rc);
 877
 878                         /* Make posted receives complete */
 879                         kiblnd_abort_receives(conn);
 880
 881                         /* correct # of posted buffers
 882                          * NB locking needed now I'm racing with completion */
 883                         spin_lock_irqsave(&sched->ibs_lock, flags);
 884                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 885                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 886
 887                         /* cmid will be destroyed by CM(ofed) after cm_callback
 888                          * returned, so we can't refer it anymore
 889                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 890                         rdma_destroy_qp(conn->ibc_cmid);
 891                         conn->ibc_cmid = NULL;
 892
 893                         /* Drop my own and unused rxbuffer refcounts */
 894                         while (i++ <= IBLND_RX_MSGS(version))
 895                                 kiblnd_conn_decref(conn);
 896
 897                         return NULL;
 898                 }
 899         }
 900
 901         /* Init successful! */
 902         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 903                  state == IBLND_CONN_PASSIVE_WAIT);
 904         conn->ibc_state = state;
 905
 906         /* 1 more conn */
 907         atomic_inc(&net->ibn_nconns);
 908         return conn;
 909
 910  failed_2:
 911         kiblnd_destroy_conn(conn);
 912  failed_1:
 913         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 914  failed_0:
 915         return NULL;
 916 }
 917
 918 void
 919 kiblnd_destroy_conn (kib_conn_t *conn)
 920 {
 921         struct rdma_cm_id *cmid = conn->ibc_cmid;
 922         kib_peer_t      *peer = conn->ibc_peer;
 923         int             rc;
 924
 925         LASSERT (!in_interrupt());
 926         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 927         LASSERT (list_empty(&conn->ibc_early_rxs));
 928         LASSERT (list_empty(&conn->ibc_tx_noops));
 929         LASSERT (list_empty(&conn->ibc_tx_queue));
 930         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 931         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 932         LASSERT (list_empty(&conn->ibc_active_txs));
 933         LASSERT (conn->ibc_noops_posted == 0);
 934         LASSERT (conn->ibc_nsends_posted == 0);
 935
 936         switch (conn->ibc_state) {
 937         default:
 938                 /* conn must be completely disengaged from the network */
 939                 LBUG();
 940
 941         case IBLND_CONN_DISCONNECTED:
 942                 /* connvars should have been freed already */
 943                 LASSERT (conn->ibc_connvars == NULL);
 944                 break;
 945
 946         case IBLND_CONN_INIT:
 947                 break;
 948         }
 949
 950         /* conn->ibc_cmid might be destroyed by CM already */
 951         if (cmid != NULL && cmid->qp != NULL)
 952                 rdma_destroy_qp(cmid);
 953
 954         if (conn->ibc_cq != NULL) {
 955                 rc = ib_destroy_cq(conn->ibc_cq);
 956                 if (rc != 0)
 957                         CWARN("Error destroying CQ: %d\n", rc);
 958         }
 959
 960         if (conn->ibc_rx_pages != NULL)
 961                 kiblnd_unmap_rx_descs(conn);
 962
 963         if (conn->ibc_rxs != NULL) {
 964                 LIBCFS_FREE(conn->ibc_rxs,
 965                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 966         }
 967
 968         if (conn->ibc_connvars != NULL)
 969                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 970
 971         if (conn->ibc_hdev != NULL)
 972                 kiblnd_hdev_decref(conn->ibc_hdev);
 973
 974         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 975         if (conn->ibc_state != IBLND_CONN_INIT) {
 976                 kib_net_t *net = peer->ibp_ni->ni_data;
 977
 978                 kiblnd_peer_decref(peer);
 979                 rdma_destroy_id(cmid);
 980                 atomic_dec(&net->ibn_nconns);
 981         }
 982
 983         LIBCFS_FREE(conn, sizeof(*conn));
 984 }
 985
 986 int
 987 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 988 {
 989         kib_conn_t           *conn;
 990         struct list_head             *ctmp;
 991         struct list_head             *cnxt;
 992         int                  count = 0;
 993
 994         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 995                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 996
 997                 CDEBUG(D_NET, "Closing conn -> %s, "
 998                               "version: %x, reason: %d\n",
 999                        libcfs_nid2str(peer->ibp_nid),
1000                        conn->ibc_version, why);
1001
1002                 kiblnd_close_conn_locked(conn, why);
1003                 count++;
1004         }
1005
1006         return count;
1007 }
1008
1009 int
1010 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1011                                  int version, __u64 incarnation)
1012 {
1013         kib_conn_t           *conn;
1014         struct list_head             *ctmp;
1015         struct list_head             *cnxt;
1016         int                  count = 0;
1017
1018         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1019                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1020
1021                 if (conn->ibc_version     == version &&
1022                     conn->ibc_incarnation == incarnation)
1023                         continue;
1024
1025                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1026                               "incarnation:"LPX64"(%x, "LPX64")\n",
1027                        libcfs_nid2str(peer->ibp_nid),
1028                        conn->ibc_version, conn->ibc_incarnation,
1029                        version, incarnation);
1030
1031                 kiblnd_close_conn_locked(conn, -ESTALE);
1032                 count++;
1033         }
1034
1035         return count;
1036 }
1037
1038 int
1039 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1040 {
1041         kib_peer_t           *peer;
1042         struct list_head             *ptmp;
1043         struct list_head             *pnxt;
1044         int                  lo;
1045         int                  hi;
1046         int                  i;
1047         unsigned long      flags;
1048         int                  count = 0;
1049
1050         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1051
1052         if (nid != LNET_NID_ANY)
1053                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1054         else {
1055                 lo = 0;
1056                 hi = kiblnd_data.kib_peer_hash_size - 1;
1057         }
1058
1059         for (i = lo; i <= hi; i++) {
1060                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1061
1062                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1063                         LASSERT (peer->ibp_connecting > 0 ||
1064                                  peer->ibp_accepting > 0 ||
1065                                  !list_empty(&peer->ibp_conns));
1066
1067                         if (peer->ibp_ni != ni)
1068                                 continue;
1069
1070                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1071                                 continue;
1072
1073                         count += kiblnd_close_peer_conns_locked(peer, 0);
1074                 }
1075         }
1076
1077         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1078
1079         /* wildcards always succeed */
1080         if (nid == LNET_NID_ANY)
1081                 return 0;
1082
1083         return (count == 0) ? -ENOENT : 0;
1084 }
1085
1086 int
1087 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1088 {
1089         struct libcfs_ioctl_data *data = arg;
1090         int                    rc = -EINVAL;
1091
1092         switch(cmd) {
1093         case IOC_LIBCFS_GET_PEER: {
1094                 lnet_nid_t   nid = 0;
1095                 int       count = 0;
1096
1097                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1098                                           &nid, &count);
1099                 data->ioc_nid    = nid;
1100                 data->ioc_count  = count;
1101                 break;
1102         }
1103
1104         case IOC_LIBCFS_DEL_PEER: {
1105                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1106                 break;
1107         }
1108         case IOC_LIBCFS_GET_CONN: {
1109                 kib_conn_t *conn;
1110
1111                 rc = 0;
1112                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1113                 if (conn == NULL) {
1114                         rc = -ENOENT;
1115                         break;
1116                 }
1117
1118                 LASSERT (conn->ibc_cmid != NULL);
1119                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1120                 if (conn->ibc_cmid->route.path_rec == NULL)
1121                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1122                 else
1123                         data->ioc_u32[0] =
1124                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1125                 kiblnd_conn_decref(conn);
1126                 break;
1127         }
1128         case IOC_LIBCFS_CLOSE_CONNECTION: {
1129                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1130                 break;
1131         }
1132
1133         default:
1134                 break;
1135         }
1136
1137         return rc;
1138 }
1139
1140 void
1141 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1142 {
1143         cfs_time_t      last_alive = 0;
1144         cfs_time_t      now = cfs_time_current();
1145         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1146         kib_peer_t      *peer;
1147         unsigned long   flags;
1148
1149         read_lock_irqsave(glock, flags);
1150
1151         peer = kiblnd_find_peer_locked(nid);
1152         if (peer != NULL) {
1153                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1154                          peer->ibp_accepting > 0 ||
1155                          !list_empty(&peer->ibp_conns));  /* active conn */
1156                 last_alive = peer->ibp_last_alive;
1157         }
1158
1159         read_unlock_irqrestore(glock, flags);
1160
1161         if (last_alive != 0)
1162                 *when = last_alive;
1163
1164         /* peer is not persistent in hash, trigger peer creation
