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, Whamcloud, Inc.
  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         cfs_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         cfs_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         cfs_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                         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 461                                                    flags);
 462                         return 0;
 463                 }
 464         }
 465
 466         cfs_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         cfs_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         cfs_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         cfs_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                                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 576                                                            flags);
 577                                 return conn;
 578                         }
 579                 }
 580         }
 581
 582         cfs_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         cfs_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         cfs_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         cfs_rwlock_t           *glock = &kiblnd_data.kib_global_lock;
 729         kib_net_t              *net = peer->ibp_ni->ni_data;
 730         kib_dev_t              *dev = net->ibn_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         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 744         sched = kiblnd_data.kib_scheds[cpt];
 745
 746         LASSERT(sched->ibs_nthreads > 0);
 747
 748         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 749                          sizeof(*init_qp_attr));
 750         if (init_qp_attr == NULL) {
 751                 CERROR("Can't allocate qp_attr for %s\n",
 752                        libcfs_nid2str(peer->ibp_nid));
 753                 goto failed_0;
 754         }
 755
 756         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 757         if (conn == NULL) {
 758                 CERROR("Can't allocate connection for %s\n",
 759                        libcfs_nid2str(peer->ibp_nid));
 760                 goto failed_1;
 761         }
 762
 763         conn->ibc_state = IBLND_CONN_INIT;
 764         conn->ibc_version = version;
 765         conn->ibc_peer = peer;                  /* I take the caller's ref */
 766         cmid->context = conn;                   /* for future CM callbacks */
 767         conn->ibc_cmid = cmid;
 768
 769         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 770         CFS_INIT_LIST_HEAD(&conn->ibc_tx_noops);
 771         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 772         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 773         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 774         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 775         cfs_spin_lock_init(&conn->ibc_lock);
 776
 777         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 778                          sizeof(*conn->ibc_connvars));
 779         if (conn->ibc_connvars == NULL) {
 780                 CERROR("Can't allocate in-progress connection state\n");
 781                 goto failed_2;
 782         }
 783
 784         cfs_write_lock_irqsave(glock, flags);
 785         if (dev->ibd_failover) {
 786                 cfs_write_unlock_irqrestore(glock, flags);
 787                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 788                 goto failed_2;
 789         }
 790
 791         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 792                 /* wakeup failover thread and teardown connection */
 793                 if (kiblnd_dev_can_failover(dev)) {
 794                         cfs_list_add_tail(&dev->ibd_fail_list,
 795                                       &kiblnd_data.kib_failed_devs);
 796                         cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
 797                 }
 798
 799                 cfs_write_unlock_irqrestore(glock, flags);
 800                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 801                        cmid->device->name, dev->ibd_ifname);
 802                 goto failed_2;
 803         }
 804
 805         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 806         conn->ibc_hdev = dev->ibd_hdev;
 807
 808         kiblnd_setup_mtu_locked(cmid);
 809
 810         cfs_write_unlock_irqrestore(glock, flags);
 811
 812         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 813                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 814         if (conn->ibc_rxs == NULL) {
 815                 CERROR("Cannot allocate RX buffers\n");
 816                 goto failed_2;
 817         }
 818
 819         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 820                                 IBLND_RX_MSG_PAGES(version));
 821         if (rc != 0)
 822                 goto failed_2;
 823
 824         kiblnd_map_rx_descs(conn);
 825
 826 #ifdef HAVE_OFED_IB_COMP_VECTOR
 827         cq = ib_create_cq(cmid->device,
 828                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 829                           IBLND_CQ_ENTRIES(version),
 830                           kiblnd_get_completion_vector(conn, cpt));
 831 #else
 832         cq = ib_create_cq(cmid->device,
 833                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 834                           IBLND_CQ_ENTRIES(version));
 835 #endif
 836         if (IS_ERR(cq)) {
 837                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 838                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 839                 goto failed_2;
 840         }
 841
 842         conn->ibc_cq = cq;
 843
 844         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 845         if (rc != 0) {
 846                 CERROR("Can't request completion notificiation: %d\n", rc);
 847                 goto failed_2;
 848         }
 849
 850         init_qp_attr->event_handler = kiblnd_qp_event;
