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         cfs_atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 350
 351         CFS_INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 352         CFS_INIT_LIST_HEAD(&peer->ibp_conns);
 353         CFS_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         cfs_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 (cfs_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 (cfs_list_empty(&peer->ibp_conns));
 380         LASSERT (cfs_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         cfs_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         cfs_list_t       *peer_list = kiblnd_nid2peerlist(nid);
 397         cfs_list_t       *tmp;
 398         kib_peer_t       *peer;
 399
 400         cfs_list_for_each (tmp, peer_list) {
 401
 402                 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
 403
 404                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 405                          peer->ibp_accepting > 0 ||
 406                          !cfs_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                        cfs_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 (cfs_list_empty(&peer->ibp_conns));
 424
 425         LASSERT (kiblnd_peer_active(peer));
 426         cfs_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         cfs_list_t            *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                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 445
 446                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 447                         LASSERT (peer->ibp_connecting > 0 ||
 448                                  peer->ibp_accepting > 0 ||
 449                                  !cfs_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 = cfs_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         cfs_list_t           *ctmp;
 474         cfs_list_t           *cnxt;
 475         kib_conn_t           *conn;
 476
 477         if (cfs_list_empty(&peer->ibp_conns)) {
 478                 kiblnd_unlink_peer_locked(peer);
 479         } else {
 480                 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 481                         conn = cfs_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         CFS_LIST_HEAD         (zombies);
 495         cfs_list_t            *ptmp;
 496         cfs_list_t            *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                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 515                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 516                         LASSERT (peer->ibp_connecting > 0 ||
 517                                  peer->ibp_accepting > 0 ||
 518                                  !cfs_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 (!cfs_list_empty(&peer->ibp_tx_queue)) {
 527                                 LASSERT (cfs_list_empty(&peer->ibp_conns));
 528
 529                                 cfs_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         cfs_list_t            *ptmp;
 550         kib_conn_t            *conn;
 551         cfs_list_t            *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                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 559
 560                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 561                         LASSERT (peer->ibp_connecting > 0 ||
 562                                  peer->ibp_accepting > 0 ||
 563                                  !cfs_list_empty(&peer->ibp_conns));
 564
 565                         if (peer->ibp_ni != ni)
 566                                 continue;
 567
 568                         cfs_list_for_each (ctmp, &peer->ibp_conns) {
 569                                 if (index-- > 0)
 570                                         continue;
 571
 572                                 conn = cfs_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         cfs_list_t      *tmp;
 610         int             i;
 611
 612         spin_lock(&conn->ibc_lock);
 613
 614         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 615                cfs_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         cfs_list_for_each(tmp, &conn->ibc_early_rxs)
 625                 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
 626
 627         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 628         cfs_list_for_each(tmp, &conn->ibc_tx_noops)
 629                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 632         cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 633                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 634
 635         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 636         cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 637                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 638
 639         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 640         cfs_list_for_each(tmp, &conn->ibc_tx_queue)
 641                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 642
 643         CDEBUG(D_CONSOLE, "   active_txs:\n");
 644         cfs_list_for_each(tmp, &conn->ibc_active_txs)
 645                 kiblnd_debug_tx(cfs_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 #ifdef HAVE_OFED_IB_COMP_VECTOR
 691 static int
 692 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 693 {
 694         cpumask_t       *mask;
 695         int             vectors;
 696         int             off;
 697         int             i;
 698
 699         vectors = conn->ibc_cmid->device->num_comp_vectors;
 700         if (vectors <= 1)
 701                 return 0;
 702
 703         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 704
 705         /* hash NID to CPU id in this partition... */
 706         off = conn->ibc_peer->ibp_nid % cpus_weight(*mask);
 707         for_each_cpu_mask(i, *mask) {
 708                 if (off-- == 0)
 709                         return i % vectors;
 710         }
 711
 712         LBUG();
 713         return 1;
 714 }
 715 #endif
 716
 717 kib_conn_t *
 718 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 719                    int state, int version)
 720 {
 721         /* CAVEAT EMPTOR:
 722          * If the new conn is created successfully it takes over the caller's
 723          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 724          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 725          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 726          * to destroy 'cmid' here since I'm called from the CM which still has
 727          * its ref on 'cmid'). */
 728         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
 729         kib_net_t              *net = peer->ibp_ni->ni_data;
 730         kib_dev_t              *dev;
 731         struct ib_qp_init_attr *init_qp_attr;
 732         struct kib_sched_info   *sched;
 733         kib_conn_t              *conn;
 734         struct ib_cq            *cq;
 735         unsigned long           flags;
 736         int                     cpt;
 737         int                     rc;
 738         int                     i;
 739
 740         LASSERT(net != NULL);
 741         LASSERT(!cfs_in_interrupt());
 742
 743         dev = net->ibn_dev;
 744
 745         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 746         sched = kiblnd_data.kib_scheds[cpt];
 747
 748         LASSERT(sched->ibs_nthreads > 0);
 749
 750         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 751                          sizeof(*init_qp_attr));
 752         if (init_qp_attr == NULL) {
 753                 CERROR("Can't allocate qp_attr for %s\n",
 754                        libcfs_nid2str(peer->ibp_nid));
 755                 goto failed_0;
 756         }
 757
 758         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 759         if (conn == NULL) {
 760                 CERROR("Can't allocate connection for %s\n",
 761                        libcfs_nid2str(peer->ibp_nid));
 762                 goto failed_1;
 763         }
 764
 765         conn->ibc_state = IBLND_CONN_INIT;
 766         conn->ibc_version = version;
 767         conn->ibc_peer = peer;                  /* I take the caller's ref */
 768         cmid->context = conn;                   /* for future CM callbacks */
 769         conn->ibc_cmid = cmid;
 770
 771         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 772         CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops);
 773         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 774         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 775         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 776         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 777         spin_lock_init(&conn->ibc_lock);
 778
 779         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 780                          sizeof(*conn->ibc_connvars));
 781         if (conn->ibc_connvars == NULL) {
 782                 CERROR("Can't allocate in-progress connection state\n");
 783                 goto failed_2;
 784         }
 785
 786         write_lock_irqsave(glock, flags);
 787         if (dev->ibd_failover) {
 788                 write_unlock_irqrestore(glock, flags);
 789                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 790                 goto failed_2;
 791         }
 792
 793         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 794                 /* wakeup failover thread and teardown connection */
 795                 if (kiblnd_dev_can_failover(dev)) {
 796                         cfs_list_add_tail(&dev->ibd_fail_list,
 797                                       &kiblnd_data.kib_failed_devs);
 798                         cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
 799                 }
 800
 801                 write_unlock_irqrestore(glock, flags);
 802                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 803                        cmid->device->name, dev->ibd_ifname);
 804                 goto failed_2;
 805         }
 806
 807         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 808         conn->ibc_hdev = dev->ibd_hdev;
 809
 810         kiblnd_setup_mtu_locked(cmid);
 811
 812         write_unlock_irqrestore(glock, flags);
 813
 814         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 815                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 816         if (conn->ibc_rxs == NULL) {
 817                 CERROR("Cannot allocate RX buffers\n");
 818                 goto failed_2;
 819         }
 820
 821         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 822                                 IBLND_RX_MSG_PAGES(version));
 823         if (rc != 0)
 824                 goto failed_2;
 825
 826         kiblnd_map_rx_descs(conn);
 827
 828 #ifdef HAVE_OFED_IB_COMP_VECTOR
 829         cq = ib_create_cq(cmid->device,
 830                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 831                           IBLND_CQ_ENTRIES(version),
 832                           kiblnd_get_completion_vector(conn, cpt));
 833 #else
 834         cq = ib_create_cq(cmid->device,
 835                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 836                           IBLND_CQ_ENTRIES(version));
 837 #endif
 838         if (IS_ERR(cq)) {
 839                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 840                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 841                 goto failed_2;
 842         }
 843
 844         conn->ibc_cq = cq;
 845
 846         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 847         if (rc != 0) {
 848                 CERROR("Can't request completion notificiation: %d\n", rc);
 849                 goto failed_2;
 850         }
 851
 852         init_qp_attr->event_handler = kiblnd_qp_event;
 853         init_qp_attr->qp_context = conn;
 854         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 855         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 856         init_qp_attr->cap.max_send_sge = 1;
 857         init_qp_attr->cap.max_recv_sge = 1;
 858         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 859         init_qp_attr->qp_type = IB_QPT_RC;
 860         init_qp_attr->send_cq = cq;
 861         init_qp_attr->recv_cq = cq;
 862
 863         conn->ibc_sched = sched;
 864