1165          * and connection establishment with a NULL tx */
1166         if (peer == NULL)
1167                 kiblnd_launch_tx(ni, NULL, nid);
1168
1169         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1170                libcfs_nid2str(nid), peer,
1171                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1172         return;
1173 }
1174
1175 void
1176 kiblnd_free_pages(kib_pages_t *p)
1177 {
1178         int     npages = p->ibp_npages;
1179         int     i;
1180
1181         for (i = 0; i < npages; i++) {
1182                 if (p->ibp_pages[i] != NULL)
1183                         __free_page(p->ibp_pages[i]);
1184         }
1185
1186         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1187 }
1188
1189 int
1190 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1191 {
1192         kib_pages_t     *p;
1193         int             i;
1194
1195         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1196                          offsetof(kib_pages_t, ibp_pages[npages]));
1197         if (p == NULL) {
1198                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1199                 return -ENOMEM;
1200         }
1201
1202         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1203         p->ibp_npages = npages;
1204
1205         for (i = 0; i < npages; i++) {
1206                 p->ibp_pages[i] = alloc_pages_node(
1207                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1208                                     __GFP_IO, 0);
1209                 if (p->ibp_pages[i] == NULL) {
1210                         CERROR("Can't allocate page %d of %d\n", i, npages);
1211                         kiblnd_free_pages(p);
1212                         return -ENOMEM;
1213                 }
1214         }
1215
1216         *pp = p;
1217         return 0;
1218 }
1219
1220 void
1221 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1222 {
1223         kib_rx_t *rx;
1224         int       i;
1225
1226         LASSERT (conn->ibc_rxs != NULL);
1227         LASSERT (conn->ibc_hdev != NULL);
1228
1229         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1230                 rx = &conn->ibc_rxs[i];
1231
1232                 LASSERT (rx->rx_nob >= 0); /* not posted */
1233
1234                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1235                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1236                                                           rx->rx_msgaddr),
1237                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1238         }
1239
1240         kiblnd_free_pages(conn->ibc_rx_pages);
1241
1242         conn->ibc_rx_pages = NULL;
1243 }
1244
1245 void
1246 kiblnd_map_rx_descs(kib_conn_t *conn)
1247 {
1248         kib_rx_t       *rx;
1249         struct page    *pg;
1250         int          pg_off;
1251         int          ipg;
1252         int          i;
1253
1254         for (pg_off = ipg = i = 0;
1255              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1256                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1257                 rx = &conn->ibc_rxs[i];
1258
1259                 rx->rx_conn = conn;
1260                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1261
1262                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1263                                                        rx->rx_msg, IBLND_MSG_SIZE,
1264                                                        DMA_FROM_DEVICE);
1265                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1266                                                    rx->rx_msgaddr));
1267                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1268
1269                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1270                        i, rx->rx_msg, rx->rx_msgaddr,
1271                        lnet_page2phys(pg) + pg_off);
1272
1273                 pg_off += IBLND_MSG_SIZE;
1274                 LASSERT (pg_off <= PAGE_SIZE);
1275
1276                 if (pg_off == PAGE_SIZE) {
1277                         pg_off = 0;
1278                         ipg++;
1279                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1280                 }
1281         }
1282 }
1283
1284 static void
1285 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1286 {
1287         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1288         kib_tx_t       *tx;
1289         int          i;
1290
1291         LASSERT (tpo->tpo_pool.po_allocated == 0);
1292
1293         if (hdev == NULL)
1294                 return;
1295
1296         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1297                 tx = &tpo->tpo_tx_descs[i];
1298                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1299                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1300                                                           tx->tx_msgaddr),
1301                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1302         }
1303
1304         kiblnd_hdev_decref(hdev);
1305         tpo->tpo_hdev = NULL;
1306 }
1307
1308 static kib_hca_dev_t *
1309 kiblnd_current_hdev(kib_dev_t *dev)
1310 {
1311         kib_hca_dev_t *hdev;
1312         unsigned long  flags;
1313         int         i = 0;
1314
1315         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1316         while (dev->ibd_failover) {
1317                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1318                 if (i++ % 50 == 0)
1319                         CDEBUG(D_NET, "%s: Wait for failover\n",
1320                                dev->ibd_ifname);
1321                 schedule_timeout(cfs_time_seconds(1) / 100);
1322
1323                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1324         }
1325
1326         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1327         hdev = dev->ibd_hdev;
1328
1329         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1330
1331         return hdev;
1332 }
1333
1334 static void
1335 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1336 {
1337         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1338         kib_pool_t     *pool  = &tpo->tpo_pool;
1339         kib_net_t      *net   = pool->po_owner->ps_net;
1340         kib_dev_t      *dev;
1341         struct page    *page;
1342         kib_tx_t       *tx;
1343         int          page_offset;
1344         int          ipage;
1345         int          i;
1346
1347         LASSERT (net != NULL);
1348
1349         dev = net->ibn_dev;
1350
1351         /* pre-mapped messages are not bigger than 1 page */
1352         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1353
1354         /* No fancy arithmetic when we do the buffer calculations */
1355         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1356
1357         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1358
1359         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1360                 page = txpgs->ibp_pages[ipage];
1361                 tx = &tpo->tpo_tx_descs[i];
1362
1363                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1364                                            page_offset);
1365
1366                 tx->tx_msgaddr = kiblnd_dma_map_single(
1367                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1368                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1369                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1370                                                    tx->tx_msgaddr));
1371                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1372
1373                 list_add(&tx->tx_list, &pool->po_free_list);
1374
1375                 page_offset += IBLND_MSG_SIZE;
1376                 LASSERT (page_offset <= PAGE_SIZE);
1377
1378                 if (page_offset == PAGE_SIZE) {
1379                         page_offset = 0;
1380                         ipage++;
1381                         LASSERT (ipage <= txpgs->ibp_npages);
1382                 }
1383         }
1384 }
1385
1386 struct ib_mr *
1387 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1388 {
1389         __u64   index;
1390
1391         LASSERT (hdev->ibh_mrs[0] != NULL);
1392
1393         if (hdev->ibh_nmrs == 1)
1394                 return hdev->ibh_mrs[0];
1395
1396         index = addr >> hdev->ibh_mr_shift;
1397
1398         if (index <  hdev->ibh_nmrs &&
1399             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1400                 return hdev->ibh_mrs[index];
1401
1402         return NULL;
1403 }
1404
1405 struct ib_mr *
1406 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1407 {
1408         struct ib_mr *prev_mr;
1409         struct ib_mr *mr;
1410         int        i;
1411
1412         LASSERT (hdev->ibh_mrs[0] != NULL);
1413
1414         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1415             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1416                 return NULL;
1417
1418         if (hdev->ibh_nmrs == 1)
1419                 return hdev->ibh_mrs[0];
1420
1421         for (i = 0, mr = prev_mr = NULL;
1422              i < rd->rd_nfrags; i++) {
1423                 mr = kiblnd_find_dma_mr(hdev,
1424                                         rd->rd_frags[i].rf_addr,
1425                                         rd->rd_frags[i].rf_nob);
1426                 if (prev_mr == NULL)
1427                         prev_mr = mr;
1428
1429                 if (mr == NULL || prev_mr != mr) {
1430                         /* Can't covered by one single MR */
1431                         mr = NULL;
1432                         break;
1433                 }
1434         }
1435
1436         return mr;
1437 }
1438
1439 void
1440 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1441 {
1442         LASSERT (pool->fpo_map_count == 0);
1443
1444         if (pool->fpo_fmr_pool != NULL)
1445                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1446
1447         if (pool->fpo_hdev != NULL)
1448                 kiblnd_hdev_decref(pool->fpo_hdev);
1449
1450         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1451 }
1452
1453 void
1454 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1455 {
1456         kib_fmr_pool_t *pool;
1457
1458         while (!list_empty(head)) {
1459                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1460                 list_del(&pool->fpo_list);
1461                 kiblnd_destroy_fmr_pool(pool);
1462         }
1463 }
1464
1465 static int kiblnd_fmr_pool_size(int ncpts)