 851         init_qp_attr->qp_context = conn;
 852         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 853         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 854         init_qp_attr->cap.max_send_sge = 1;
 855         init_qp_attr->cap.max_recv_sge = 1;
 856         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 857         init_qp_attr->qp_type = IB_QPT_RC;
 858         init_qp_attr->send_cq = cq;
 859         init_qp_attr->recv_cq = cq;
 860
 861         conn->ibc_sched = sched;
 862
 863         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 864         if (rc != 0) {
 865                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 866                        rc, init_qp_attr->cap.max_send_wr,
 867                        init_qp_attr->cap.max_recv_wr);
 868                 goto failed_2;
 869         }
 870
 871         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 872
 873         /* 1 ref for caller and each rxmsg */
 874         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 875         conn->ibc_nrx = IBLND_RX_MSGS(version);
 876
 877         /* post receives */
 878         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 879                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 880                                     IBLND_POSTRX_NO_CREDIT);
 881                 if (rc != 0) {
 882                         CERROR("Can't post rxmsg: %d\n", rc);
 883
 884                         /* Make posted receives complete */
 885                         kiblnd_abort_receives(conn);
 886
 887                         /* correct # of posted buffers
 888                          * NB locking needed now I'm racing with completion */
 889                         cfs_spin_lock_irqsave(&sched->ibs_lock, flags);
 890                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 891                         cfs_spin_unlock_irqrestore(&sched->ibs_lock, flags);
 892
 893                         /* cmid will be destroyed by CM(ofed) after cm_callback
 894                          * returned, so we can't refer it anymore
 895                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 896                         rdma_destroy_qp(conn->ibc_cmid);
 897                         conn->ibc_cmid = NULL;
 898
 899                         /* Drop my own and unused rxbuffer refcounts */
 900                         while (i++ <= IBLND_RX_MSGS(version))
 901                                 kiblnd_conn_decref(conn);
 902
 903                         return NULL;
 904                 }
 905         }
 906
 907         /* Init successful! */
 908         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 909                  state == IBLND_CONN_PASSIVE_WAIT);
 910         conn->ibc_state = state;
 911
 912         /* 1 more conn */
 913         cfs_atomic_inc(&net->ibn_nconns);
 914         return conn;
 915
 916  failed_2:
 917         kiblnd_destroy_conn(conn);
 918  failed_1:
 919         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 920  failed_0:
 921         return NULL;
 922 }
 923
 924 void
 925 kiblnd_destroy_conn (kib_conn_t *conn)
 926 {
 927         struct rdma_cm_id *cmid = conn->ibc_cmid;
 928         kib_peer_t        *peer = conn->ibc_peer;
 929         int                rc;
 930
 931         LASSERT (!cfs_in_interrupt());
 932         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 933         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 934         LASSERT (cfs_list_empty(&conn->ibc_tx_noops));
 935         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 936         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 937         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 938         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 939         LASSERT (conn->ibc_noops_posted == 0);
 940         LASSERT (conn->ibc_nsends_posted == 0);
 941
 942         switch (conn->ibc_state) {
 943         default:
 944                 /* conn must be completely disengaged from the network */
 945                 LBUG();
 946
 947         case IBLND_CONN_DISCONNECTED:
 948                 /* connvars should have been freed already */
 949                 LASSERT (conn->ibc_connvars == NULL);
 950                 break;
 951
 952         case IBLND_CONN_INIT:
 953                 break;
 954         }
 955
 956         /* conn->ibc_cmid might be destroyed by CM already */
 957         if (cmid != NULL && cmid->qp != NULL)
 958                 rdma_destroy_qp(cmid);
 959
 960         if (conn->ibc_cq != NULL) {
 961                 rc = ib_destroy_cq(conn->ibc_cq);
 962                 if (rc != 0)
 963                         CWARN("Error destroying CQ: %d\n", rc);
 964         }
 965
 966         if (conn->ibc_rx_pages != NULL)
 967                 kiblnd_unmap_rx_descs(conn);
 968
 969         if (conn->ibc_rxs != NULL) {
 970                 LIBCFS_FREE(conn->ibc_rxs,
 971                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 972         }
 973
 974         if (conn->ibc_connvars != NULL)
 975                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 976
 977         if (conn->ibc_hdev != NULL)
 978                 kiblnd_hdev_decref(conn->ibc_hdev);
 979
 980         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 981         if (conn->ibc_state != IBLND_CONN_INIT) {
 982                 kib_net_t *net = peer->ibp_ni->ni_data;
 983
 984                 kiblnd_peer_decref(peer);
 985                 rdma_destroy_id(cmid);
 986                 cfs_atomic_dec(&net->ibn_nconns);
 987         }
 988
 989         LIBCFS_FREE(conn, sizeof(*conn));
 990 }
 991
 992 int
 993 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 994 {
 995         kib_conn_t             *conn;
 996         cfs_list_t             *ctmp;
 997         cfs_list_t             *cnxt;
 998         int                     count = 0;
 999
1000         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1001                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1002
1003                 CDEBUG(D_NET, "Closing conn -> %s, "
1004                               "version: %x, reason: %d\n",