 865         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 866         if (rc != 0) {
 867                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 868                        rc, init_qp_attr->cap.max_send_wr,
 869                        init_qp_attr->cap.max_recv_wr);
 870                 goto failed_2;
 871         }
 872
 873         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 874
 875         /* 1 ref for caller and each rxmsg */
 876         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 877         conn->ibc_nrx = IBLND_RX_MSGS(version);
 878
 879         /* post receives */
 880         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 881                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 882                                     IBLND_POSTRX_NO_CREDIT);
 883                 if (rc != 0) {
 884                         CERROR("Can't post rxmsg: %d\n", rc);
 885
 886                         /* Make posted receives complete */
 887                         kiblnd_abort_receives(conn);
 888
 889                         /* correct # of posted buffers
 890                          * NB locking needed now I'm racing with completion */
 891                         spin_lock_irqsave(&sched->ibs_lock, flags);
 892                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 893                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 894
 895                         /* cmid will be destroyed by CM(ofed) after cm_callback
 896                          * returned, so we can't refer it anymore
 897                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 898                         rdma_destroy_qp(conn->ibc_cmid);
 899                         conn->ibc_cmid = NULL;
 900
 901                         /* Drop my own and unused rxbuffer refcounts */
 902                         while (i++ <= IBLND_RX_MSGS(version))
 903                                 kiblnd_conn_decref(conn);
 904
 905                         return NULL;
 906                 }
 907         }
 908
 909         /* Init successful! */
 910         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 911                  state == IBLND_CONN_PASSIVE_WAIT);
 912         conn->ibc_state = state;
 913
 914         /* 1 more conn */
 915         cfs_atomic_inc(&net->ibn_nconns);
 916         return conn;
 917
 918  failed_2:
 919         kiblnd_destroy_conn(conn);
 920  failed_1:
 921         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 922  failed_0:
 923         return NULL;
 924 }
 925
 926 void
 927 kiblnd_destroy_conn (kib_conn_t *conn)
 928 {
 929         struct rdma_cm_id *cmid = conn->ibc_cmid;
 930         kib_peer_t        *peer = conn->ibc_peer;
 931         int                rc;
 932
 933         LASSERT (!cfs_in_interrupt());
 934         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 935         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 936         LASSERT (cfs_list_empty(&conn->ibc_tx_noops));
 937         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 938         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 939         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 940         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 941         LASSERT (conn->ibc_noops_posted == 0);
 942         LASSERT (conn->ibc_nsends_posted == 0);
 943
 944         switch (conn->ibc_state) {
 945         default:
 946                 /* conn must be completely disengaged from the network */
 947                 LBUG();
 948
 949         case IBLND_CONN_DISCONNECTED:
 950                 /* connvars should have been freed already */
 951                 LASSERT (conn->ibc_connvars == NULL);
 952                 break;
 953
 954         case IBLND_CONN_INIT:
 955                 break;
 956         }
 957
 958         /* conn->ibc_cmid might be destroyed by CM already */
 959         if (cmid != NULL && cmid->qp != NULL)
 960                 rdma_destroy_qp(cmid);
 961
 962         if (conn->ibc_cq != NULL) {
 963                 rc = ib_destroy_cq(conn->ibc_cq);
 964                 if (rc != 0)
 965                         CWARN("Error destroying CQ: %d\n", rc);
 966         }
 967
 968         if (conn->ibc_rx_pages != NULL)
 969                 kiblnd_unmap_rx_descs(conn);
 970
 971         if (conn->ibc_rxs != NULL) {
 972                 LIBCFS_FREE(conn->ibc_rxs,
 973                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 974         }
 975
 976         if (conn->ibc_connvars != NULL)
 977                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 978
 979         if (conn->ibc_hdev != NULL)
 980                 kiblnd_hdev_decref(conn->ibc_hdev);
 981
 982         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 983         if (conn->ibc_state != IBLND_CONN_INIT) {
 984                 kib_net_t *net = peer->ibp_ni->ni_data;
 985
 986                 kiblnd_peer_decref(peer);
 987                 rdma_destroy_id(cmid);
 988                 cfs_atomic_dec(&net->ibn_nconns);
 989         }
 990
 991         LIBCFS_FREE(conn, sizeof(*conn));
 992 }
 993
 994 int
 995 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 996 {
 997         kib_conn_t             *conn;
 998         cfs_list_t             *ctmp;
 999         cfs_list_t             *cnxt;
1000         int                     count = 0;
1001
1002         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1003                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1004
1005                 CDEBUG(D_NET, "Closing conn -> %s, "
1006                               "version: %x, reason: %d\n",
1007                        libcfs_nid2str(peer->ibp_nid),
1008                        conn->ibc_version, why);
1009
1010                 kiblnd_close_conn_locked(conn, why);
1011                 count++;
1012         }
1013
1014         return count;
1015 }
1016
1017 int
1018 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1019                                  int version, __u64 incarnation)
1020 {
1021         kib_conn_t             *conn;
1022         cfs_list_t             *ctmp;
1023         cfs_list_t             *cnxt;
1024         int                     count = 0;
1025
1026         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1027                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1028
1029                 if (conn->ibc_version     == version &&
1030                     conn->ibc_incarnation == incarnation)
1031                         continue;
1032
1033                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1034                               "incarnation:"LPX64"(%x, "LPX64")\n",
1035                        libcfs_nid2str(peer->ibp_nid),
1036                        conn->ibc_version, conn->ibc_incarnation,
1037                        version, incarnation);
1038
1039                 kiblnd_close_conn_locked(conn, -ESTALE);
1040                 count++;
1041         }
1042
1043         return count;
1044 }
1045
1046 int
1047 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1048 {
1049         kib_peer_t             *peer;
1050         cfs_list_t             *ptmp;
1051         cfs_list_t             *pnxt;
1052         int                     lo;
1053         int                     hi;
1054         int                     i;
1055         unsigned long           flags;
1056         int                     count = 0;
1057
1058         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1059
1060         if (nid != LNET_NID_ANY)
1061                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1062         else {
1063                 lo = 0;
1064                 hi = kiblnd_data.kib_peer_hash_size - 1;
1065         }
1066
1067         for (i = lo; i <= hi; i++) {
1068                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1069
1070                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1071                         LASSERT (peer->ibp_connecting > 0 ||
1072                                  peer->ibp_accepting > 0 ||
1073                                  !cfs_list_empty(&peer->ibp_conns));
1074
1075                         if (peer->ibp_ni != ni)
1076                                 continue;
1077
1078                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1079                                 continue;
1080
1081                         count += kiblnd_close_peer_conns_locked(peer, 0);
1082                 }
1083         }
1084
1085         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1086
1087         /* wildcards always succeed */
1088         if (nid == LNET_NID_ANY)
1089                 return 0;
1090
1091         return (count == 0) ? -ENOENT : 0;
1092 }
1093
1094 int
1095 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1096 {
1097         struct libcfs_ioctl_data *data = arg;
1098         int                       rc = -EINVAL;
1099
1100         switch(cmd) {
1101         case IOC_LIBCFS_GET_PEER: {
1102                 lnet_nid_t   nid = 0;
1103                 int          count = 0;
1104
1105                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1106                                           &nid, &count);
1107                 data->ioc_nid    = nid;
1108                 data->ioc_count  = count;
1109                 break;
1110         }
1111
1112         case IOC_LIBCFS_DEL_PEER: {
1113                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1114                 break;
1115         }
1116         case IOC_LIBCFS_GET_CONN: {
1117                 kib_conn_t *conn;
1118
1119                 rc = 0;
1120                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1121                 if (conn == NULL) {
1122                         rc = -ENOENT;
1123                         break;
1124                 }
1125
1126                 LASSERT (conn->ibc_cmid != NULL);
1127                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1128                 if (conn->ibc_cmid->route.path_rec == NULL)
1129                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1130                 else
1131                         data->ioc_u32[0] =
1132                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1133                 kiblnd_conn_decref(conn);
1134                 break;
1135         }
1136         case IOC_LIBCFS_CLOSE_CONNECTION: {
1137                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1138                 break;
1139         }
1140
1141         default:
1142                 break;
1143         }
1144
1145         return rc;
1146 }
1147
1148 void
1149 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1150 {
1151         cfs_time_t      last_alive = 0;
1152         cfs_time_t      now = cfs_time_current();
1153         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1154         kib_peer_t      *peer;
1155         unsigned long   flags;
1156
1157         read_lock_irqsave(glock, flags);
1158
1159         peer = kiblnd_find_peer_locked(nid);
1160         if (peer != NULL) {
1161                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1162                          peer->ibp_accepting > 0 ||
1163                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1164                 last_alive = peer->ibp_last_alive;
1165         }
1166
1167         read_unlock_irqrestore(glock, flags);
1168
1169         if (last_alive != 0)
1170                 *when = last_alive;
1171
1172         /* peer is not persistent in hash, trigger peer creation
1173          * and connection establishment with a NULL tx */
1174         if (peer == NULL)
1175                 kiblnd_launch_tx(ni, NULL, nid);
1176
1177         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1178                libcfs_nid2str(nid), peer,
1179                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1180         return;
1181 }
1182
1183 void
1184 kiblnd_free_pages(kib_pages_t *p)
1185 {
1186         int     npages = p->ibp_npages;
1187         int     i;
1188
1189         for (i = 0; i < npages; i++) {
1190                 if (p->ibp_pages[i] != NULL)
1191                         __free_page(p->ibp_pages[i]);
1192         }
1193
1194         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1195 }
1196
1197 int
1198 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1199 {
1200         kib_pages_t     *p;