1466 {
1467         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1468
1469         return max(IBLND_FMR_POOL, size);
1470 }
1471
1472 static int kiblnd_fmr_flush_trigger(int ncpts)
1473 {
1474         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1475
1476         return max(IBLND_FMR_POOL_FLUSH, size);
1477 }
1478
1479 int
1480 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1481 {
1482         /* FMR pool for RDMA */
1483         kib_dev_t              *dev = fps->fps_net->ibn_dev;
1484         kib_fmr_pool_t    *fpo;
1485         struct ib_fmr_pool_param param = {
1486                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1487                 .page_shift     = PAGE_SHIFT,
1488                 .access     = (IB_ACCESS_LOCAL_WRITE |
1489                                       IB_ACCESS_REMOTE_WRITE),
1490                 .pool_size         = fps->fps_pool_size,
1491                 .dirty_watermark   = fps->fps_flush_trigger,
1492                 .flush_function    = NULL,
1493                 .flush_arg       = NULL,
1494                 .cache       = !!*kiblnd_tunables.kib_fmr_cache};
1495         int rc;
1496
1497         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1498         if (fpo == NULL)
1499                 return -ENOMEM;
1500
1501         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1502
1503         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1504         if (IS_ERR(fpo->fpo_fmr_pool)) {
1505                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1506                 CERROR("Failed to create FMR pool: %d\n", rc);
1507
1508                 kiblnd_hdev_decref(fpo->fpo_hdev);
1509                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1510                 return rc;
1511         }
1512
1513         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1514         fpo->fpo_owner    = fps;
1515         *pp_fpo = fpo;
1516
1517         return 0;
1518 }
1519
1520 static void
1521 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1522 {
1523         if (fps->fps_net == NULL) /* intialized? */
1524                 return;
1525
1526         spin_lock(&fps->fps_lock);
1527
1528         while (!list_empty(&fps->fps_pool_list)) {
1529                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1530                                                  kib_fmr_pool_t, fpo_list);
1531                 fpo->fpo_failed = 1;
1532                 list_del(&fpo->fpo_list);
1533                 if (fpo->fpo_map_count == 0)
1534                         list_add(&fpo->fpo_list, zombies);
1535                 else
1536                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1537         }
1538
1539         spin_unlock(&fps->fps_lock);
1540 }
1541
1542 static void
1543 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1544 {
1545         if (fps->fps_net != NULL) { /* initialized? */
1546                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1547                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1548         }
1549 }
1550
1551 static int
1552 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1553                         int pool_size, int flush_trigger)
1554 {
1555         kib_fmr_pool_t *fpo;
1556         int          rc;
1557
1558         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1559
1560         fps->fps_net = net;
1561         fps->fps_cpt = cpt;
1562         fps->fps_pool_size = pool_size;
1563         fps->fps_flush_trigger = flush_trigger;
1564         spin_lock_init(&fps->fps_lock);
1565         INIT_LIST_HEAD(&fps->fps_pool_list);
1566         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1567
1568         rc = kiblnd_create_fmr_pool(fps, &fpo);
1569         if (rc == 0)
1570                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1571
1572         return rc;
1573 }
1574
1575 static int
1576 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1577 {
1578         if (fpo->fpo_map_count != 0) /* still in use */
1579                 return 0;
1580         if (fpo->fpo_failed)
1581                 return 1;
1582         return cfs_time_aftereq(now, fpo->fpo_deadline);
1583 }
1584
1585 void
1586 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1587 {
1588         LIST_HEAD     (zombies);
1589         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1590         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1591         cfs_time_t       now = cfs_time_current();
1592         kib_fmr_pool_t    *tmp;
1593         int             rc;
1594
1595         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1596         LASSERT (rc == 0);
1597
1598         if (status != 0) {
1599                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1600                 LASSERT (rc == 0);
1601         }
1602
1603         fmr->fmr_pool = NULL;
1604         fmr->fmr_pfmr = NULL;
1605
1606         spin_lock(&fps->fps_lock);
1607         fpo->fpo_map_count --;  /* decref the pool */
1608
1609         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1610                 /* the first pool is persistent */
1611                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1612                         continue;
1613
1614                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1615                         list_move(&fpo->fpo_list, &zombies);
1616                         fps->fps_version ++;
1617                 }
1618         }
1619         spin_unlock(&fps->fps_lock);
1620
1621         if (!list_empty(&zombies))
1622                 kiblnd_destroy_fmr_pool_list(&zombies);
1623 }
1624
1625 int
1626 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1627                     __u64 iov, kib_fmr_t *fmr)
1628 {
1629         struct ib_pool_fmr *pfmr;
1630         kib_fmr_pool_t     *fpo;
1631         __u64          version;
1632         int              rc;
1633
1634  again:
1635         spin_lock(&fps->fps_lock);
1636         version = fps->fps_version;
1637         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1638                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1639                 fpo->fpo_map_count++;
1640                 spin_unlock(&fps->fps_lock);
1641
1642                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1643                                             pages, npages, iov);
1644                 if (likely(!IS_ERR(pfmr))) {
1645                         fmr->fmr_pool = fpo;
1646                         fmr->fmr_pfmr = pfmr;
1647                         return 0;
1648                 }
1649
1650                 spin_lock(&fps->fps_lock);
1651                 fpo->fpo_map_count--;
1652                 if (PTR_ERR(pfmr) != -EAGAIN) {
1653                         spin_unlock(&fps->fps_lock);
1654                         return PTR_ERR(pfmr);
1655                 }
1656
1657                 /* EAGAIN and ... */
1658                 if (version != fps->fps_version) {
1659                         spin_unlock(&fps->fps_lock);
1660                         goto again;
1661                 }
1662         }
1663
1664         if (fps->fps_increasing) {
1665                 spin_unlock(&fps->fps_lock);
1666                 CDEBUG(D_NET, "Another thread is allocating new "
1667                        "FMR pool, waiting for her to complete\n");
1668                 schedule();
1669                 goto again;
1670
1671         }
1672
1673         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1674                 /* someone failed recently */
1675                 spin_unlock(&fps->fps_lock);
1676                 return -EAGAIN;
1677         }
1678
1679         fps->fps_increasing = 1;
1680         spin_unlock(&fps->fps_lock);
1681
1682         CDEBUG(D_NET, "Allocate new FMR pool\n");
1683         rc = kiblnd_create_fmr_pool(fps, &fpo);
1684         spin_lock(&fps->fps_lock);
1685         fps->fps_increasing = 0;
1686         if (rc == 0) {
1687                 fps->fps_version++;
1688                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1689         } else {
1690                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1691         }
1692         spin_unlock(&fps->fps_lock);
1693
1694         goto again;
1695 }
1696
1697 static void
1698 kiblnd_fini_pool(kib_pool_t *pool)
1699 {
1700         LASSERT (list_empty(&pool->po_free_list));
1701         LASSERT (pool->po_allocated == 0);
1702
1703         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1704 }
1705
1706 static void
1707 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1708 {
1709         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1710
1711         memset(pool, 0, sizeof(kib_pool_t));
1712         INIT_LIST_HEAD(&pool->po_free_list);
1713         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1714         pool->po_owner    = ps;
1715         pool->po_size     = size;
1716 }
1717
1718 void
1719 kiblnd_destroy_pool_list(struct list_head *head)
1720 {
1721         kib_pool_t *pool;
1722
1723         while (!list_empty(head)) {
1724                 pool = list_entry(head->next, kib_pool_t, po_list);
1725                 list_del(&pool->po_list);
1726
1727                 LASSERT (pool->po_owner != NULL);
1728                 pool->po_owner->ps_pool_destroy(pool);
1729         }
1730 }
1731
1732 static void
1733 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1734 {
1735         if (ps->ps_net == NULL) /* intialized? */
1736                 return;
1737
1738         spin_lock(&ps->ps_lock);
1739         while (!list_empty(&ps->ps_pool_list)) {
1740                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1741                                             kib_pool_t, po_list);
1742                 po->po_failed = 1;
1743                 list_del(&po->po_list);
1744                 if (po->po_allocated == 0)
1745                         list_add(&po->po_list, zombies);
1746                 else
1747                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1748         }
1749         spin_unlock(&ps->ps_lock);
1750 }
1751
1752 static void
1753 kiblnd_fini_poolset(kib_poolset_t *ps)
1754 {
1755         if (ps->ps_net != NULL) { /* initialized? */
1756                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1757                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1758         }
1759 }
1760
1761 static int
1762 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1763                     kib_net_t *net, char *name, int size,
1764                     kib_ps_pool_create_t po_create,