1005                        libcfs_nid2str(peer->ibp_nid),
1006                        conn->ibc_version, why);
1007
1008                 kiblnd_close_conn_locked(conn, why);
1009                 count++;
1010         }
1011
1012         return count;
1013 }
1014
1015 int
1016 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1017                                  int version, __u64 incarnation)
1018 {
1019         kib_conn_t             *conn;
1020         cfs_list_t             *ctmp;
1021         cfs_list_t             *cnxt;
1022         int                     count = 0;
1023
1024         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1025                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
1026
1027                 if (conn->ibc_version     == version &&
1028                     conn->ibc_incarnation == incarnation)
1029                         continue;
1030
1031                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1032                               "incarnation:"LPX64"(%x, "LPX64")\n",
1033                        libcfs_nid2str(peer->ibp_nid),
1034                        conn->ibc_version, conn->ibc_incarnation,
1035                        version, incarnation);
1036
1037                 kiblnd_close_conn_locked(conn, -ESTALE);
1038                 count++;
1039         }
1040
1041         return count;
1042 }
1043
1044 int
1045 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1046 {
1047         kib_peer_t             *peer;
1048         cfs_list_t             *ptmp;
1049         cfs_list_t             *pnxt;
1050         int                     lo;
1051         int                     hi;
1052         int                     i;
1053         unsigned long           flags;
1054         int                     count = 0;
1055
1056         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1057
1058         if (nid != LNET_NID_ANY)
1059                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1060         else {
1061                 lo = 0;
1062                 hi = kiblnd_data.kib_peer_hash_size - 1;
1063         }
1064
1065         for (i = lo; i <= hi; i++) {
1066                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1067
1068                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1069                         LASSERT (peer->ibp_connecting > 0 ||
1070                                  peer->ibp_accepting > 0 ||
1071                                  !cfs_list_empty(&peer->ibp_conns));
1072
1073                         if (peer->ibp_ni != ni)
1074                                 continue;
1075
1076                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1077                                 continue;
1078
1079                         count += kiblnd_close_peer_conns_locked(peer, 0);
1080                 }
1081         }
1082
1083         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1084
1085         /* wildcards always succeed */
1086         if (nid == LNET_NID_ANY)
1087                 return 0;
1088
1089         return (count == 0) ? -ENOENT : 0;
1090 }
1091
1092 int
1093 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1094 {
1095         struct libcfs_ioctl_data *data = arg;
1096         int                       rc = -EINVAL;
1097
1098         switch(cmd) {
1099         case IOC_LIBCFS_GET_PEER: {
1100                 lnet_nid_t   nid = 0;
1101                 int          count = 0;
1102
1103                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1104                                           &nid, &count);
1105                 data->ioc_nid    = nid;
1106                 data->ioc_count  = count;
1107                 break;
1108         }
1109
1110         case IOC_LIBCFS_DEL_PEER: {
1111                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1112                 break;
1113         }
1114         case IOC_LIBCFS_GET_CONN: {
1115                 kib_conn_t *conn;
1116
1117                 rc = 0;
1118                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1119                 if (conn == NULL) {
1120                         rc = -ENOENT;
1121                         break;
1122                 }
1123
1124                 LASSERT (conn->ibc_cmid != NULL);
1125                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1126                 if (conn->ibc_cmid->route.path_rec == NULL)
1127                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1128                 else
1129                         data->ioc_u32[0] =
1130                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1131                 kiblnd_conn_decref(conn);
1132                 break;
1133         }
1134         case IOC_LIBCFS_CLOSE_CONNECTION: {
1135                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1136                 break;
1137         }
1138
1139         default:
1140                 break;
1141         }
1142
1143         return rc;
1144 }
1145
1146 void
1147 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1148 {
1149         cfs_time_t         last_alive = 0;
1150         cfs_time_t         now = cfs_time_current();
1151         cfs_rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1152         kib_peer_t        *peer;
1153         unsigned long      flags;
1154
1155         cfs_read_lock_irqsave(glock, flags);
1156
1157         peer = kiblnd_find_peer_locked(nid);
1158         if (peer != NULL) {
1159                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1160                          peer->ibp_accepting > 0 ||
1161                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1162                 last_alive = peer->ibp_last_alive;
1163         }
1164
1165         cfs_read_unlock_irqrestore(glock, flags);
1166
1167         if (last_alive != 0)
1168                 *when = last_alive;
1169
1170         /* peer is not persistent in hash, trigger peer creation
1171          * and connection establishment with a NULL tx */
1172         if (peer == NULL)
1173                 kiblnd_launch_tx(ni, NULL, nid);
1174