1201         int             i;
1202
1203         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1204                          offsetof(kib_pages_t, ibp_pages[npages]));
1205         if (p == NULL) {
1206                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1207                 return -ENOMEM;
1208         }
1209
1210         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1211         p->ibp_npages = npages;
1212
1213         for (i = 0; i < npages; i++) {
1214                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1215                                                      __GFP_IO);
1216                 if (p->ibp_pages[i] == NULL) {
1217                         CERROR("Can't allocate page %d of %d\n", i, npages);
1218                         kiblnd_free_pages(p);
1219                         return -ENOMEM;
1220                 }
1221         }
1222
1223         *pp = p;
1224         return 0;
1225 }
1226
1227 void
1228 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1229 {
1230         kib_rx_t *rx;
1231         int       i;
1232
1233         LASSERT (conn->ibc_rxs != NULL);
1234         LASSERT (conn->ibc_hdev != NULL);
1235
1236         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1237                 rx = &conn->ibc_rxs[i];
1238
1239                 LASSERT (rx->rx_nob >= 0); /* not posted */
1240
1241                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1242                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1243                                                           rx->rx_msgaddr),
1244                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1245         }
1246
1247         kiblnd_free_pages(conn->ibc_rx_pages);
1248
1249         conn->ibc_rx_pages = NULL;
1250 }
1251
1252 void
1253 kiblnd_map_rx_descs(kib_conn_t *conn)
1254 {
1255         kib_rx_t       *rx;
1256         struct page    *pg;
1257         int             pg_off;
1258         int             ipg;
1259         int             i;
1260
1261         for (pg_off = ipg = i = 0;
1262              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1263                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1264                 rx = &conn->ibc_rxs[i];
1265
1266                 rx->rx_conn = conn;
1267                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1268
1269                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1270                                                        rx->rx_msg, IBLND_MSG_SIZE,
1271                                                        DMA_FROM_DEVICE);
1272                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1273                                                    rx->rx_msgaddr));
1274                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1275
1276                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1277                        i, rx->rx_msg, rx->rx_msgaddr,
1278                        lnet_page2phys(pg) + pg_off);
1279
1280                 pg_off += IBLND_MSG_SIZE;
1281                 LASSERT (pg_off <= PAGE_SIZE);
1282
1283                 if (pg_off == PAGE_SIZE) {
1284                         pg_off = 0;
1285                         ipg++;
1286                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1287                 }
1288         }
1289 }
1290
1291 static void
1292 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1293 {
1294         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1295         kib_tx_t       *tx;
1296         int             i;
1297
1298         LASSERT (tpo->tpo_pool.po_allocated == 0);
1299
1300         if (hdev == NULL)
1301                 return;
1302
1303         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1304                 tx = &tpo->tpo_tx_descs[i];
1305                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1306                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1307                                                           tx->tx_msgaddr),
1308                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1309         }
1310
1311         kiblnd_hdev_decref(hdev);
1312         tpo->tpo_hdev = NULL;
1313 }
1314
1315 static kib_hca_dev_t *
1316 kiblnd_current_hdev(kib_dev_t *dev)
1317 {
1318         kib_hca_dev_t *hdev;
1319         unsigned long  flags;
1320         int            i = 0;
1321
1322         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1323         while (dev->ibd_failover) {
1324                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1325                 if (i++ % 50 == 0)
1326                         CDEBUG(D_NET, "%s: Wait for failover\n",
1327                                dev->ibd_ifname);
1328                 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1329
1330                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1331         }
1332
1333         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1334         hdev = dev->ibd_hdev;
1335
1336         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1337
1338         return hdev;
1339 }
1340
1341 static void
1342 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1343 {
1344         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1345         kib_pool_t     *pool  = &tpo->tpo_pool;
1346         kib_net_t      *net   = pool->po_owner->ps_net;
1347         kib_dev_t      *dev;
1348         struct page    *page;
1349         kib_tx_t       *tx;
1350         int             page_offset;
1351         int             ipage;
1352         int             i;
1353
1354         LASSERT (net != NULL);
1355
1356         dev = net->ibn_dev;
1357
1358         /* pre-mapped messages are not bigger than 1 page */
1359         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1360
1361         /* No fancy arithmetic when we do the buffer calculations */
1362         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1363
1364         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1365
1366         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1367                 page = txpgs->ibp_pages[ipage];
1368                 tx = &tpo->tpo_tx_descs[i];
1369
1370                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1371                                            page_offset);
1372
1373                 tx->tx_msgaddr = kiblnd_dma_map_single(
1374                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1375                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1376                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1377                                                    tx->tx_msgaddr));
1378                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1379
1380                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1381
1382                 page_offset += IBLND_MSG_SIZE;
1383                 LASSERT (page_offset <= PAGE_SIZE);
1384
1385                 if (page_offset == PAGE_SIZE) {
1386                         page_offset = 0;
1387                         ipage++;
1388                         LASSERT (ipage <= txpgs->ibp_npages);
1389                 }
1390         }
1391 }
1392
1393 struct ib_mr *
1394 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1395 {
1396         __u64   index;
1397
1398         LASSERT (hdev->ibh_mrs[0] != NULL);
1399
1400         if (hdev->ibh_nmrs == 1)
1401                 return hdev->ibh_mrs[0];
1402
1403         index = addr >> hdev->ibh_mr_shift;
1404
1405         if (index <  hdev->ibh_nmrs &&
1406             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1407                 return hdev->ibh_mrs[index];
1408
1409         return NULL;
1410 }
1411
1412 struct ib_mr *
1413 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1414 {
1415         struct ib_mr *prev_mr;
1416         struct ib_mr *mr;
1417         int           i;
1418
1419         LASSERT (hdev->ibh_mrs[0] != NULL);
1420
1421         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1422             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1423                 return NULL;
1424
1425         if (hdev->ibh_nmrs == 1)
1426                 return hdev->ibh_mrs[0];
1427
1428         for (i = 0, mr = prev_mr = NULL;
1429              i < rd->rd_nfrags; i++) {
1430                 mr = kiblnd_find_dma_mr(hdev,
1431                                         rd->rd_frags[i].rf_addr,
1432                                         rd->rd_frags[i].rf_nob);
1433                 if (prev_mr == NULL)
1434                         prev_mr = mr;
1435
1436                 if (mr == NULL || prev_mr != mr) {
1437                         /* Can't covered by one single MR */
1438                         mr = NULL;
1439                         break;
1440                 }
1441         }
1442
1443         return mr;
1444 }
1445
1446 void
1447 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1448 {
1449         LASSERT (pool->fpo_map_count == 0);
1450
1451         if (pool->fpo_fmr_pool != NULL)
1452                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1453
1454         if (pool->fpo_hdev != NULL)
1455                 kiblnd_hdev_decref(pool->fpo_hdev);
1456
1457         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1458 }
1459
1460 void
1461 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1462 {
1463         kib_fmr_pool_t *pool;
1464
1465         while (!cfs_list_empty(head)) {
1466                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1467                 cfs_list_del(&pool->fpo_list);
1468                 kiblnd_destroy_fmr_pool(pool);
1469         }
1470 }
1471
1472 static int kiblnd_fmr_pool_size(int ncpts)
1473 {
1474         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1475
1476         return max(IBLND_FMR_POOL, size);
1477 }
1478
1479 static int kiblnd_fmr_flush_trigger(int ncpts)
1480 {
1481         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1482
1483         return max(IBLND_FMR_POOL_FLUSH, size);
1484 }
1485
1486 int
1487 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1488 {
1489         /* FMR pool for RDMA */
1490         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1491         kib_fmr_pool_t          *fpo;
1492         struct ib_fmr_pool_param param = {
1493                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1494                 .page_shift        = PAGE_SHIFT,
1495                 .access            = (IB_ACCESS_LOCAL_WRITE |
1496                                       IB_ACCESS_REMOTE_WRITE),
1497                 .pool_size         = fps->fps_pool_size,
1498                 .dirty_watermark   = fps->fps_flush_trigger,
1499                 .flush_function    = NULL,
1500                 .flush_arg         = NULL,
1501                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1502         int rc;
1503
1504         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1505         if (fpo == NULL)
1506                 return -ENOMEM;
1507
1508         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1509
1510         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1511         if (IS_ERR(fpo->fpo_fmr_pool)) {
1512                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1513                 CERROR("Failed to create FMR pool: %d\n", rc);
1514
1515                 kiblnd_hdev_decref(fpo->fpo_hdev);
1516                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1517                 return rc;
1518         }
1519
1520         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1521         fpo->fpo_owner    = fps;
1522         *pp_fpo = fpo;
1523
1524         return 0;
1525 }
1526
1527 static void
1528 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1529 {
1530         if (fps->fps_net == NULL) /* intialized? */
1531                 return;
1532
1533         spin_lock(&fps->fps_lock);
1534
1535         while (!cfs_list_empty(&fps->fps_pool_list)) {
1536                 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1537                                                  kib_fmr_pool_t, fpo_list);
1538                 fpo->fpo_failed = 1;
1539                 cfs_list_del(&fpo->fpo_list);
1540                 if (fpo->fpo_map_count == 0)
1541                         cfs_list_add(&fpo->fpo_list, zombies);
1542                 else
1543                         cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1544         }