1765                     kib_ps_pool_destroy_t po_destroy,
1766                     kib_ps_node_init_t nd_init,
1767                     kib_ps_node_fini_t nd_fini)
1768 {
1769         kib_pool_t      *pool;
1770         int             rc;
1771
1772         memset(ps, 0, sizeof(kib_poolset_t));
1773
1774         ps->ps_cpt          = cpt;
1775         ps->ps_net        = net;
1776         ps->ps_pool_create  = po_create;
1777         ps->ps_pool_destroy = po_destroy;
1778         ps->ps_node_init    = nd_init;
1779         ps->ps_node_fini    = nd_fini;
1780         ps->ps_pool_size    = size;
1781         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1782             >= sizeof(ps->ps_name))
1783                 return -E2BIG;
1784         spin_lock_init(&ps->ps_lock);
1785         INIT_LIST_HEAD(&ps->ps_pool_list);
1786         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1787
1788         rc = ps->ps_pool_create(ps, size, &pool);
1789         if (rc == 0)
1790                 list_add(&pool->po_list, &ps->ps_pool_list);
1791         else
1792                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1793
1794         return rc;
1795 }
1796
1797 static int
1798 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1799 {
1800         if (pool->po_allocated != 0) /* still in use */
1801                 return 0;
1802         if (pool->po_failed)
1803                 return 1;
1804         return cfs_time_aftereq(now, pool->po_deadline);
1805 }
1806
1807 void
1808 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1809 {
1810         LIST_HEAD  (zombies);
1811         kib_poolset_t  *ps = pool->po_owner;
1812         kib_pool_t     *tmp;
1813         cfs_time_t      now = cfs_time_current();
1814
1815         spin_lock(&ps->ps_lock);
1816
1817         if (ps->ps_node_fini != NULL)
1818                 ps->ps_node_fini(pool, node);
1819
1820         LASSERT (pool->po_allocated > 0);
1821         list_add(node, &pool->po_free_list);
1822         pool->po_allocated --;
1823
1824         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1825                 /* the first pool is persistent */
1826                 if (ps->ps_pool_list.next == &pool->po_list)
1827                         continue;
1828
1829                 if (kiblnd_pool_is_idle(pool, now))
1830                         list_move(&pool->po_list, &zombies);
1831         }
1832         spin_unlock(&ps->ps_lock);
1833
1834         if (!list_empty(&zombies))
1835                 kiblnd_destroy_pool_list(&zombies);
1836 }
1837
1838 struct list_head *
1839 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1840 {
1841         struct list_head            *node;
1842         kib_pool_t          *pool;
1843         int                 rc;
1844
1845  again:
1846         spin_lock(&ps->ps_lock);
1847         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1848                 if (list_empty(&pool->po_free_list))
1849                         continue;
1850
1851                 pool->po_allocated ++;
1852                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1853                 node = pool->po_free_list.next;
1854                 list_del(node);
1855
1856                 if (ps->ps_node_init != NULL) {
1857                         /* still hold the lock */
1858                         ps->ps_node_init(pool, node);
1859                 }
1860                 spin_unlock(&ps->ps_lock);
1861                 return node;
1862         }
1863
1864         /* no available tx pool and ... */
1865         if (ps->ps_increasing) {
1866                 /* another thread is allocating a new pool */
1867                 spin_unlock(&ps->ps_lock);
1868                 CDEBUG(D_NET, "Another thread is allocating new "
1869                        "%s pool, waiting for her to complete\n",
1870                        ps->ps_name);
1871                 schedule();
1872                 goto again;
1873         }
1874
1875         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1876                 /* someone failed recently */
1877                 spin_unlock(&ps->ps_lock);
1878                 return NULL;
1879         }
1880
1881         ps->ps_increasing = 1;
1882         spin_unlock(&ps->ps_lock);
1883
1884         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1885
1886         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1887
1888         spin_lock(&ps->ps_lock);
1889         ps->ps_increasing = 0;
1890         if (rc == 0) {
1891                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1892         } else {
1893                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1894                 CERROR("Can't allocate new %s pool because out of memory\n",
1895                        ps->ps_name);
1896         }
1897         spin_unlock(&ps->ps_lock);
1898
1899         goto again;
1900 }
1901
1902 void
1903 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1904 {
1905         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1906         struct ib_mr    *mr  = pmr->pmr_mr;
1907
1908         pmr->pmr_mr = NULL;
1909         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1910         if (mr != NULL)
1911                 ib_dereg_mr(mr);
1912 }
1913
1914 int
1915 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1916                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1917 {
1918         kib_phys_mr_t *pmr;
1919         struct list_head    *node;
1920         int         rc;
1921         int         i;
1922
1923         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1924         if (node == NULL) {
1925                 CERROR("Failed to allocate PMR descriptor\n");
1926                 return -ENOMEM;
1927         }
1928
1929         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1930         if (pmr->pmr_pool->ppo_hdev != hdev) {
1931                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1932                 return -EAGAIN;
1933         }
1934
1935         for (i = 0; i < rd->rd_nfrags; i ++) {
1936                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1937                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1938         }
1939
1940         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1941                                      pmr->pmr_ipb, rd->rd_nfrags,
1942                                      IB_ACCESS_LOCAL_WRITE |
1943                                      IB_ACCESS_REMOTE_WRITE,
1944                                      iova);
1945         if (!IS_ERR(pmr->pmr_mr)) {
1946                 pmr->pmr_iova = *iova;
1947                 *pp_pmr = pmr;
1948                 return 0;
1949         }
1950
1951         rc = PTR_ERR(pmr->pmr_mr);
1952         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1953
1954         pmr->pmr_mr = NULL;
1955         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1956
1957         return rc;
1958 }
1959
1960 static void
1961 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1962 {
1963         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1964         kib_phys_mr_t  *pmr;
1965
1966         LASSERT (pool->po_allocated == 0);
1967
1968         while (!list_empty(&pool->po_free_list)) {
1969                 pmr = list_entry(pool->po_free_list.next,
1970                                      kib_phys_mr_t, pmr_list);
1971
1972                 LASSERT (pmr->pmr_mr == NULL);
1973                 list_del(&pmr->pmr_list);
1974
1975                 if (pmr->pmr_ipb != NULL) {
1976                         LIBCFS_FREE(pmr->pmr_ipb,
1977                                     IBLND_MAX_RDMA_FRAGS *
1978                                     sizeof(struct ib_phys_buf));
1979                 }
1980
1981                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1982         }
1983
1984         kiblnd_fini_pool(pool);
1985         if (ppo->ppo_hdev != NULL)
1986                 kiblnd_hdev_decref(ppo->ppo_hdev);
1987
1988         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1989 }
1990
1991 static inline int kiblnd_pmr_pool_size(int ncpts)
1992 {
1993         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1994
1995         return max(IBLND_PMR_POOL, size);
1996 }
1997
1998 static int
1999 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2000 {
2001         struct kib_pmr_pool     *ppo;
2002         struct kib_pool         *pool;
2003         kib_phys_mr_t           *pmr;
2004         int                     i;
2005
2006         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2007                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2008         if (ppo == NULL) {
2009                 CERROR("Failed to allocate PMR pool\n");
2010                 return -ENOMEM;
2011         }
2012
2013         pool = &ppo->ppo_pool;
2014         kiblnd_init_pool(ps, pool, size);
2015
2016         for (i = 0; i < size; i++) {
2017                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2018                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2019                 if (pmr == NULL)
2020                         break;
2021
2022                 pmr->pmr_pool = ppo;
2023                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2024                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2025                 if (pmr->pmr_ipb == NULL)
2026                         break;
2027
2028                 list_add(&pmr->pmr_list, &pool->po_free_list);
2029         }
2030
2031         if (i < size) {
2032                 ps->ps_pool_destroy(pool);
2033                 return -ENOMEM;
2034         }
2035
2036         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2037         *pp_po = pool;
2038         return 0;
2039 }
2040
2041 static void
2042 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2043 {
2044         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2045         int          i;
2046
2047         LASSERT (pool->po_allocated == 0);
2048
2049         if (tpo->tpo_tx_pages != NULL) {
2050                 kiblnd_unmap_tx_pool(tpo);
2051                 kiblnd_free_pages(tpo->tpo_tx_pages);
2052         }
2053
2054         if (tpo->tpo_tx_descs == NULL)
2055                 goto out;
2056
2057         for (i = 0; i < pool->po_size; i++) {
2058                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2059
2060                 list_del(&tx->tx_list);
2061                 if (tx->tx_pages != NULL)
2062                         LIBCFS_FREE(tx->tx_pages,
2063                                     LNET_MAX_IOV *