1175         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1176                libcfs_nid2str(nid), peer,
1177                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1178         return;
1179 }
1180
1181 void
1182 kiblnd_free_pages(kib_pages_t *p)
1183 {
1184         int     npages = p->ibp_npages;
1185         int     i;
1186
1187         for (i = 0; i < npages; i++) {
1188                 if (p->ibp_pages[i] != NULL)
1189                         cfs_free_page(p->ibp_pages[i]);
1190         }
1191
1192         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1193 }
1194
1195 int
1196 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1197 {
1198         kib_pages_t     *p;
1199         int             i;
1200
1201         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1202                          offsetof(kib_pages_t, ibp_pages[npages]));
1203         if (p == NULL) {
1204                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1205                 return -ENOMEM;
1206         }
1207
1208         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1209         p->ibp_npages = npages;
1210
1211         for (i = 0; i < npages; i++) {
1212                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1213                                                      CFS_ALLOC_IO);
1214                 if (p->ibp_pages[i] == NULL) {
1215                         CERROR("Can't allocate page %d of %d\n", i, npages);
1216                         kiblnd_free_pages(p);
1217                         return -ENOMEM;
1218                 }
1219         }
1220
1221         *pp = p;
1222         return 0;
1223 }
1224
1225 void
1226 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1227 {
1228         kib_rx_t *rx;
1229         int       i;
1230
1231         LASSERT (conn->ibc_rxs != NULL);
1232         LASSERT (conn->ibc_hdev != NULL);
1233
1234         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1235                 rx = &conn->ibc_rxs[i];
1236
1237                 LASSERT (rx->rx_nob >= 0); /* not posted */
1238
1239                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1240                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1241                                                           rx->rx_msgaddr),
1242                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1243         }
1244
1245         kiblnd_free_pages(conn->ibc_rx_pages);
1246
1247         conn->ibc_rx_pages = NULL;
1248 }
1249
1250 void
1251 kiblnd_map_rx_descs(kib_conn_t *conn)
1252 {
1253         kib_rx_t       *rx;
1254         struct page    *pg;
1255         int             pg_off;
1256         int             ipg;
1257         int             i;
1258
1259         for (pg_off = ipg = i = 0;
1260              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1261                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1262                 rx = &conn->ibc_rxs[i];
1263
1264                 rx->rx_conn = conn;
1265                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1266
1267                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1268                                                        rx->rx_msg, IBLND_MSG_SIZE,
1269                                                        DMA_FROM_DEVICE);
1270                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1271                                                    rx->rx_msgaddr));
1272                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1273
1274                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1275                        i, rx->rx_msg, rx->rx_msgaddr,
1276                        lnet_page2phys(pg) + pg_off);
1277
1278                 pg_off += IBLND_MSG_SIZE;
1279                 LASSERT (pg_off <= PAGE_SIZE);
1280
1281                 if (pg_off == PAGE_SIZE) {
1282                         pg_off = 0;
1283                         ipg++;
1284                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1285                 }
1286         }
1287 }
1288
1289 static void
1290 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1291 {
1292         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1293         kib_tx_t       *tx;
1294         int             i;
1295
1296         LASSERT (tpo->tpo_pool.po_allocated == 0);
1297
1298         if (hdev == NULL)
1299                 return;
1300
1301         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1302                 tx = &tpo->tpo_tx_descs[i];
1303                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1304                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1305                                                           tx->tx_msgaddr),
1306                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1307         }
1308
1309         kiblnd_hdev_decref(hdev);
1310         tpo->tpo_hdev = NULL;
1311 }
1312
1313 static kib_hca_dev_t *
1314 kiblnd_current_hdev(kib_dev_t *dev)
1315 {
1316         kib_hca_dev_t *hdev;
1317         unsigned long  flags;
1318         int            i = 0;
1319
1320         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1321         while (dev->ibd_failover) {
1322                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1323                 if (i++ % 50 == 0)
1324                         CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
1325                 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1326
1327                 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1328         }
1329
1330         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331         hdev = dev->ibd_hdev;
1332
1333         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1334
1335         return hdev;
1336 }
1337
1338 static void
1339 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1340 {
1341         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1342         kib_pool_t     *pool  = &tpo->tpo_pool;
1343         kib_net_t      *net   = pool->po_owner->ps_net;
1344         kib_dev_t      *dev   = net->ibn_dev;
1345         struct page    *page;
1346         kib_tx_t       *tx;