1545
1546         spin_unlock(&fps->fps_lock);
1547 }
1548
1549 static void
1550 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1551 {
1552         if (fps->fps_net != NULL) { /* initialized? */
1553                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1554                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1555         }
1556 }
1557
1558 static int
1559 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1560                         int pool_size, int flush_trigger)
1561 {
1562         kib_fmr_pool_t *fpo;
1563         int             rc;
1564
1565         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1566
1567         fps->fps_net = net;
1568         fps->fps_cpt = cpt;
1569         fps->fps_pool_size = pool_size;
1570         fps->fps_flush_trigger = flush_trigger;
1571         spin_lock_init(&fps->fps_lock);
1572         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1573         CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1574
1575         rc = kiblnd_create_fmr_pool(fps, &fpo);
1576         if (rc == 0)
1577                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1578
1579         return rc;
1580 }
1581
1582 static int
1583 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1584 {
1585         if (fpo->fpo_map_count != 0) /* still in use */
1586                 return 0;
1587         if (fpo->fpo_failed)
1588                 return 1;
1589         return cfs_time_aftereq(now, fpo->fpo_deadline);
1590 }
1591
1592 void
1593 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1594 {
1595         CFS_LIST_HEAD     (zombies);
1596         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1597         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1598         cfs_time_t         now = cfs_time_current();
1599         kib_fmr_pool_t    *tmp;
1600         int                rc;
1601
1602         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1603         LASSERT (rc == 0);
1604
1605         if (status != 0) {
1606                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1607                 LASSERT (rc == 0);
1608         }
1609
1610         fmr->fmr_pool = NULL;
1611         fmr->fmr_pfmr = NULL;
1612
1613         spin_lock(&fps->fps_lock);
1614         fpo->fpo_map_count --;  /* decref the pool */
1615
1616         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1617                 /* the first pool is persistent */
1618                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1619                         continue;
1620
1621                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1622                         cfs_list_move(&fpo->fpo_list, &zombies);
1623                         fps->fps_version ++;
1624                 }
1625         }
1626         spin_unlock(&fps->fps_lock);
1627
1628         if (!cfs_list_empty(&zombies))
1629                 kiblnd_destroy_fmr_pool_list(&zombies);
1630 }
1631
1632 int
1633 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1634                     __u64 iov, kib_fmr_t *fmr)
1635 {
1636         struct ib_pool_fmr *pfmr;
1637         kib_fmr_pool_t     *fpo;
1638         __u64               version;
1639         int                 rc;
1640
1641  again:
1642         spin_lock(&fps->fps_lock);
1643         version = fps->fps_version;
1644         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1645                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1646                 fpo->fpo_map_count++;
1647                 spin_unlock(&fps->fps_lock);
1648
1649                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1650                                             pages, npages, iov);
1651                 if (likely(!IS_ERR(pfmr))) {
1652                         fmr->fmr_pool = fpo;
1653                         fmr->fmr_pfmr = pfmr;
1654                         return 0;
1655                 }
1656
1657                 spin_lock(&fps->fps_lock);
1658                 fpo->fpo_map_count--;
1659                 if (PTR_ERR(pfmr) != -EAGAIN) {
1660                         spin_unlock(&fps->fps_lock);
1661                         return PTR_ERR(pfmr);
1662                 }
1663
1664                 /* EAGAIN and ... */
1665                 if (version != fps->fps_version) {
1666                         spin_unlock(&fps->fps_lock);
1667                         goto again;
1668                 }
1669         }
1670
1671         if (fps->fps_increasing) {
1672                 spin_unlock(&fps->fps_lock);
1673                 CDEBUG(D_NET, "Another thread is allocating new "
1674                        "FMR pool, waiting for her to complete\n");
1675                 cfs_schedule();
1676                 goto again;
1677
1678         }
1679
1680         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1681                 /* someone failed recently */
1682                 spin_unlock(&fps->fps_lock);
1683                 return -EAGAIN;
1684         }
1685
1686         fps->fps_increasing = 1;
1687         spin_unlock(&fps->fps_lock);
1688
1689         CDEBUG(D_NET, "Allocate new FMR pool\n");
1690         rc = kiblnd_create_fmr_pool(fps, &fpo);
1691         spin_lock(&fps->fps_lock);
1692         fps->fps_increasing = 0;
1693         if (rc == 0) {
1694                 fps->fps_version++;
1695                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1696         } else {
1697                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1698         }
1699         spin_unlock(&fps->fps_lock);
1700
1701         goto again;
1702 }
1703
1704 static void
1705 kiblnd_fini_pool(kib_pool_t *pool)
1706 {
1707         LASSERT (cfs_list_empty(&pool->po_free_list));
1708         LASSERT (pool->po_allocated == 0);
1709
1710         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1711 }
1712
1713 static void
1714 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1715 {
1716         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1717
1718         memset(pool, 0, sizeof(kib_pool_t));
1719         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1720         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1721         pool->po_owner    = ps;
1722         pool->po_size     = size;
1723 }
1724
1725 void
1726 kiblnd_destroy_pool_list(cfs_list_t *head)
1727 {
1728         kib_pool_t *pool;
1729
1730         while (!cfs_list_empty(head)) {
1731                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1732                 cfs_list_del(&pool->po_list);
1733
1734                 LASSERT (pool->po_owner != NULL);
1735                 pool->po_owner->ps_pool_destroy(pool);
1736         }
1737 }
1738
1739 static void
1740 kiblnd_fail_poolset(kib_poolset_t *ps, cfs_list_t *zombies)
1741 {
1742         if (ps->ps_net == NULL) /* intialized? */
1743                 return;
1744
1745         spin_lock(&ps->ps_lock);
1746         while (!cfs_list_empty(&ps->ps_pool_list)) {
1747                 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1748                                             kib_pool_t, po_list);
1749                 po->po_failed = 1;
1750                 cfs_list_del(&po->po_list);
1751                 if (po->po_allocated == 0)
1752                         cfs_list_add(&po->po_list, zombies);
1753                 else
1754                         cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1755         }
1756         spin_unlock(&ps->ps_lock);
1757 }
1758
1759 static void
1760 kiblnd_fini_poolset(kib_poolset_t *ps)
1761 {
1762         if (ps->ps_net != NULL) { /* initialized? */
1763                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1764                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1765         }
1766 }
1767
1768 static int
1769 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1770                     kib_net_t *net, char *name, int size,
1771                     kib_ps_pool_create_t po_create,
1772                     kib_ps_pool_destroy_t po_destroy,
1773                     kib_ps_node_init_t nd_init,
1774                     kib_ps_node_fini_t nd_fini)
1775 {
1776         kib_pool_t      *pool;
1777         int             rc;
1778
1779         memset(ps, 0, sizeof(kib_poolset_t));
1780
1781         ps->ps_cpt          = cpt;
1782         ps->ps_net          = net;
1783         ps->ps_pool_create  = po_create;
1784         ps->ps_pool_destroy = po_destroy;
1785         ps->ps_node_init    = nd_init;
1786         ps->ps_node_fini    = nd_fini;
1787         ps->ps_pool_size    = size;
1788         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1789             >= sizeof(ps->ps_name))
1790                 return -E2BIG;
1791         spin_lock_init(&ps->ps_lock);
1792         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1793         CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1794
1795         rc = ps->ps_pool_create(ps, size, &pool);
1796         if (rc == 0)
1797                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1798         else
1799                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1800
1801         return rc;
1802 }
1803
1804 static int
1805 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1806 {
1807         if (pool->po_allocated != 0) /* still in use */
1808                 return 0;
1809         if (pool->po_failed)
1810                 return 1;
1811         return cfs_time_aftereq(now, pool->po_deadline);
1812 }
1813
1814 void
1815 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1816 {
1817         CFS_LIST_HEAD  (zombies);
1818         kib_poolset_t  *ps = pool->po_owner;
1819         kib_pool_t     *tmp;
1820         cfs_time_t      now = cfs_time_current();
1821
1822         spin_lock(&ps->ps_lock);
1823
1824         if (ps->ps_node_fini != NULL)
1825                 ps->ps_node_fini(pool, node);
1826
1827         LASSERT (pool->po_allocated > 0);
1828         cfs_list_add(node, &pool->po_free_list);
1829         pool->po_allocated --;
1830
1831         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1832                 /* the first pool is persistent */
1833                 if (ps->ps_pool_list.next == &pool->po_list)
1834                         continue;
1835
1836                 if (kiblnd_pool_is_idle(pool, now))
1837                         cfs_list_move(&pool->po_list, &zombies);
1838         }
1839         spin_unlock(&ps->ps_lock);
1840
1841         if (!cfs_list_empty(&zombies))
1842                 kiblnd_destroy_pool_list(&zombies);
1843 }
1844
1845 cfs_list_t *
1846 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1847 {
1848         cfs_list_t            *node;
1849         kib_pool_t            *pool;
1850         int                    rc;
1851
1852  again:
1853         spin_lock(&ps->ps_lock);
1854         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1855                 if (cfs_list_empty(&pool->po_free_list))
1856                         continue;
1857
1858                 pool->po_allocated ++;
1859                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1860                 node = pool->po_free_list.next;
1861                 cfs_list_del(node);
1862
1863                 if (ps->ps_node_init != NULL) {
1864                         /* still hold the lock */
1865                         ps->ps_node_init(pool, node);
1866                 }
1867                 spin_unlock(&ps->ps_lock);
1868                 return node;
1869         }
1870
1871         /* no available tx pool and ... */
1872         if (ps->ps_increasing) {
1873                 /* another thread is allocating a new pool */
1874                 spin_unlock(&ps->ps_lock);
1875                 CDEBUG(D_NET, "Another thread is allocating new "
1876                        "%s pool, waiting for her to complete\n",
1877                        ps->ps_name);
1878                 cfs_schedule();
1879                 goto again;
1880         }
1881
1882         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1883                 /* someone failed recently */
1884                 spin_unlock(&ps->ps_lock);
1885                 return NULL;
1886         }
1887
1888         ps->ps_increasing = 1;
1889         spin_unlock(&ps->ps_lock);
1890
1891         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1892
1893         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1894