2064                                     sizeof(*tx->tx_pages));
2065                 if (tx->tx_frags != NULL)
2066                         LIBCFS_FREE(tx->tx_frags,
2067                                     IBLND_MAX_RDMA_FRAGS *
2068                                             sizeof(*tx->tx_frags));
2069                 if (tx->tx_wrq != NULL)
2070                         LIBCFS_FREE(tx->tx_wrq,
2071                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2072                                     sizeof(*tx->tx_wrq));
2073                 if (tx->tx_sge != NULL)
2074                         LIBCFS_FREE(tx->tx_sge,
2075                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2076                                     sizeof(*tx->tx_sge));
2077                 if (tx->tx_rd != NULL)
2078                         LIBCFS_FREE(tx->tx_rd,
2079                                     offsetof(kib_rdma_desc_t,
2080                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2081         }
2082
2083         LIBCFS_FREE(tpo->tpo_tx_descs,
2084                     pool->po_size * sizeof(kib_tx_t));
2085 out:
2086         kiblnd_fini_pool(pool);
2087         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2088 }
2089
2090 static int kiblnd_tx_pool_size(int ncpts)
2091 {
2092         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2093
2094         return max(IBLND_TX_POOL, ntx);
2095 }
2096
2097 static int
2098 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2099 {
2100         int         i;
2101         int         npg;
2102         kib_pool_t    *pool;
2103         kib_tx_pool_t *tpo;
2104
2105         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2106         if (tpo == NULL) {
2107                 CERROR("Failed to allocate TX pool\n");
2108                 return -ENOMEM;
2109         }
2110
2111         pool = &tpo->tpo_pool;
2112         kiblnd_init_pool(ps, pool, size);
2113         tpo->tpo_tx_descs = NULL;
2114         tpo->tpo_tx_pages = NULL;
2115
2116         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2117         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2118                 CERROR("Can't allocate tx pages: %d\n", npg);
2119                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2120                 return -ENOMEM;
2121         }
2122
2123         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2124                          size * sizeof(kib_tx_t));
2125         if (tpo->tpo_tx_descs == NULL) {
2126                 CERROR("Can't allocate %d tx descriptors\n", size);
2127                 ps->ps_pool_destroy(pool);
2128                 return -ENOMEM;
2129         }
2130
2131         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2132
2133         for (i = 0; i < size; i++) {
2134                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2135
2136                 tx->tx_pool = tpo;
2137                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2138                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2139                                          lnet_cpt_table(), ps->ps_cpt,
2140                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2141                         if (tx->tx_pages == NULL)
2142                                 break;
2143                 }
2144
2145                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2146                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2147                 if (tx->tx_frags == NULL)
2148                         break;
2149
2150                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2151
2152                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2153                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2154                                  sizeof(*tx->tx_wrq));
2155                 if (tx->tx_wrq == NULL)
2156                         break;
2157
2158                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2159                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2160                                  sizeof(*tx->tx_sge));
2161                 if (tx->tx_sge == NULL)
2162                         break;
2163
2164                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2165                                  offsetof(kib_rdma_desc_t,
2166                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2167                 if (tx->tx_rd == NULL)
2168                         break;
2169         }
2170
2171         if (i == size) {
2172                 kiblnd_map_tx_pool(tpo);
2173                 *pp_po = pool;
2174                 return 0;
2175         }
2176
2177         ps->ps_pool_destroy(pool);
2178         return -ENOMEM;
2179 }
2180
2181 static void
2182 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2183 {
2184         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2185                                              tps_poolset);
2186         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2187
2188         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2189 }
2190
2191 void
2192 kiblnd_net_fini_pools(kib_net_t *net)
2193 {
2194         int     i;
2195
2196         cfs_cpt_for_each(i, lnet_cpt_table()) {
2197                 kib_tx_poolset_t        *tps;
2198                 kib_fmr_poolset_t       *fps;
2199                 kib_pmr_poolset_t       *pps;
2200
2201                 if (net->ibn_tx_ps != NULL) {
2202                         tps = net->ibn_tx_ps[i];
2203                         kiblnd_fini_poolset(&tps->tps_poolset);
2204                 }
2205
2206                 if (net->ibn_fmr_ps != NULL) {
2207                         fps = net->ibn_fmr_ps[i];
2208                         kiblnd_fini_fmr_poolset(fps);
2209                 }
2210
2211                 if (net->ibn_pmr_ps != NULL) {
2212                         pps = net->ibn_pmr_ps[i];
2213                         kiblnd_fini_poolset(&pps->pps_poolset);
2214                 }
2215         }
2216
2217         if (net->ibn_tx_ps != NULL) {
2218                 cfs_percpt_free(net->ibn_tx_ps);
2219                 net->ibn_tx_ps = NULL;
2220         }
2221
2222         if (net->ibn_fmr_ps != NULL) {
2223                 cfs_percpt_free(net->ibn_fmr_ps);
2224                 net->ibn_fmr_ps = NULL;
2225         }
2226
2227         if (net->ibn_pmr_ps != NULL) {
2228                 cfs_percpt_free(net->ibn_pmr_ps);
2229                 net->ibn_pmr_ps = NULL;
2230         }
2231 }
2232
2233 int
2234 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2235 {
2236         unsigned long   flags;
2237         int             cpt;
2238         int             rc;
2239         int             i;
2240
2241         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2242         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2243             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2244                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2245                                            flags);
2246                 goto create_tx_pool;
2247         }
2248
2249         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2250
2251         if (*kiblnd_tunables.kib_fmr_pool_size <
2252             *kiblnd_tunables.kib_ntx / 4) {
2253                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2254                        *kiblnd_tunables.kib_fmr_pool_size,
2255                        *kiblnd_tunables.kib_ntx / 4);
2256                 rc = -EINVAL;
2257                 goto failed;
2258         }
2259
2260         /* TX pool must be created later than FMR/PMR, see LU-2268
2261          * for details */
2262         LASSERT(net->ibn_tx_ps == NULL);
2263
2264         /* premapping can fail if ibd_nmr > 1, so we always create
2265          * FMR/PMR pool and map-on-demand if premapping failed */
2266
2267         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2268                                            sizeof(kib_fmr_poolset_t));
2269         if (net->ibn_fmr_ps == NULL) {
2270                 CERROR("Failed to allocate FMR pool array\n");
2271                 rc = -ENOMEM;
2272                 goto failed;
2273         }
2274
2275         for (i = 0; i < ncpts; i++) {
2276                 cpt = (cpts == NULL) ? i : cpts[i];
2277                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2278                                              kiblnd_fmr_pool_size(ncpts),
2279                                              kiblnd_fmr_flush_trigger(ncpts));
2280                 if (rc == -ENOSYS && i == 0) /* no FMR */
2281                         break; /* create PMR pool */
2282
2283                 if (rc != 0) { /* a real error */
2284                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2285                                cpt, rc);
2286                         goto failed;
2287                 }
2288         }
2289
2290         if (i > 0) {
2291                 LASSERT(i == ncpts);
2292                 goto create_tx_pool;
2293         }
2294
2295         cfs_percpt_free(net->ibn_fmr_ps);
2296         net->ibn_fmr_ps = NULL;
2297
2298         CWARN("Device does not support FMR, failing back to PMR\n");
2299
2300         if (*kiblnd_tunables.kib_pmr_pool_size <
2301             *kiblnd_tunables.kib_ntx / 4) {
2302                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2303                        *kiblnd_tunables.kib_pmr_pool_size,
2304                        *kiblnd_tunables.kib_ntx / 4);
2305                 rc = -EINVAL;
2306                 goto failed;
2307         }
2308
2309         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2310                                            sizeof(kib_pmr_poolset_t));
2311         if (net->ibn_pmr_ps == NULL) {
2312                 CERROR("Failed to allocate PMR pool array\n");
2313                 rc = -ENOMEM;
2314                 goto failed;
2315         }
2316
2317         for (i = 0; i < ncpts; i++) {
2318                 cpt = (cpts == NULL) ? i : cpts[i];
2319                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2320                                          cpt, net, "PMR",
2321                                          kiblnd_pmr_pool_size(ncpts),
2322                                          kiblnd_create_pmr_pool,
2323                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2324                 if (rc != 0) {
2325                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2326                                cpt, rc);