1347         int             page_offset;
1348         int             ipage;
1349         int             i;
1350
1351         LASSERT (net != NULL);
1352
1353         /* pre-mapped messages are not bigger than 1 page */
1354         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1355
1356         /* No fancy arithmetic when we do the buffer calculations */
1357         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1358
1359         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1360
1361         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1362                 page = txpgs->ibp_pages[ipage];
1363                 tx = &tpo->tpo_tx_descs[i];
1364
1365                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1366                                            page_offset);
1367
1368                 tx->tx_msgaddr = kiblnd_dma_map_single(
1369                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1370                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1371                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1372                                                    tx->tx_msgaddr));
1373                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1374
1375                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1376
1377                 page_offset += IBLND_MSG_SIZE;
1378                 LASSERT (page_offset <= PAGE_SIZE);
1379
1380                 if (page_offset == PAGE_SIZE) {
1381                         page_offset = 0;
1382                         ipage++;
1383                         LASSERT (ipage <= txpgs->ibp_npages);
1384                 }
1385         }
1386 }
1387
1388 struct ib_mr *
1389 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1390 {
1391         __u64   index;
1392
1393         LASSERT (hdev->ibh_mrs[0] != NULL);
1394
1395         if (hdev->ibh_nmrs == 1)
1396                 return hdev->ibh_mrs[0];
1397
1398         index = addr >> hdev->ibh_mr_shift;
1399
1400         if (index <  hdev->ibh_nmrs &&
1401             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1402                 return hdev->ibh_mrs[index];
1403
1404         return NULL;
1405 }
1406
1407 struct ib_mr *
1408 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1409 {
1410         struct ib_mr *prev_mr;
1411         struct ib_mr *mr;
1412         int           i;
1413
1414         LASSERT (hdev->ibh_mrs[0] != NULL);
1415
1416         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1417             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1418                 return NULL;
1419
1420         if (hdev->ibh_nmrs == 1)
1421                 return hdev->ibh_mrs[0];
1422
1423         for (i = 0, mr = prev_mr = NULL;
1424              i < rd->rd_nfrags; i++) {
1425                 mr = kiblnd_find_dma_mr(hdev,
1426                                         rd->rd_frags[i].rf_addr,
1427                                         rd->rd_frags[i].rf_nob);
1428                 if (prev_mr == NULL)
1429                         prev_mr = mr;
1430
1431                 if (mr == NULL || prev_mr != mr) {
1432                         /* Can't covered by one single MR */
1433                         mr = NULL;
1434                         break;
1435                 }
1436         }
1437
1438         return mr;
1439 }
1440
1441 void
1442 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1443 {
1444         LASSERT (pool->fpo_map_count == 0);
1445
1446         if (pool->fpo_fmr_pool != NULL)
1447                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1448
1449         if (pool->fpo_hdev != NULL)
1450                 kiblnd_hdev_decref(pool->fpo_hdev);
1451
1452         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1453 }
1454
1455 void
1456 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1457 {
1458         kib_fmr_pool_t *pool;
1459
1460         while (!cfs_list_empty(head)) {
1461                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1462                 cfs_list_del(&pool->fpo_list);
1463                 kiblnd_destroy_fmr_pool(pool);
1464         }
1465 }
1466
1467 static int kiblnd_fmr_pool_size(int ncpts)
1468 {
1469         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1470
1471         return max(IBLND_FMR_POOL, size);
1472 }
1473
1474 static int kiblnd_fmr_flush_trigger(int ncpts)
1475 {
1476         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1477
1478         return max(IBLND_FMR_POOL_FLUSH, size);
1479 }
1480
1481 int
1482 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1483 {
1484         /* FMR pool for RDMA */
1485         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1486         kib_fmr_pool_t          *fpo;
1487         struct ib_fmr_pool_param param = {
1488                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1489                 .page_shift        = PAGE_SHIFT,
1490                 .access            = (IB_ACCESS_LOCAL_WRITE |
1491                                       IB_ACCESS_REMOTE_WRITE),
1492                 .pool_size         = fps->fps_pool_size,
1493                 .dirty_watermark   = fps->fps_flush_trigger,
1494                 .flush_function    = NULL,
1495                 .flush_arg         = NULL,
1496                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1497         int rc;
1498
1499         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1500         if (fpo == NULL)
1501                 return -ENOMEM;
1502
1503         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1504
1505         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1506         if (IS_ERR(fpo->fpo_fmr_pool)) {
1507                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1508                 CERROR("Failed to create FMR pool: %d\n", rc);
1509
1510                 kiblnd_hdev_decref(fpo->fpo_hdev);
1511                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1512                 return rc;
1513         }
1514
1515         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1516         fpo->fpo_owner    = fps;
1517         *pp_fpo = fpo;