1895         spin_lock(&ps->ps_lock);
1896         ps->ps_increasing = 0;
1897         if (rc == 0) {
1898                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1899         } else {
1900                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1901                 CERROR("Can't allocate new %s pool because out of memory\n",
1902                        ps->ps_name);
1903         }
1904         spin_unlock(&ps->ps_lock);
1905
1906         goto again;
1907 }
1908
1909 void
1910 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1911 {
1912         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1913         struct ib_mr        *mr  = pmr->pmr_mr;
1914
1915         pmr->pmr_mr = NULL;
1916         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1917         if (mr != NULL)
1918                 ib_dereg_mr(mr);
1919 }
1920
1921 int
1922 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1923                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1924 {
1925         kib_phys_mr_t *pmr;
1926         cfs_list_t    *node;
1927         int            rc;
1928         int            i;
1929
1930         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1931         if (node == NULL) {
1932                 CERROR("Failed to allocate PMR descriptor\n");
1933                 return -ENOMEM;
1934         }
1935
1936         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1937         if (pmr->pmr_pool->ppo_hdev != hdev) {
1938                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1939                 return -EAGAIN;
1940         }
1941
1942         for (i = 0; i < rd->rd_nfrags; i ++) {
1943                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1944                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1945         }
1946
1947         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1948                                      pmr->pmr_ipb, rd->rd_nfrags,
1949                                      IB_ACCESS_LOCAL_WRITE |
1950                                      IB_ACCESS_REMOTE_WRITE,
1951                                      iova);
1952         if (!IS_ERR(pmr->pmr_mr)) {
1953                 pmr->pmr_iova = *iova;
1954                 *pp_pmr = pmr;
1955                 return 0;
1956         }
1957
1958         rc = PTR_ERR(pmr->pmr_mr);
1959         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1960
1961         pmr->pmr_mr = NULL;
1962         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1963
1964         return rc;
1965 }
1966
1967 static void
1968 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1969 {
1970         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1971         kib_phys_mr_t  *pmr;
1972
1973         LASSERT (pool->po_allocated == 0);
1974
1975         while (!cfs_list_empty(&pool->po_free_list)) {
1976                 pmr = cfs_list_entry(pool->po_free_list.next,
1977                                      kib_phys_mr_t, pmr_list);
1978
1979                 LASSERT (pmr->pmr_mr == NULL);
1980                 cfs_list_del(&pmr->pmr_list);
1981
1982                 if (pmr->pmr_ipb != NULL) {
1983                         LIBCFS_FREE(pmr->pmr_ipb,
1984                                     IBLND_MAX_RDMA_FRAGS *
1985                                     sizeof(struct ib_phys_buf));
1986                 }
1987
1988                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1989         }
1990
1991         kiblnd_fini_pool(pool);
1992         if (ppo->ppo_hdev != NULL)
1993                 kiblnd_hdev_decref(ppo->ppo_hdev);
1994
1995         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1996 }
1997
1998 static inline int kiblnd_pmr_pool_size(int ncpts)
1999 {
2000         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
2001
2002         return max(IBLND_PMR_POOL, size);
2003 }
2004
2005 static int
2006 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2007 {
2008         struct kib_pmr_pool     *ppo;
2009         struct kib_pool         *pool;
2010         kib_phys_mr_t           *pmr;
2011         int                     i;
2012
2013         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2014                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2015         if (ppo == NULL) {
2016                 CERROR("Failed to allocate PMR pool\n");
2017                 return -ENOMEM;
2018         }
2019
2020         pool = &ppo->ppo_pool;
2021         kiblnd_init_pool(ps, pool, size);
2022
2023         for (i = 0; i < size; i++) {
2024                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2025                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2026                 if (pmr == NULL)
2027                         break;
2028
2029                 pmr->pmr_pool = ppo;
2030                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2031                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2032                 if (pmr->pmr_ipb == NULL)
2033                         break;
2034
2035                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
2036         }
2037
2038         if (i < size) {
2039                 ps->ps_pool_destroy(pool);
2040                 return -ENOMEM;
2041         }
2042
2043         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2044         *pp_po = pool;
2045         return 0;
2046 }
2047
2048 static void
2049 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2050 {
2051         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2052         int             i;
2053
2054         LASSERT (pool->po_allocated == 0);
2055
2056         if (tpo->tpo_tx_pages != NULL) {
2057                 kiblnd_unmap_tx_pool(tpo);
2058                 kiblnd_free_pages(tpo->tpo_tx_pages);
2059         }
2060
2061         if (tpo->tpo_tx_descs == NULL)
2062                 goto out;
2063
2064         for (i = 0; i < pool->po_size; i++) {
2065                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2066
2067                 cfs_list_del(&tx->tx_list);
2068                 if (tx->tx_pages != NULL)
2069                         LIBCFS_FREE(tx->tx_pages,
2070                                     LNET_MAX_IOV *
2071                                     sizeof(*tx->tx_pages));
2072                 if (tx->tx_frags != NULL)
2073                         LIBCFS_FREE(tx->tx_frags,
2074                                     IBLND_MAX_RDMA_FRAGS *
2075                                             sizeof(*tx->tx_frags));
2076                 if (tx->tx_wrq != NULL)
2077                         LIBCFS_FREE(tx->tx_wrq,
2078                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2079                                     sizeof(*tx->tx_wrq));
2080                 if (tx->tx_sge != NULL)
2081                         LIBCFS_FREE(tx->tx_sge,
2082                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2083                                     sizeof(*tx->tx_sge));
2084                 if (tx->tx_rd != NULL)
2085                         LIBCFS_FREE(tx->tx_rd,
2086                                     offsetof(kib_rdma_desc_t,
2087                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2088         }
2089
2090         LIBCFS_FREE(tpo->tpo_tx_descs,
2091                     pool->po_size * sizeof(kib_tx_t));
2092 out:
2093         kiblnd_fini_pool(pool);
2094         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2095 }
2096
2097 static int kiblnd_tx_pool_size(int ncpts)
2098 {
2099         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2100
2101         return max(IBLND_TX_POOL, ntx);
2102 }
2103
2104 static int
2105 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2106 {
2107         int            i;
2108         int            npg;
2109         kib_pool_t    *pool;
2110         kib_tx_pool_t *tpo;
2111
2112         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2113         if (tpo == NULL) {
2114                 CERROR("Failed to allocate TX pool\n");
2115                 return -ENOMEM;
2116         }
2117
2118         pool = &tpo->tpo_pool;
2119         kiblnd_init_pool(ps, pool, size);
2120         tpo->tpo_tx_descs = NULL;
2121         tpo->tpo_tx_pages = NULL;
2122
2123         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2124         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2125                 CERROR("Can't allocate tx pages: %d\n", npg);
2126                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2127                 return -ENOMEM;
2128         }
2129
2130         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2131                          size * sizeof(kib_tx_t));
2132         if (tpo->tpo_tx_descs == NULL) {
2133                 CERROR("Can't allocate %d tx descriptors\n", size);
2134                 ps->ps_pool_destroy(pool);
2135                 return -ENOMEM;
2136         }
2137
2138         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2139
2140         for (i = 0; i < size; i++) {
2141                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2142
2143                 tx->tx_pool = tpo;
2144                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2145                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2146                                          lnet_cpt_table(), ps->ps_cpt,
2147                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2148                         if (tx->tx_pages == NULL)
2149                                 break;
2150                 }
2151
2152                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2153                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2154                 if (tx->tx_frags == NULL)
2155                         break;
2156
2157                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2158
2159                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2160                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2161                                  sizeof(*tx->tx_wrq));
2162                 if (tx->tx_wrq == NULL)
2163                         break;
2164
2165                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2166                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2167                                  sizeof(*tx->tx_sge));
2168                 if (tx->tx_sge == NULL)
2169                         break;
2170
2171                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2172                                  offsetof(kib_rdma_desc_t,
2173                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2174                 if (tx->tx_rd == NULL)
2175                         break;
2176         }
2177
2178         if (i == size) {
2179                 kiblnd_map_tx_pool(tpo);
2180                 *pp_po = pool;
2181                 return 0;
2182         }
2183
2184         ps->ps_pool_destroy(pool);
2185         return -ENOMEM;
2186 }
2187
2188 static void
2189 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2190 {
2191         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2192                                              tps_poolset);
2193         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
2194
2195         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2196 }
2197
2198 void
2199 kiblnd_net_fini_pools(kib_net_t *net)
2200 {
2201         int     i;
2202
2203         cfs_cpt_for_each(i, lnet_cpt_table()) {
2204                 kib_tx_poolset_t        *tps;
2205                 kib_fmr_poolset_t       *fps;
2206                 kib_pmr_poolset_t       *pps;
2207
2208                 if (net->ibn_tx_ps != NULL) {
2209                         tps = net->ibn_tx_ps[i];
2210                         kiblnd_fini_poolset(&tps->tps_poolset);
2211                 }
2212
2213                 if (net->ibn_fmr_ps != NULL) {
2214                         fps = net->ibn_fmr_ps[i];
2215                         kiblnd_fini_fmr_poolset(fps);
2216                 }
2217
2218                 if (net->ibn_pmr_ps != NULL) {
2219                         pps = net->ibn_pmr_ps[i];
2220                         kiblnd_fini_poolset(&pps->pps_poolset);
2221                 }
2222         }
2223
2224         if (net->ibn_tx_ps != NULL) {
2225                 cfs_percpt_free(net->ibn_tx_ps);
2226                 net->ibn_tx_ps = NULL;
2227         }
2228