2327                         goto failed;
2328                 }
2329         }
2330
2331  create_tx_pool:
2332         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2333                                           sizeof(kib_tx_poolset_t));
2334         if (net->ibn_tx_ps == NULL) {
2335                 CERROR("Failed to allocate tx pool array\n");
2336                 rc = -ENOMEM;
2337                 goto failed;
2338         }
2339
2340         for (i = 0; i < ncpts; i++) {
2341                 cpt = (cpts == NULL) ? i : cpts[i];
2342                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2343                                          cpt, net, "TX",
2344                                          kiblnd_tx_pool_size(ncpts),
2345                                          kiblnd_create_tx_pool,
2346                                          kiblnd_destroy_tx_pool,
2347                                          kiblnd_tx_init, NULL);
2348                 if (rc != 0) {
2349                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2350                                cpt, rc);
2351                         goto failed;
2352                 }
2353         }
2354
2355         return 0;
2356  failed:
2357         kiblnd_net_fini_pools(net);
2358         LASSERT(rc != 0);
2359         return rc;
2360 }
2361
2362 static int
2363 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2364 {
2365         struct ib_device_attr *attr;
2366         int                 rc;
2367
2368         /* It's safe to assume a HCA can handle a page size
2369          * matching that of the native system */
2370         hdev->ibh_page_shift = PAGE_SHIFT;
2371         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2372         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2373
2374         LIBCFS_ALLOC(attr, sizeof(*attr));
2375         if (attr == NULL) {
2376                 CERROR("Out of memory\n");
2377                 return -ENOMEM;
2378         }
2379
2380         rc = ib_query_device(hdev->ibh_ibdev, attr);
2381         if (rc == 0)
2382                 hdev->ibh_mr_size = attr->max_mr_size;
2383
2384         LIBCFS_FREE(attr, sizeof(*attr));
2385
2386         if (rc != 0) {
2387                 CERROR("Failed to query IB device: %d\n", rc);
2388                 return rc;
2389         }
2390
2391         if (hdev->ibh_mr_size == ~0ULL) {
2392                 hdev->ibh_mr_shift = 64;
2393                 return 0;
2394         }
2395
2396         for (hdev->ibh_mr_shift = 0;
2397              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2398                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2399                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2400                         return 0;
2401         }
2402
2403         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2404         return -EINVAL;
2405 }
2406
2407 void
2408 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2409 {
2410         int     i;
2411
2412         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2413                 return;
2414
2415         for (i = 0; i < hdev->ibh_nmrs; i++) {
2416                 if (hdev->ibh_mrs[i] == NULL)
2417                         break;
2418
2419                 ib_dereg_mr(hdev->ibh_mrs[i]);
2420         }
2421
2422         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2423         hdev->ibh_mrs  = NULL;
2424         hdev->ibh_nmrs = 0;
2425 }
2426
2427 void
2428 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2429 {
2430         kiblnd_hdev_cleanup_mrs(hdev);
2431
2432         if (hdev->ibh_pd != NULL)
2433                 ib_dealloc_pd(hdev->ibh_pd);
2434
2435         if (hdev->ibh_cmid != NULL)
2436                 rdma_destroy_id(hdev->ibh_cmid);
2437
2438         LIBCFS_FREE(hdev, sizeof(*hdev));
2439 }
2440
2441 int
2442 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2443 {
2444         struct ib_mr *mr;
2445         int        i;
2446         int        rc;
2447         __u64    mm_size;
2448         __u64    mr_size;
2449         int        acflags = IB_ACCESS_LOCAL_WRITE |
2450                                 IB_ACCESS_REMOTE_WRITE;
2451
2452         rc = kiblnd_hdev_get_attr(hdev);
2453         if (rc != 0)
2454                 return rc;
2455
2456         if (hdev->ibh_mr_shift == 64) {
2457                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2458                 if (hdev->ibh_mrs == NULL) {
2459                         CERROR("Failed to allocate MRs table\n");
2460                         return -ENOMEM;
2461                 }
2462
2463                 hdev->ibh_mrs[0] = NULL;
2464                 hdev->ibh_nmrs   = 1;
2465
2466                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2467                 if (IS_ERR(mr)) {
2468                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2469                         kiblnd_hdev_cleanup_mrs(hdev);
2470                         return PTR_ERR(mr);
2471                 }
2472
2473                 hdev->ibh_mrs[0] = mr;
2474
2475                 goto out;
2476         }
2477
2478         mr_size = (1ULL << hdev->ibh_mr_shift);
2479         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2480
2481         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2482
2483         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2484                 /* it's 4T..., assume we will re-code at that time */
2485                 CERROR("Can't support memory size: x"LPX64
2486                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2487                 return -EINVAL;
2488         }
2489
2490         /* create an array of MRs to cover all memory */
2491         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2492         if (hdev->ibh_mrs == NULL) {
2493                 CERROR("Failed to allocate MRs' table\n");
2494                 return -ENOMEM;
2495         }
2496
2497         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2498
2499         for (i = 0; i < hdev->ibh_nmrs; i++) {
2500                 struct ib_phys_buf ipb;
2501                 __u64         iova;
2502
2503                 ipb.size = hdev->ibh_mr_size;
2504                 ipb.addr = i * mr_size;
2505                 iova     = ipb.addr;
2506
2507                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2508                 if (IS_ERR(mr)) {
2509                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2510                                " size "LPX64" : %ld\n",
2511                                ipb.addr, ipb.size, PTR_ERR(mr));
2512                         kiblnd_hdev_cleanup_mrs(hdev);
2513                         return PTR_ERR(mr);
2514                 }
2515
2516                 LASSERT (iova == ipb.addr);
2517
2518                 hdev->ibh_mrs[i] = mr;
2519         }
2520
2521 out:
2522         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2523                 LCONSOLE_INFO("Register global MR array, MR size: "
2524                               LPX64", array size: %d\n",
2525                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2526         return 0;
2527 }
2528
2529 static int
2530 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2531 {       /* DUMMY */
2532         return 0;
2533 }
2534
2535 static int
2536 kiblnd_dev_need_failover(kib_dev_t *dev)
2537 {
2538         struct rdma_cm_id  *cmid;
2539         struct sockaddr_in  srcaddr;
2540         struct sockaddr_in  dstaddr;
2541         int              rc;
2542
2543         if (dev->ibd_hdev == NULL || /* initializing */
2544             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2545             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2546                 return 1;
2547
2548         /* XXX: it's UGLY, but I don't have better way to find
2549          * ib-bonding HCA failover because:
2550          *
2551          * a. no reliable CM event for HCA failover...
2552          * b. no OFED API to get ib_device for current net_device...
2553          *
2554          * We have only two choices at this point:
2555          *
2556          * a. rdma_bind_addr(), it will conflict with listener cmid
2557          * b. rdma_resolve_addr() to zero addr */
2558         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2559                                      IB_QPT_RC);
2560         if (IS_ERR(cmid)) {
2561                 rc = PTR_ERR(cmid);
2562                 CERROR("Failed to create cmid for failover: %d\n", rc);
2563                 return rc;
2564         }
2565
2566         memset(&srcaddr, 0, sizeof(srcaddr));
2567         srcaddr.sin_family      = AF_INET;
2568         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2569
2570         memset(&dstaddr, 0, sizeof(dstaddr));
2571         dstaddr.sin_family = AF_INET;
2572         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2573                                (struct sockaddr *)&dstaddr, 1);
2574         if (rc != 0 || cmid->device == NULL) {
2575                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2576                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2577                        cmid->device, rc);
2578                 rdma_destroy_id(cmid);
2579                 return rc;
2580         }
2581
2582         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2583                 /* don't need device failover */
2584                 rdma_destroy_id(cmid);
2585                 return 0;
2586         }
2587
2588         return 1;
2589 }
2590
2591 int
2592 kiblnd_dev_failover(kib_dev_t *dev)
2593 {
2594         LIST_HEAD      (zombie_tpo);
2595         LIST_HEAD      (zombie_ppo);
2596         LIST_HEAD      (zombie_fpo);
2597         struct rdma_cm_id  *cmid  = NULL;
2598         kib_hca_dev_t      *hdev  = NULL;
2599         kib_hca_dev_t      *old;
2600         struct ib_pd       *pd;
2601         kib_net_t         *net;
2602         struct sockaddr_in  addr;
2603         unsigned long       flags;
2604         int              rc = 0;
2605         int                 i;
2606
2607         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2608                  dev->ibd_can_failover ||
2609                  dev->ibd_hdev == NULL);
2610
2611         rc = kiblnd_dev_need_failover(dev);
2612         if (rc <= 0)
2613                 goto out;
2614
2615         if (dev->ibd_hdev != NULL &&
2616             dev->ibd_hdev->ibh_cmid != NULL) {
2617                 /* XXX it's not good to close old listener at here,
2618                  * because we can fail to create new listener.