1518
1519         return 0;
1520 }
1521
1522 static void
1523 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1524 {
1525         if (fps->fps_net == NULL) /* intialized? */
1526                 return;
1527
1528         cfs_spin_lock(&fps->fps_lock);
1529
1530         while (!cfs_list_empty(&fps->fps_pool_list)) {
1531                 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1532                                                  kib_fmr_pool_t, fpo_list);
1533                 fpo->fpo_failed = 1;
1534                 cfs_list_del(&fpo->fpo_list);
1535                 if (fpo->fpo_map_count == 0)
1536                         cfs_list_add(&fpo->fpo_list, zombies);
1537                 else
1538                         cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1539         }
1540
1541         cfs_spin_unlock(&fps->fps_lock);
1542 }
1543
1544 static void
1545 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1546 {
1547         if (fps->fps_net != NULL) { /* initialized? */
1548                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1549                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1550         }
1551 }
1552
1553 static int
1554 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1555                         int pool_size, int flush_trigger)
1556 {
1557         kib_fmr_pool_t *fpo;
1558         int             rc;
1559
1560         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1561
1562         fps->fps_net = net;
1563         fps->fps_cpt = cpt;
1564         fps->fps_pool_size = pool_size;
1565         fps->fps_flush_trigger = flush_trigger;
1566         cfs_spin_lock_init(&fps->fps_lock);
1567         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1568         CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1569
1570         rc = kiblnd_create_fmr_pool(fps, &fpo);
1571         if (rc == 0)
1572                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1573
1574         return rc;
1575 }
1576
1577 static int
1578 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1579 {
1580         if (fpo->fpo_map_count != 0) /* still in use */
1581                 return 0;
1582         if (fpo->fpo_failed)
1583                 return 1;
1584         return cfs_time_aftereq(now, fpo->fpo_deadline);
1585 }
1586
1587 void
1588 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1589 {
1590         CFS_LIST_HEAD     (zombies);
1591         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1592         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1593         cfs_time_t         now = cfs_time_current();
1594         kib_fmr_pool_t    *tmp;
1595         int                rc;
1596
1597         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1598         LASSERT (rc == 0);
1599
1600         if (status != 0) {
1601                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1602                 LASSERT (rc == 0);
1603         }
1604
1605         fmr->fmr_pool = NULL;
1606         fmr->fmr_pfmr = NULL;
1607
1608         cfs_spin_lock(&fps->fps_lock);
1609         fpo->fpo_map_count --;  /* decref the pool */
1610
1611         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1612                 /* the first pool is persistent */
1613                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1614                         continue;
1615
1616                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1617                         cfs_list_move(&fpo->fpo_list, &zombies);
1618                         fps->fps_version ++;
1619                 }
1620         }
1621         cfs_spin_unlock(&fps->fps_lock);
1622
1623         if (!cfs_list_empty(&zombies))
1624                 kiblnd_destroy_fmr_pool_list(&zombies);
1625 }
1626
1627 int
1628 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1629                     __u64 iov, kib_fmr_t *fmr)
1630 {
1631         struct ib_pool_fmr *pfmr;
1632         kib_fmr_pool_t     *fpo;
1633         __u64               version;
1634         int                 rc;
1635
1636  again:
1637         cfs_spin_lock(&fps->fps_lock);
1638         version = fps->fps_version;
1639         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1640                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1641                 fpo->fpo_map_count ++;
1642                 cfs_spin_unlock(&fps->fps_lock);
1643
1644                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1645                                             pages, npages, iov);
1646                 if (likely(!IS_ERR(pfmr))) {
1647                         fmr->fmr_pool = fpo;
1648                         fmr->fmr_pfmr = pfmr;
1649                         return 0;
1650                 }
1651
1652                 cfs_spin_lock(&fps->fps_lock);
1653                 fpo->fpo_map_count --;
1654                 if (PTR_ERR(pfmr) != -EAGAIN) {
1655                         cfs_spin_unlock(&fps->fps_lock);
1656                         return PTR_ERR(pfmr);
1657                 }
1658
1659                 /* EAGAIN and ... */
1660                 if (version != fps->fps_version) {
1661                         cfs_spin_unlock(&fps->fps_lock);
1662                         goto again;
1663                 }
1664         }
1665
1666         if (fps->fps_increasing) {
1667                 cfs_spin_unlock(&fps->fps_lock);
1668                 CDEBUG(D_NET, "Another thread is allocating new "
1669                               "FMR pool, waiting for her to complete\n");
1670                 cfs_schedule();
1671                 goto again;
1672
1673         }
1674
1675         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1676                 /* someone failed recently */
1677                 cfs_spin_unlock(&fps->fps_lock);
1678                 return -EAGAIN;
1679         }
1680
1681         fps->fps_increasing = 1;
1682         cfs_spin_unlock(&fps->fps_lock);
1683
1684         CDEBUG(D_NET, "Allocate new FMR pool\n");
1685         rc = kiblnd_create_fmr_pool(fps, &fpo);
1686         cfs_spin_lock(&fps->fps_lock);
1687         fps->fps_increasing = 0;