2229         if (net->ibn_fmr_ps != NULL) {
2230                 cfs_percpt_free(net->ibn_fmr_ps);
2231                 net->ibn_fmr_ps = NULL;
2232         }
2233
2234         if (net->ibn_pmr_ps != NULL) {
2235                 cfs_percpt_free(net->ibn_pmr_ps);
2236                 net->ibn_pmr_ps = NULL;
2237         }
2238 }
2239
2240 int
2241 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2242 {
2243         unsigned long   flags;
2244         int             cpt;
2245         int             rc;
2246         int             i;
2247
2248         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2249         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2250             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2251                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2252                                            flags);
2253                 goto create_tx_pool;
2254         }
2255
2256         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2257
2258         if (*kiblnd_tunables.kib_fmr_pool_size <
2259             *kiblnd_tunables.kib_ntx / 4) {
2260                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2261                        *kiblnd_tunables.kib_fmr_pool_size,
2262                        *kiblnd_tunables.kib_ntx / 4);
2263                 rc = -EINVAL;
2264                 goto failed;
2265         }
2266
2267         /* TX pool must be created later than FMR/PMR, see LU-2268
2268          * for details */
2269         LASSERT(net->ibn_tx_ps == NULL);
2270
2271         /* premapping can fail if ibd_nmr > 1, so we always create
2272          * FMR/PMR pool and map-on-demand if premapping failed */
2273
2274         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2275                                            sizeof(kib_fmr_poolset_t));
2276         if (net->ibn_fmr_ps == NULL) {
2277                 CERROR("Failed to allocate FMR pool array\n");
2278                 rc = -ENOMEM;
2279                 goto failed;
2280         }
2281
2282         for (i = 0; i < ncpts; i++) {
2283                 cpt = (cpts == NULL) ? i : cpts[i];
2284                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2285                                              kiblnd_fmr_pool_size(ncpts),
2286                                              kiblnd_fmr_flush_trigger(ncpts));
2287                 if (rc == -ENOSYS && i == 0) /* no FMR */
2288                         break; /* create PMR pool */
2289
2290                 if (rc != 0) { /* a real error */
2291                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2292                                cpt, rc);
2293                         goto failed;
2294                 }
2295         }
2296
2297         if (i > 0) {
2298                 LASSERT(i == ncpts);
2299                 goto create_tx_pool;
2300         }
2301
2302         cfs_percpt_free(net->ibn_fmr_ps);
2303         net->ibn_fmr_ps = NULL;
2304
2305         CWARN("Device does not support FMR, failing back to PMR\n");
2306
2307         if (*kiblnd_tunables.kib_pmr_pool_size <
2308             *kiblnd_tunables.kib_ntx / 4) {
2309                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2310                        *kiblnd_tunables.kib_pmr_pool_size,
2311                        *kiblnd_tunables.kib_ntx / 4);
2312                 rc = -EINVAL;
2313                 goto failed;
2314         }
2315
2316         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2317                                            sizeof(kib_pmr_poolset_t));
2318         if (net->ibn_pmr_ps == NULL) {
2319                 CERROR("Failed to allocate PMR pool array\n");
2320                 rc = -ENOMEM;
2321                 goto failed;
2322         }
2323
2324         for (i = 0; i < ncpts; i++) {
2325                 cpt = (cpts == NULL) ? i : cpts[i];
2326                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2327                                          cpt, net, "PMR",
2328                                          kiblnd_pmr_pool_size(ncpts),
2329                                          kiblnd_create_pmr_pool,
2330                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2331                 if (rc != 0) {
2332                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2333                                cpt, rc);
2334                         goto failed;
2335                 }
2336         }
2337
2338  create_tx_pool:
2339         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2340                                           sizeof(kib_tx_poolset_t));
2341         if (net->ibn_tx_ps == NULL) {
2342                 CERROR("Failed to allocate tx pool array\n");
2343                 rc = -ENOMEM;
2344                 goto failed;
2345         }
2346
2347         for (i = 0; i < ncpts; i++) {
2348                 cpt = (cpts == NULL) ? i : cpts[i];
2349                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2350                                          cpt, net, "TX",
2351                                          kiblnd_tx_pool_size(ncpts),
2352                                          kiblnd_create_tx_pool,
2353                                          kiblnd_destroy_tx_pool,
2354                                          kiblnd_tx_init, NULL);
2355                 if (rc != 0) {
2356                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2357                                cpt, rc);
2358                         goto failed;
2359                 }
2360         }
2361
2362         return 0;
2363  failed:
2364         kiblnd_net_fini_pools(net);
2365         LASSERT(rc != 0);
2366         return rc;
2367 }
2368
2369 static int
2370 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2371 {
2372         struct ib_device_attr *attr;
2373         int                    rc;
2374
2375         /* It's safe to assume a HCA can handle a page size
2376          * matching that of the native system */
2377         hdev->ibh_page_shift = PAGE_SHIFT;
2378         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2379         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2380
2381         LIBCFS_ALLOC(attr, sizeof(*attr));
2382         if (attr == NULL) {
2383                 CERROR("Out of memory\n");
2384                 return -ENOMEM;
2385         }
2386
2387         rc = ib_query_device(hdev->ibh_ibdev, attr);
2388         if (rc == 0)
2389                 hdev->ibh_mr_size = attr->max_mr_size;
2390
2391         LIBCFS_FREE(attr, sizeof(*attr));
2392
2393         if (rc != 0) {
2394                 CERROR("Failed to query IB device: %d\n", rc);
2395                 return rc;
2396         }
2397
2398         if (hdev->ibh_mr_size == ~0ULL) {
2399                 hdev->ibh_mr_shift = 64;
2400                 return 0;
2401         }
2402
2403         for (hdev->ibh_mr_shift = 0;
2404              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2405                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2406                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2407                         return 0;
2408         }
2409
2410         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2411         return -EINVAL;
2412 }
2413
2414 void
2415 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2416 {
2417         int     i;
2418
2419         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2420                 return;
2421
2422         for (i = 0; i < hdev->ibh_nmrs; i++) {
2423                 if (hdev->ibh_mrs[i] == NULL)
2424                         break;
2425
2426                 ib_dereg_mr(hdev->ibh_mrs[i]);
2427         }
2428
2429         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2430         hdev->ibh_mrs  = NULL;
2431         hdev->ibh_nmrs = 0;
2432 }
2433
2434 void
2435 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2436 {
2437         kiblnd_hdev_cleanup_mrs(hdev);
2438
2439         if (hdev->ibh_pd != NULL)
2440                 ib_dealloc_pd(hdev->ibh_pd);
2441
2442         if (hdev->ibh_cmid != NULL)
2443                 rdma_destroy_id(hdev->ibh_cmid);
2444
2445         LIBCFS_FREE(hdev, sizeof(*hdev));
2446 }
2447
2448 int
2449 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2450 {
2451         struct ib_mr *mr;
2452         int           i;
2453         int           rc;
2454         __u64         mm_size;
2455         __u64         mr_size;
2456         int           acflags = IB_ACCESS_LOCAL_WRITE |
2457                                 IB_ACCESS_REMOTE_WRITE;
2458
2459         rc = kiblnd_hdev_get_attr(hdev);
2460         if (rc != 0)
2461                 return rc;
2462
2463         if (hdev->ibh_mr_shift == 64) {
2464                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2465                 if (hdev->ibh_mrs == NULL) {
2466                         CERROR("Failed to allocate MRs table\n");
2467                         return -ENOMEM;
2468                 }
2469
2470                 hdev->ibh_mrs[0] = NULL;
2471                 hdev->ibh_nmrs   = 1;
2472
2473                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2474                 if (IS_ERR(mr)) {
2475                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2476                         kiblnd_hdev_cleanup_mrs(hdev);
2477                         return PTR_ERR(mr);
2478                 }
2479
2480                 hdev->ibh_mrs[0] = mr;
2481
2482                 goto out;
2483         }
2484
2485         mr_size = (1ULL << hdev->ibh_mr_shift);
2486         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2487
2488         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2489
2490         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2491                 /* it's 4T..., assume we will re-code at that time */
2492                 CERROR("Can't support memory size: x"LPX64
2493                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2494                 return -EINVAL;
2495         }
2496
2497         /* create an array of MRs to cover all memory */
2498         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2499         if (hdev->ibh_mrs == NULL) {
2500                 CERROR("Failed to allocate MRs' table\n");
2501                 return -ENOMEM;
2502         }
2503
2504         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2505
2506         for (i = 0; i < hdev->ibh_nmrs; i++) {
2507                 struct ib_phys_buf ipb;
2508                 __u64              iova;
2509
2510                 ipb.size = hdev->ibh_mr_size;
2511                 ipb.addr = i * mr_size;
2512                 iova     = ipb.addr;
2513
2514                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2515                 if (IS_ERR(mr)) {
2516                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2517                                " size "LPX64" : %ld\n",
2518                                ipb.addr, ipb.size, PTR_ERR(mr));
2519                         kiblnd_hdev_cleanup_mrs(hdev);
2520                         return PTR_ERR(mr);
2521                 }
2522
2523                 LASSERT (iova == ipb.addr);
2524
2525                 hdev->ibh_mrs[i] = mr;
2526         }
2527
2528 out:
2529         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2530                 LCONSOLE_INFO("Register global MR array, MR size: "
2531                               LPX64", array size: %d\n",
2532                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2533         return 0;
2534 }
2535
2536 static int
2537 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2538 {       /* DUMMY */
2539         return 0;
2540 }
2541
2542 static int
2543 kiblnd_dev_need_failover(kib_dev_t *dev)
2544 {
2545         struct rdma_cm_id  *cmid;
2546         struct sockaddr_in  srcaddr;
2547         struct sockaddr_in  dstaddr;
2548         int                 rc;
2549
2550         if (dev->ibd_hdev == NULL || /* initializing */
2551             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2552             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2553                 return 1;
2554
2555         /* XXX: it's UGLY, but I don't have better way to find
2556          * ib-bonding HCA failover because:
2557          *
2558          * a. no reliable CM event for HCA failover...
2559          * b. no OFED API to get ib_device for current net_device...