2619                  * But we have to close it now, otherwise rdma_bind_addr
2620                  * will return EADDRINUSE... How crap! */
2621                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2622
2623                 cmid = dev->ibd_hdev->ibh_cmid;
2624                 /* make next schedule of kiblnd_dev_need_failover()
2625                  * return 1 for me */
2626                 dev->ibd_hdev->ibh_cmid  = NULL;
2627                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2628
2629                 rdma_destroy_id(cmid);
2630         }
2631
2632         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2633                                      IB_QPT_RC);
2634         if (IS_ERR(cmid)) {
2635                 rc = PTR_ERR(cmid);
2636                 CERROR("Failed to create cmid for failover: %d\n", rc);
2637                 goto out;
2638         }
2639
2640         memset(&addr, 0, sizeof(addr));
2641         addr.sin_family      = AF_INET;
2642         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2643         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2644
2645         /* Bind to failover device or port */
2646         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2647         if (rc != 0 || cmid->device == NULL) {
2648                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2649                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2650                        cmid->device, rc);
2651                 rdma_destroy_id(cmid);
2652                 goto out;
2653         }
2654
2655         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2656         if (hdev == NULL) {
2657                 CERROR("Failed to allocate kib_hca_dev\n");
2658                 rdma_destroy_id(cmid);
2659                 rc = -ENOMEM;
2660                 goto out;
2661         }
2662
2663         atomic_set(&hdev->ibh_ref, 1);
2664         hdev->ibh_dev   = dev;
2665         hdev->ibh_cmid  = cmid;
2666         hdev->ibh_ibdev = cmid->device;
2667
2668         pd = ib_alloc_pd(cmid->device);
2669         if (IS_ERR(pd)) {
2670                 rc = PTR_ERR(pd);
2671                 CERROR("Can't allocate PD: %d\n", rc);
2672                 goto out;
2673         }
2674
2675         hdev->ibh_pd = pd;
2676
2677         rc = rdma_listen(cmid, 0);
2678         if (rc != 0) {
2679                 CERROR("Can't start new listener: %d\n", rc);
2680                 goto out;
2681         }
2682
2683         rc = kiblnd_hdev_setup_mrs(hdev);
2684         if (rc != 0) {
2685                 CERROR("Can't setup device: %d\n", rc);
2686                 goto out;
2687         }
2688
2689         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2690
2691         old = dev->ibd_hdev;
2692         dev->ibd_hdev = hdev; /* take over the refcount */
2693         hdev = old;
2694
2695         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2696                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2697                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2698                                             &zombie_tpo);
2699
2700                         if (net->ibn_fmr_ps != NULL) {
2701                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2702                                                         &zombie_fpo);
2703
2704                         } else if (net->ibn_pmr_ps != NULL) {
2705                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2706                                                     pps_poolset, &zombie_ppo);
2707                         }
2708                 }
2709         }
2710
2711         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2712  out:
2713         if (!list_empty(&zombie_tpo))
2714                 kiblnd_destroy_pool_list(&zombie_tpo);
2715         if (!list_empty(&zombie_ppo))
2716                 kiblnd_destroy_pool_list(&zombie_ppo);
2717         if (!list_empty(&zombie_fpo))
2718                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2719         if (hdev != NULL)
2720                 kiblnd_hdev_decref(hdev);
2721
2722         if (rc != 0)
2723                 dev->ibd_failed_failover++;
2724         else
2725                 dev->ibd_failed_failover = 0;
2726
2727         return rc;
2728 }
2729
2730 void
2731 kiblnd_destroy_dev (kib_dev_t *dev)
2732 {
2733         LASSERT (dev->ibd_nnets == 0);
2734         LASSERT (list_empty(&dev->ibd_nets));
2735
2736         list_del(&dev->ibd_fail_list);
2737         list_del(&dev->ibd_list);
2738
2739         if (dev->ibd_hdev != NULL)
2740                 kiblnd_hdev_decref(dev->ibd_hdev);
2741
2742         LIBCFS_FREE(dev, sizeof(*dev));
2743 }
2744
2745 kib_dev_t *
2746 kiblnd_create_dev(char *ifname)
2747 {
2748         struct net_device *netdev;
2749         kib_dev_t        *dev;
2750         __u32         netmask;
2751         __u32         ip;
2752         int             up;
2753         int             rc;
2754
2755         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2756         if (rc != 0) {
2757                 CERROR("Can't query IPoIB interface %s: %d\n",
2758                        ifname, rc);
2759                 return NULL;
2760         }
2761
2762         if (!up) {
2763                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2764                 return NULL;
2765         }
2766
2767         LIBCFS_ALLOC(dev, sizeof(*dev));
2768         if (dev == NULL)
2769                 return NULL;
2770
2771         memset(dev, 0, sizeof(*dev));
2772         netdev = dev_get_by_name(&init_net, ifname);
2773         if (netdev == NULL) {
2774                 dev->ibd_can_failover = 0;
2775         } else {
2776                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2777                 dev_put(netdev);
2778         }
2779
2780         INIT_LIST_HEAD(&dev->ibd_nets);
2781         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2782         INIT_LIST_HEAD(&dev->ibd_fail_list);
2783         dev->ibd_ifip = ip;
2784         strcpy(&dev->ibd_ifname[0], ifname);
2785
2786         /* initialize the device */
2787         rc = kiblnd_dev_failover(dev);
2788         if (rc != 0) {
2789                 CERROR("Can't initialize device: %d\n", rc);
2790                 LIBCFS_FREE(dev, sizeof(*dev));
2791                 return NULL;
2792         }
2793
2794         list_add_tail(&dev->ibd_list,
2795                           &kiblnd_data.kib_devs);
2796         return dev;
2797 }
2798
2799 void
2800 kiblnd_base_shutdown(void)
2801 {
2802         struct kib_sched_info   *sched;
2803         int                     i;
2804
2805         LASSERT (list_empty(&kiblnd_data.kib_devs));
2806
2807         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2808                atomic_read(&libcfs_kmemory));
2809
2810         switch (kiblnd_data.kib_init) {
2811         default:
2812                 LBUG();
2813
2814         case IBLND_INIT_ALL:
2815         case IBLND_INIT_DATA:
2816                 LASSERT (kiblnd_data.kib_peers != NULL);
2817                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2818                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2819                 }
2820                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2821                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2822
2823                 /* flag threads to terminate; wake and wait for them to die */
2824                 kiblnd_data.kib_shutdown = 1;
2825
2826                 /* NB: we really want to stop scheduler threads net by net
2827                  * instead of the whole module, this should be improved
2828                  * with dynamic configuration LNet */
2829                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2830                         wake_up_all(&sched->ibs_waitq);
2831
2832                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2833                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2834
2835                 i = 2;
2836                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2837                         i++;
2838                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2839                                "Waiting for %d threads to terminate\n",
2840                                atomic_read(&kiblnd_data.kib_nthreads));
2841                         cfs_pause(cfs_time_seconds(1));
2842                 }
2843
2844                 /* fall through */
2845
2846         case IBLND_INIT_NOTHING:
2847                 break;
2848         }
2849
2850         if (kiblnd_data.kib_peers != NULL) {
2851                 LIBCFS_FREE(kiblnd_data.kib_peers,
2852                             sizeof(struct list_head) *
2853                             kiblnd_data.kib_peer_hash_size);
2854         }
2855
2856         if (kiblnd_data.kib_scheds != NULL)
2857                 cfs_percpt_free(kiblnd_data.kib_scheds);
2858
2859         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2860                atomic_read(&libcfs_kmemory));
2861
2862         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2863         module_put(THIS_MODULE);
2864 }
2865
2866 void
2867 kiblnd_shutdown (lnet_ni_t *ni)
2868 {
2869         kib_net_t       *net = ni->ni_data;
2870         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2871         int            i;
2872         unsigned long     flags;
2873
2874         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2875
2876         if (net == NULL)
2877                 goto out;
2878
2879         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2880                atomic_read(&libcfs_kmemory));
2881
2882         write_lock_irqsave(g_lock, flags);
2883         net->ibn_shutdown = 1;
2884         write_unlock_irqrestore(g_lock, flags);
2885
2886         switch (net->ibn_init) {
2887         default:
2888                 LBUG();
2889
2890         case IBLND_INIT_ALL:
2891                 /* nuke all existing peers within this net */
2892                 kiblnd_del_peer(ni, LNET_NID_ANY);
2893
2894                 /* Wait for all peer state to clean up */
2895                 i = 2;
2896                 while (atomic_read(&net->ibn_npeers) != 0) {
2897                         i++;
2898                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2899                                "%s: waiting for %d peers to disconnect\n",
2900                                libcfs_nid2str(ni->ni_nid),
2901                                atomic_read(&net->ibn_npeers));
2902                         cfs_pause(cfs_time_seconds(1));
2903                 }
2904
2905                 kiblnd_net_fini_pools(net);
2906
2907                 write_lock_irqsave(g_lock, flags);
2908                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2909                 net->ibn_dev->ibd_nnets--;
2910                 list_del(&net->ibn_list);
2911                 write_unlock_irqrestore(g_lock, flags);
2912
2913                 /* fall through */
2914
2915         case IBLND_INIT_NOTHING:
2916                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2917
2918                 if (net->ibn_dev != NULL &&
2919                     net->ibn_dev->ibd_nnets == 0)
2920                         kiblnd_destroy_dev(net->ibn_dev);
2921
2922                 break;
2923         }
2924
2925         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2926                atomic_read(&libcfs_kmemory));
2927
2928         net->ibn_init = IBLND_INIT_NOTHING;
2929         ni->ni_data = NULL;
2930
2931         LIBCFS_FREE(net, sizeof(*net));