1688         if (rc == 0) {
1689                 fps->fps_version ++;
1690                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1691         } else {
1692                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1693         }
1694         cfs_spin_unlock(&fps->fps_lock);
1695
1696         goto again;
1697 }
1698
1699 static void
1700 kiblnd_fini_pool(kib_pool_t *pool)
1701 {
1702         LASSERT (cfs_list_empty(&pool->po_free_list));
1703         LASSERT (pool->po_allocated == 0);
1704
1705         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1706 }
1707
1708 static void
1709 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1710 {
1711         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1712
1713         memset(pool, 0, sizeof(kib_pool_t));
1714         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1715         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1716         pool->po_owner    = ps;
1717         pool->po_size     = size;
1718 }
1719
1720 void
1721 kiblnd_destroy_pool_list(cfs_list_t *head)
1722 {
1723         kib_pool_t *pool;
1724
1725         while (!cfs_list_empty(head)) {
1726                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1727                 cfs_list_del(&pool->po_list);
1728
1729                 LASSERT (pool->po_owner != NULL);
1730                 pool->po_owner->ps_pool_destroy(pool);
1731         }
1732 }
1733
1734 static void
1735 kiblnd_fail_poolset(kib_poolset_t *ps, cfs_list_t *zombies)
1736 {
1737         if (ps->ps_net == NULL) /* intialized? */
1738                 return;
1739
1740         cfs_spin_lock(&ps->ps_lock);
1741         while (!cfs_list_empty(&ps->ps_pool_list)) {
1742                 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1743                                             kib_pool_t, po_list);
1744                 po->po_failed = 1;
1745                 cfs_list_del(&po->po_list);
1746                 if (po->po_allocated == 0)
1747                         cfs_list_add(&po->po_list, zombies);
1748                 else
1749                         cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1750         }
1751         cfs_spin_unlock(&ps->ps_lock);
1752 }
1753
1754 static void
1755 kiblnd_fini_poolset(kib_poolset_t *ps)
1756 {
1757         if (ps->ps_net != NULL) { /* initialized? */
1758                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1759                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1760         }
1761 }
1762
1763 static int
1764 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1765                     kib_net_t *net, char *name, int size,
1766                     kib_ps_pool_create_t po_create,
1767                     kib_ps_pool_destroy_t po_destroy,
1768                     kib_ps_node_init_t nd_init,
1769                     kib_ps_node_fini_t nd_fini)
1770 {
1771         kib_pool_t      *pool;
1772         int             rc;
1773
1774         memset(ps, 0, sizeof(kib_poolset_t));
1775
1776         ps->ps_cpt          = cpt;
1777         ps->ps_net          = net;
1778         ps->ps_pool_create  = po_create;
1779         ps->ps_pool_destroy = po_destroy;
1780         ps->ps_node_init    = nd_init;
1781         ps->ps_node_fini    = nd_fini;
1782         ps->ps_pool_size    = size;
1783         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1784         cfs_spin_lock_init(&ps->ps_lock);
1785         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1786         CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1787
1788         rc = ps->ps_pool_create(ps, size, &pool);
1789         if (rc == 0)
1790                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1791         else
1792                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1793
1794         return rc;
1795 }
1796
1797 static int
1798 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1799 {
1800         if (pool->po_allocated != 0) /* still in use */
1801                 return 0;
1802         if (pool->po_failed)
1803                 return 1;
1804         return cfs_time_aftereq(now, pool->po_deadline);
1805 }
1806
1807 void
1808 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1809 {
1810         CFS_LIST_HEAD  (zombies);
1811         kib_poolset_t  *ps = pool->po_owner;
1812         kib_pool_t     *tmp;
1813         cfs_time_t      now = cfs_time_current();
1814
1815         cfs_spin_lock(&ps->ps_lock);
1816
1817         if (ps->ps_node_fini != NULL)
1818                 ps->ps_node_fini(pool, node);
1819
1820         LASSERT (pool->po_allocated > 0);
1821         cfs_list_add(node, &pool->po_free_list);
1822         pool->po_allocated --;
1823
1824         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1825                 /* the first pool is persistent */
1826                 if (ps->ps_pool_list.next == &pool->po_list)
1827                         continue;
1828
1829                 if (kiblnd_pool_is_idle(pool, now))
1830                         cfs_list_move(&pool->po_list, &zombies);
1831         }
1832         cfs_spin_unlock(&ps->ps_lock);
1833
1834         if (!cfs_list_empty(&zombies))
1835                 kiblnd_destroy_pool_list(&zombies);
1836 }
1837
1838 cfs_list_t *
1839 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1840 {
1841         cfs_list_t            *node;
1842         kib_pool_t            *pool;
1843         int                    rc;
1844
1845  again:
1846         cfs_spin_lock(&ps->ps_lock);
1847         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1848                 if (cfs_list_empty(&pool->po_free_list))
1849                         continue;
1850
1851                 pool->po_allocated ++;
1852                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1853                 node = pool->po_free_list.next;
1854                 cfs_list_del(node);
1855
1856                 if (ps->ps_node_init != NULL) {
1857                         /* still hold the lock */
1858                         ps->ps_node_init(pool, node);
1859                 }
1860                 cfs_spin_unlock(&ps->ps_lock);
1861                 return node;