2560          *
2561          * We have only two choices at this point:
2562          *
2563          * a. rdma_bind_addr(), it will conflict with listener cmid
2564          * b. rdma_resolve_addr() to zero addr */
2565         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2566                                      IB_QPT_RC);
2567         if (IS_ERR(cmid)) {
2568                 rc = PTR_ERR(cmid);
2569                 CERROR("Failed to create cmid for failover: %d\n", rc);
2570                 return rc;
2571         }
2572
2573         memset(&srcaddr, 0, sizeof(srcaddr));
2574         srcaddr.sin_family      = AF_INET;
2575         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2576
2577         memset(&dstaddr, 0, sizeof(dstaddr));
2578         dstaddr.sin_family = AF_INET;
2579         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2580                                (struct sockaddr *)&dstaddr, 1);
2581         if (rc != 0 || cmid->device == NULL) {
2582                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2583                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2584                        cmid->device, rc);
2585                 rdma_destroy_id(cmid);
2586                 return rc;
2587         }
2588
2589         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2590                 /* don't need device failover */
2591                 rdma_destroy_id(cmid);
2592                 return 0;
2593         }
2594
2595         return 1;
2596 }
2597
2598 int
2599 kiblnd_dev_failover(kib_dev_t *dev)
2600 {
2601         CFS_LIST_HEAD      (zombie_tpo);
2602         CFS_LIST_HEAD      (zombie_ppo);
2603         CFS_LIST_HEAD      (zombie_fpo);
2604         struct rdma_cm_id  *cmid  = NULL;
2605         kib_hca_dev_t      *hdev  = NULL;
2606         kib_hca_dev_t      *old;
2607         struct ib_pd       *pd;
2608         kib_net_t          *net;
2609         struct sockaddr_in  addr;
2610         unsigned long       flags;
2611         int                 rc = 0;
2612         int                 i;
2613
2614         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2615                  dev->ibd_can_failover ||
2616                  dev->ibd_hdev == NULL);
2617
2618         rc = kiblnd_dev_need_failover(dev);
2619         if (rc <= 0)
2620                 goto out;
2621
2622         if (dev->ibd_hdev != NULL &&
2623             dev->ibd_hdev->ibh_cmid != NULL) {
2624                 /* XXX it's not good to close old listener at here,
2625                  * because we can fail to create new listener.
2626                  * But we have to close it now, otherwise rdma_bind_addr
2627                  * will return EADDRINUSE... How crap! */
2628                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2629
2630                 cmid = dev->ibd_hdev->ibh_cmid;
2631                 /* make next schedule of kiblnd_dev_need_failover()
2632                  * return 1 for me */
2633                 dev->ibd_hdev->ibh_cmid  = NULL;
2634                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2635
2636                 rdma_destroy_id(cmid);
2637         }
2638
2639         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2640                                      IB_QPT_RC);
2641         if (IS_ERR(cmid)) {
2642                 rc = PTR_ERR(cmid);
2643                 CERROR("Failed to create cmid for failover: %d\n", rc);
2644                 goto out;
2645         }
2646
2647         memset(&addr, 0, sizeof(addr));
2648         addr.sin_family      = AF_INET;
2649         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2650         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2651
2652         /* Bind to failover device or port */
2653         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2654         if (rc != 0 || cmid->device == NULL) {
2655                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2656                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2657                        cmid->device, rc);
2658                 rdma_destroy_id(cmid);
2659                 goto out;
2660         }
2661
2662         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2663         if (hdev == NULL) {
2664                 CERROR("Failed to allocate kib_hca_dev\n");
2665                 rdma_destroy_id(cmid);
2666                 rc = -ENOMEM;
2667                 goto out;
2668         }
2669
2670         atomic_set(&hdev->ibh_ref, 1);
2671         hdev->ibh_dev   = dev;
2672         hdev->ibh_cmid  = cmid;
2673         hdev->ibh_ibdev = cmid->device;
2674
2675         pd = ib_alloc_pd(cmid->device);
2676         if (IS_ERR(pd)) {
2677                 rc = PTR_ERR(pd);
2678                 CERROR("Can't allocate PD: %d\n", rc);
2679                 goto out;
2680         }
2681
2682         hdev->ibh_pd = pd;
2683
2684         rc = rdma_listen(cmid, 0);
2685         if (rc != 0) {
2686                 CERROR("Can't start new listener: %d\n", rc);
2687                 goto out;
2688         }
2689
2690         rc = kiblnd_hdev_setup_mrs(hdev);
2691         if (rc != 0) {
2692                 CERROR("Can't setup device: %d\n", rc);
2693                 goto out;
2694         }
2695
2696         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2697
2698         old = dev->ibd_hdev;
2699         dev->ibd_hdev = hdev; /* take over the refcount */
2700         hdev = old;
2701
2702         cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2703                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2704                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2705                                             &zombie_tpo);
2706
2707                         if (net->ibn_fmr_ps != NULL) {
2708                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2709                                                         &zombie_fpo);
2710
2711                         } else if (net->ibn_pmr_ps != NULL) {
2712                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2713                                                     pps_poolset, &zombie_ppo);
2714                         }
2715                 }
2716         }
2717
2718         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2719  out:
2720         if (!cfs_list_empty(&zombie_tpo))
2721                 kiblnd_destroy_pool_list(&zombie_tpo);
2722         if (!cfs_list_empty(&zombie_ppo))
2723                 kiblnd_destroy_pool_list(&zombie_ppo);
2724         if (!cfs_list_empty(&zombie_fpo))
2725                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2726         if (hdev != NULL)
2727                 kiblnd_hdev_decref(hdev);
2728
2729         if (rc != 0)
2730                 dev->ibd_failed_failover++;
2731         else
2732                 dev->ibd_failed_failover = 0;
2733
2734         return rc;
2735 }
2736
2737 void
2738 kiblnd_destroy_dev (kib_dev_t *dev)
2739 {
2740         LASSERT (dev->ibd_nnets == 0);
2741         LASSERT (cfs_list_empty(&dev->ibd_nets));
2742
2743         cfs_list_del(&dev->ibd_fail_list);
2744         cfs_list_del(&dev->ibd_list);
2745
2746         if (dev->ibd_hdev != NULL)
2747                 kiblnd_hdev_decref(dev->ibd_hdev);
2748
2749         LIBCFS_FREE(dev, sizeof(*dev));
2750 }
2751
2752 kib_dev_t *
2753 kiblnd_create_dev(char *ifname)
2754 {
2755         struct net_device *netdev;
2756         kib_dev_t         *dev;
2757         __u32              netmask;
2758         __u32              ip;
2759         int                up;
2760         int                rc;
2761
2762         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2763         if (rc != 0) {
2764                 CERROR("Can't query IPoIB interface %s: %d\n",
2765                        ifname, rc);
2766                 return NULL;
2767         }
2768
2769         if (!up) {
2770                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2771                 return NULL;
2772         }
2773
2774         LIBCFS_ALLOC(dev, sizeof(*dev));
2775         if (dev == NULL)
2776                 return NULL;
2777
2778         memset(dev, 0, sizeof(*dev));
2779 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2780         netdev = dev_get_by_name(&init_net, ifname);
2781 #else
2782         netdev = dev_get_by_name(ifname);
2783 #endif
2784         if (netdev == NULL) {
2785                 dev->ibd_can_failover = 0;
2786         } else {
2787                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2788                 dev_put(netdev);
2789         }
2790
2791         CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2792         CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2793         CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2794         dev->ibd_ifip = ip;
2795         strcpy(&dev->ibd_ifname[0], ifname);
2796
2797         /* initialize the device */
2798         rc = kiblnd_dev_failover(dev);
2799         if (rc != 0) {
2800                 CERROR("Can't initialize device: %d\n", rc);
2801                 LIBCFS_FREE(dev, sizeof(*dev));
2802                 return NULL;
2803         }
2804
2805         cfs_list_add_tail(&dev->ibd_list,
2806                           &kiblnd_data.kib_devs);
2807         return dev;
2808 }
2809
2810 void
2811 kiblnd_base_shutdown(void)
2812 {
2813         struct kib_sched_info   *sched;
2814         int                     i;
2815
2816         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2817
2818         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2819                cfs_atomic_read(&libcfs_kmemory));
2820
2821         switch (kiblnd_data.kib_init) {
2822         default:
2823                 LBUG();
2824
2825         case IBLND_INIT_ALL:
2826         case IBLND_INIT_DATA:
2827                 LASSERT (kiblnd_data.kib_peers != NULL);
2828                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2829                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2830                 }
2831                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2832                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2833
2834                 /* flag threads to terminate; wake and wait for them to die */
2835                 kiblnd_data.kib_shutdown = 1;
2836
2837                 /* NB: we really want to stop scheduler threads net by net
2838                  * instead of the whole module, this should be improved
2839                  * with dynamic configuration LNet */
2840                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2841                         cfs_waitq_broadcast(&sched->ibs_waitq);
2842
2843                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2844                 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2845
2846                 i = 2;
2847                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2848                         i++;
2849                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2850                                "Waiting for %d threads to terminate\n",
2851                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2852                         cfs_pause(cfs_time_seconds(1));
2853                 }
2854
2855                 /* fall through */
2856
2857         case IBLND_INIT_NOTHING:
2858                 break;
2859         }
2860
2861         if (kiblnd_data.kib_peers != NULL) {
2862                 LIBCFS_FREE(kiblnd_data.kib_peers,
2863                             sizeof(cfs_list_t) *
2864                             kiblnd_data.kib_peer_hash_size);
2865         }
2866
2867         if (kiblnd_data.kib_scheds != NULL)
2868                 cfs_percpt_free(kiblnd_data.kib_scheds);
2869
2870         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2871                cfs_atomic_read(&libcfs_kmemory));
2872
2873         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2874         PORTAL_MODULE_UNUSE;
2875 }
2876
2877 void
2878 kiblnd_shutdown (lnet_ni_t *ni)
2879 {
2880         kib_net_t        *net = ni->ni_data;
2881         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2882         int               i;
2883         unsigned long     flags;
2884
2885         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2886
2887         if (net == NULL)
2888                 goto out;
2889
2890         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2891                cfs_atomic_read(&libcfs_kmemory));
2892
2893         write_lock_irqsave(g_lock, flags);
2894         net->ibn_shutdown = 1;
2895         write_unlock_irqrestore(g_lock, flags);
2896
2897         switch (net->ibn_init) {
2898         default:
2899                 LBUG();
2900
2901         case IBLND_INIT_ALL:
2902                 /* nuke all existing peers within this net */
2903                 kiblnd_del_peer(ni, LNET_NID_ANY);
2904
2905                 /* Wait for all peer state to clean up */
2906                 i = 2;