2932
2933 out:
2934         if (list_empty(&kiblnd_data.kib_devs))
2935                 kiblnd_base_shutdown();
2936         return;
2937 }
2938
2939 int
2940 kiblnd_base_startup(void)
2941 {
2942         struct kib_sched_info   *sched;
2943         int                     rc;
2944         int                     i;
2945
2946         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2947
2948         try_module_get(THIS_MODULE);
2949         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2950
2951         rwlock_init(&kiblnd_data.kib_global_lock);
2952
2953         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2954         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2955
2956         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2957         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2958                      sizeof(struct list_head) *
2959                             kiblnd_data.kib_peer_hash_size);
2960         if (kiblnd_data.kib_peers == NULL) {
2961                 goto failed;
2962         }
2963         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2964                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2965
2966         spin_lock_init(&kiblnd_data.kib_connd_lock);
2967         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2968         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2969         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2970         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2971
2972         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2973                                                   sizeof(*sched));
2974         if (kiblnd_data.kib_scheds == NULL)
2975                 goto failed;
2976
2977         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2978                 int     nthrs;
2979
2980                 spin_lock_init(&sched->ibs_lock);
2981                 INIT_LIST_HEAD(&sched->ibs_conns);
2982                 init_waitqueue_head(&sched->ibs_waitq);
2983
2984                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2985                 if (*kiblnd_tunables.kib_nscheds > 0) {
2986                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2987                 } else {
2988                         /* max to half of CPUs, another half is reserved for
2989                          * upper layer modules */
2990                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2991                 }
2992
2993                 sched->ibs_nthreads_max = nthrs;
2994                 sched->ibs_cpt = i;
2995         }
2996
2997         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2998
2999         /* lists/ptrs/locks initialised */
3000         kiblnd_data.kib_init = IBLND_INIT_DATA;
3001         /*****************************************************/
3002
3003         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3004         if (rc != 0) {
3005                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3006                 goto failed;
3007         }
3008
3009         if (*kiblnd_tunables.kib_dev_failover != 0)
3010                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3011                                          "kiblnd_failover");
3012
3013         if (rc != 0) {
3014                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3015                 goto failed;
3016         }
3017
3018         /* flag everything initialised */
3019         kiblnd_data.kib_init = IBLND_INIT_ALL;
3020         /*****************************************************/
3021
3022         return 0;
3023
3024  failed:
3025         kiblnd_base_shutdown();
3026         return -ENETDOWN;
3027 }
3028
3029 int
3030 kiblnd_start_schedulers(struct kib_sched_info *sched)
3031 {
3032         int     rc = 0;
3033         int     nthrs;
3034         int     i;
3035
3036         if (sched->ibs_nthreads == 0) {
3037                 if (*kiblnd_tunables.kib_nscheds > 0) {
3038                         nthrs = sched->ibs_nthreads_max;
3039                 } else {
3040                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3041                                                sched->ibs_cpt);
3042                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3043                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3044                 }
3045         } else {
3046                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3047                 /* increase one thread if there is new interface */
3048                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3049         }
3050
3051         for (i = 0; i < nthrs; i++) {
3052                 long    id;
3053                 char    name[20];
3054                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3055                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3056                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3057                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3058                 if (rc == 0)
3059                         continue;
3060
3061                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3062                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3063                 break;
3064         }
3065
3066         sched->ibs_nthreads += i;
3067         return rc;
3068 }
3069
3070 int
3071 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3072 {
3073         int     cpt;
3074         int     rc;
3075         int     i;
3076
3077         for (i = 0; i < ncpts; i++) {
3078                 struct kib_sched_info *sched;
3079
3080                 cpt = (cpts == NULL) ? i : cpts[i];
3081                 sched = kiblnd_data.kib_scheds[cpt];
3082
3083                 if (!newdev && sched->ibs_nthreads > 0)
3084                         continue;
3085
3086                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3087                 if (rc != 0) {
3088                         CERROR("Failed to start scheduler threads for %s\n",
3089                                dev->ibd_ifname);
3090                         return rc;
3091                 }
3092         }
3093         return 0;
3094 }
3095
3096 kib_dev_t *
3097 kiblnd_dev_search(char *ifname)
3098 {
3099         kib_dev_t       *alias = NULL;
3100         kib_dev_t       *dev;
3101         char            *colon;
3102         char            *colon2;
3103
3104         colon = strchr(ifname, ':');
3105         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3106                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3107                         return dev;
3108
3109                 if (alias != NULL)
3110                         continue;
3111
3112                 colon2 = strchr(dev->ibd_ifname, ':');
3113                 if (colon != NULL)
3114                         *colon = 0;
3115                 if (colon2 != NULL)
3116                         *colon2 = 0;
3117
3118                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3119                         alias = dev;
3120
3121                 if (colon != NULL)
3122                         *colon = ':';
3123                 if (colon2 != NULL)
3124                         *colon2 = ':';
3125         }
3126         return alias;
3127 }
3128
3129 int
3130 kiblnd_startup (lnet_ni_t *ni)
3131 {
3132         char                 *ifname;
3133         kib_dev_t               *ibdev = NULL;
3134         kib_net_t               *net;
3135         struct timeval      tv;
3136         unsigned long        flags;
3137         int                    rc;
3138         int                       newdev;
3139
3140         LASSERT (ni->ni_lnd == &the_o2iblnd);
3141
3142         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3143                 rc = kiblnd_base_startup();
3144                 if (rc != 0)
3145                         return rc;
3146         }
3147
3148         LIBCFS_ALLOC(net, sizeof(*net));
3149         ni->ni_data = net;
3150         if (net == NULL)
3151                 goto failed;
3152
3153         memset(net, 0, sizeof(*net));
3154
3155         do_gettimeofday(&tv);
3156         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3157
3158         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3159         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3160         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3161         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3162
3163         if (ni->ni_interfaces[0] != NULL) {
3164                 /* Use the IPoIB interface specified in 'networks=' */
3165
3166                 CLASSERT (LNET_MAX_INTERFACES > 1);
3167                 if (ni->ni_interfaces[1] != NULL) {
3168                         CERROR("Multiple interfaces not supported\n");
3169                         goto failed;
3170                 }
3171
3172                 ifname = ni->ni_interfaces[0];
3173         } else {
3174                 ifname = *kiblnd_tunables.kib_default_ipif;
3175         }
3176
3177         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3178                 CERROR("IPoIB interface name too long: %s\n", ifname);
3179                 goto failed;
3180         }
3181
3182         ibdev = kiblnd_dev_search(ifname);
3183
3184         newdev = ibdev == NULL;
3185         /* hmm...create kib_dev even for alias */
3186         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3187                 ibdev = kiblnd_create_dev(ifname);
3188
3189         if (ibdev == NULL)
3190                 goto failed;
3191
3192         net->ibn_dev = ibdev;
3193         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3194
3195         rc = kiblnd_dev_start_threads(ibdev, newdev,
3196                                       ni->ni_cpts, ni->ni_ncpts);
3197         if (rc != 0)
3198                 goto failed;
3199
3200         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3201         if (rc != 0) {
3202                 CERROR("Failed to initialize NI pools: %d\n", rc);
3203                 goto failed;
3204         }
3205
3206         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3207         ibdev->ibd_nnets++;
3208         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3209         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3210
3211         net->ibn_init = IBLND_INIT_ALL;
3212
3213         return 0;
3214
3215 failed:
3216         if (net->ibn_dev == NULL && ibdev != NULL)
3217                 kiblnd_destroy_dev(ibdev);
3218
3219         kiblnd_shutdown(ni);
3220
3221         CDEBUG(D_NET, "kiblnd_startup failed\n");
3222         return -ENETDOWN;
3223 }
3224
3225 void __exit
3226 kiblnd_module_fini (void)
3227 {
3228         lnet_unregister_lnd(&the_o2iblnd);
3229         kiblnd_tunables_fini();
3230 }
3231
3232 int __init
3233 kiblnd_module_init (void)
3234 {
3235         int    rc;
3236
3237         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3238         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3239                   <= IBLND_MSG_SIZE);
3240         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3241                   <= IBLND_MSG_SIZE);
3242
3243         rc = kiblnd_tunables_init();
3244         if (rc != 0)
3245                 return rc;
3246
3247         lnet_register_lnd(&the_o2iblnd);
3248
3249         return 0;
3250 }
3251
3252 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3253 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3254 MODULE_LICENSE("GPL");
3255
3256 module_init(kiblnd_module_init);
3257 module_exit(kiblnd_module_fini);