2907                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2908                         i++;
2909                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2910                                "%s: waiting for %d peers to disconnect\n",
2911                                libcfs_nid2str(ni->ni_nid),
2912                                cfs_atomic_read(&net->ibn_npeers));
2913                         cfs_pause(cfs_time_seconds(1));
2914                 }
2915
2916                 kiblnd_net_fini_pools(net);
2917
2918                 write_lock_irqsave(g_lock, flags);
2919                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2920                 net->ibn_dev->ibd_nnets--;
2921                 cfs_list_del(&net->ibn_list);
2922                 write_unlock_irqrestore(g_lock, flags);
2923
2924                 /* fall through */
2925
2926         case IBLND_INIT_NOTHING:
2927                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2928
2929                 if (net->ibn_dev != NULL &&
2930                     net->ibn_dev->ibd_nnets == 0)
2931                         kiblnd_destroy_dev(net->ibn_dev);
2932
2933                 break;
2934         }
2935
2936         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2937                cfs_atomic_read(&libcfs_kmemory));
2938
2939         net->ibn_init = IBLND_INIT_NOTHING;
2940         ni->ni_data = NULL;
2941
2942         LIBCFS_FREE(net, sizeof(*net));
2943
2944 out:
2945         if (cfs_list_empty(&kiblnd_data.kib_devs))
2946                 kiblnd_base_shutdown();
2947         return;
2948 }
2949
2950 int
2951 kiblnd_base_startup(void)
2952 {
2953         struct kib_sched_info   *sched;
2954         int                     rc;
2955         int                     i;
2956
2957         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2958
2959         PORTAL_MODULE_USE;
2960         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2961
2962         rwlock_init(&kiblnd_data.kib_global_lock);
2963
2964         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2965         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2966
2967         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2968         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2969                      sizeof(cfs_list_t) *
2970                             kiblnd_data.kib_peer_hash_size);
2971         if (kiblnd_data.kib_peers == NULL) {
2972                 goto failed;
2973         }
2974         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2975                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2976
2977         spin_lock_init(&kiblnd_data.kib_connd_lock);
2978         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2979         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2980         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2981         cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2982
2983         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2984                                                   sizeof(*sched));
2985         if (kiblnd_data.kib_scheds == NULL)
2986                 goto failed;
2987
2988         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2989                 int     nthrs;
2990
2991                 spin_lock_init(&sched->ibs_lock);
2992                 CFS_INIT_LIST_HEAD(&sched->ibs_conns);
2993                 cfs_waitq_init(&sched->ibs_waitq);
2994
2995                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2996                 if (*kiblnd_tunables.kib_nscheds > 0) {
2997                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2998                 } else {
2999                         /* max to half of CPUs, another half is reserved for
3000                          * upper layer modules */
3001                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3002                 }
3003
3004                 sched->ibs_nthreads_max = nthrs;
3005                 sched->ibs_cpt = i;
3006         }
3007
3008         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3009
3010         /* lists/ptrs/locks initialised */
3011         kiblnd_data.kib_init = IBLND_INIT_DATA;
3012         /*****************************************************/
3013
3014         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3015         if (rc != 0) {
3016                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3017                 goto failed;
3018         }
3019
3020         if (*kiblnd_tunables.kib_dev_failover != 0)
3021                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3022                                          "kiblnd_failover");
3023
3024         if (rc != 0) {
3025                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3026                 goto failed;
3027         }
3028
3029         /* flag everything initialised */
3030         kiblnd_data.kib_init = IBLND_INIT_ALL;
3031         /*****************************************************/
3032
3033         return 0;
3034
3035  failed:
3036         kiblnd_base_shutdown();
3037         return -ENETDOWN;
3038 }
3039
3040 int
3041 kiblnd_start_schedulers(struct kib_sched_info *sched)
3042 {
3043         int     rc = 0;
3044         int     nthrs;
3045         int     i;
3046
3047         if (sched->ibs_nthreads == 0) {
3048                 if (*kiblnd_tunables.kib_nscheds > 0) {
3049                         nthrs = sched->ibs_nthreads_max;
3050                 } else {
3051                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3052                                                sched->ibs_cpt);
3053                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3054                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3055                 }
3056         } else {
3057                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3058                 /* increase one thread if there is new interface */
3059                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3060         }
3061
3062         for (i = 0; i < nthrs; i++) {
3063                 long    id;
3064                 char    name[20];
3065                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3066                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3067                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3068                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3069                 if (rc == 0)
3070                         continue;
3071
3072                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3073                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3074                 break;
3075         }
3076
3077         sched->ibs_nthreads += i;
3078         return rc;
3079 }
3080
3081 int
3082 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3083 {
3084         int     cpt;
3085         int     rc;
3086         int     i;
3087
3088         for (i = 0; i < ncpts; i++) {
3089                 struct kib_sched_info *sched;
3090
3091                 cpt = (cpts == NULL) ? i : cpts[i];
3092                 sched = kiblnd_data.kib_scheds[cpt];
3093
3094                 if (!newdev && sched->ibs_nthreads > 0)
3095                         continue;
3096
3097                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3098                 if (rc != 0) {
3099                         CERROR("Failed to start scheduler threads for %s\n",
3100                                dev->ibd_ifname);
3101                         return rc;
3102                 }
3103         }
3104         return 0;
3105 }
3106
3107 kib_dev_t *
3108 kiblnd_dev_search(char *ifname)
3109 {
3110         kib_dev_t       *alias = NULL;
3111         kib_dev_t       *dev;
3112         char            *colon;
3113         char            *colon2;
3114
3115         colon = strchr(ifname, ':');
3116         cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3117                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3118                         return dev;
3119
3120                 if (alias != NULL)
3121                         continue;
3122
3123                 colon2 = strchr(dev->ibd_ifname, ':');
3124                 if (colon != NULL)
3125                         *colon = 0;
3126                 if (colon2 != NULL)
3127                         *colon2 = 0;
3128
3129                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3130                         alias = dev;
3131
3132                 if (colon != NULL)
3133                         *colon = ':';
3134                 if (colon2 != NULL)
3135                         *colon2 = ':';
3136         }
3137         return alias;
3138 }
3139
3140 int
3141 kiblnd_startup (lnet_ni_t *ni)
3142 {
3143         char                     *ifname;
3144         kib_dev_t                *ibdev = NULL;
3145         kib_net_t                *net;
3146         struct timeval            tv;
3147         unsigned long             flags;
3148         int                       rc;
3149         int                       newdev;
3150
3151         LASSERT (ni->ni_lnd == &the_o2iblnd);
3152
3153         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3154                 rc = kiblnd_base_startup();
3155                 if (rc != 0)
3156                         return rc;
3157         }
3158
3159         LIBCFS_ALLOC(net, sizeof(*net));
3160         ni->ni_data = net;
3161         if (net == NULL)
3162                 goto failed;
3163
3164         memset(net, 0, sizeof(*net));
3165
3166         cfs_gettimeofday(&tv);
3167         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3168
3169         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3170         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3171         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3172         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3173
3174         if (ni->ni_interfaces[0] != NULL) {
3175                 /* Use the IPoIB interface specified in 'networks=' */
3176
3177                 CLASSERT (LNET_MAX_INTERFACES > 1);
3178                 if (ni->ni_interfaces[1] != NULL) {
3179                         CERROR("Multiple interfaces not supported\n");
3180                         goto failed;
3181                 }
3182
3183                 ifname = ni->ni_interfaces[0];
3184         } else {
3185                 ifname = *kiblnd_tunables.kib_default_ipif;
3186         }
3187
3188         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3189                 CERROR("IPoIB interface name too long: %s\n", ifname);
3190                 goto failed;
3191         }
3192
3193         ibdev = kiblnd_dev_search(ifname);
3194
3195         newdev = ibdev == NULL;
3196         /* hmm...create kib_dev even for alias */
3197         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3198                 ibdev = kiblnd_create_dev(ifname);
3199
3200         if (ibdev == NULL)
3201                 goto failed;
3202
3203         net->ibn_dev = ibdev;
3204         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3205
3206         rc = kiblnd_dev_start_threads(ibdev, newdev,
3207                                       ni->ni_cpts, ni->ni_ncpts);
3208         if (rc != 0)
3209                 goto failed;
3210
3211         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3212         if (rc != 0) {
3213                 CERROR("Failed to initialize NI pools: %d\n", rc);
3214                 goto failed;
3215         }
3216
3217         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3218         ibdev->ibd_nnets++;
3219         cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3220         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3221
3222         net->ibn_init = IBLND_INIT_ALL;
3223
3224         return 0;
3225
3226 failed:
3227         if (net->ibn_dev == NULL && ibdev != NULL)
3228                 kiblnd_destroy_dev(ibdev);
3229
3230         kiblnd_shutdown(ni);
3231
3232         CDEBUG(D_NET, "kiblnd_startup failed\n");
3233         return -ENETDOWN;
3234 }
3235
3236 void __exit
3237 kiblnd_module_fini (void)
3238 {
3239         lnet_unregister_lnd(&the_o2iblnd);
3240         kiblnd_tunables_fini();
3241 }
3242
3243 int __init
3244 kiblnd_module_init (void)
3245 {
3246         int    rc;
3247
3248         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3249         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3250                   <= IBLND_MSG_SIZE);
3251         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3252                   <= IBLND_MSG_SIZE);
3253
3254         rc = kiblnd_tunables_init();
3255         if (rc != 0)
3256                 return rc;
3257
3258         lnet_register_lnd(&the_o2iblnd);
3259
3260         return 0;
3261 }
3262
3263 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3264 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3265 MODULE_LICENSE("GPL");
3266
3267 module_init(kiblnd_module_init);
3268 module_exit(kiblnd_module_fini);