1862         }
1863
1864         /* no available tx pool and ... */
1865         if (ps->ps_increasing) {
1866                 /* another thread is allocating a new pool */
1867                 cfs_spin_unlock(&ps->ps_lock);
1868                 CDEBUG(D_NET, "Another thread is allocating new "
1869                        "%s pool, waiting for her to complete\n",
1870                        ps->ps_name);
1871                 cfs_schedule();
1872                 goto again;
1873         }
1874
1875         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1876                 /* someone failed recently */
1877                 cfs_spin_unlock(&ps->ps_lock);
1878                 return NULL;
1879         }
1880
1881         ps->ps_increasing = 1;
1882         cfs_spin_unlock(&ps->ps_lock);
1883
1884         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1885
1886         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1887
1888         cfs_spin_lock(&ps->ps_lock);
1889         ps->ps_increasing = 0;
1890         if (rc == 0) {
1891                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1892         } else {
1893                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1894                 CERROR("Can't allocate new %s pool because out of memory\n",
1895                        ps->ps_name);
1896         }
1897         cfs_spin_unlock(&ps->ps_lock);
1898
1899         goto again;
1900 }
1901
1902 void
1903 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1904 {
1905         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1906         struct ib_mr        *mr  = pmr->pmr_mr;
1907
1908         pmr->pmr_mr = NULL;
1909         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1910         if (mr != NULL)
1911                 ib_dereg_mr(mr);
1912 }
1913
1914 int
1915 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1916                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1917 {
1918         kib_phys_mr_t *pmr;
1919         cfs_list_t    *node;
1920         int            rc;
1921         int            i;
1922
1923         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1924         if (node == NULL) {
1925                 CERROR("Failed to allocate PMR descriptor\n");
1926                 return -ENOMEM;
1927         }
1928
1929         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1930         if (pmr->pmr_pool->ppo_hdev != hdev) {
1931                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1932                 return -EAGAIN;
1933         }
1934
1935         for (i = 0; i < rd->rd_nfrags; i ++) {
1936                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1937                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1938         }
1939
1940         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1941                                      pmr->pmr_ipb, rd->rd_nfrags,
1942                                      IB_ACCESS_LOCAL_WRITE |
1943                                      IB_ACCESS_REMOTE_WRITE,
1944                                      iova);
1945         if (!IS_ERR(pmr->pmr_mr)) {
1946                 pmr->pmr_iova = *iova;
1947                 *pp_pmr = pmr;
1948                 return 0;
1949         }
1950
1951         rc = PTR_ERR(pmr->pmr_mr);
1952         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1953
1954         pmr->pmr_mr = NULL;
1955         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1956
1957         return rc;
1958 }
1959
1960 static void
1961 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1962 {
1963         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1964         kib_phys_mr_t  *pmr;
1965
1966         LASSERT (pool->po_allocated == 0);
1967
1968         while (!cfs_list_empty(&pool->po_free_list)) {
1969                 pmr = cfs_list_entry(pool->po_free_list.next,
1970                                      kib_phys_mr_t, pmr_list);
1971
1972                 LASSERT (pmr->pmr_mr == NULL);
1973                 cfs_list_del(&pmr->pmr_list);
1974
1975                 if (pmr->pmr_ipb != NULL) {
1976                         LIBCFS_FREE(pmr->pmr_ipb,
1977                                     IBLND_MAX_RDMA_FRAGS *
1978                                     sizeof(struct ib_phys_buf));
1979                 }
1980
1981                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1982         }
1983
1984         kiblnd_fini_pool(pool);
1985         if (ppo->ppo_hdev != NULL)
1986                 kiblnd_hdev_decref(ppo->ppo_hdev);
1987
1988         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1989 }
1990
1991 static inline int kiblnd_pmr_pool_size(int ncpts)
1992 {
1993         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1994
1995         return max(IBLND_PMR_POOL, size);
1996 }
1997
1998 static int
1999 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2000 {
2001         struct kib_pmr_pool     *ppo;
2002         struct kib_pool         *pool;
2003         kib_phys_mr_t           *pmr;
2004         int                     i;
2005
2006         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2007                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2008         if (ppo == NULL) {
2009                 CERROR("Failed to allocate PMR pool\n");
2010                 return -ENOMEM;
2011         }
2012
2013         pool = &ppo->ppo_pool;
2014         kiblnd_init_pool(ps, pool, size);
2015
2016         for (i = 0; i < size; i++) {
2017                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2018                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2019                 if (pmr == NULL)
2020                         break;
2021
2022                 pmr->pmr_pool = ppo;
2023                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2024                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2025                 if (pmr->pmr_ipb == NULL)
2026                         break;
2027
2028                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
2029         }
2030
2031         if (i < size) {
2032                 ps->ps_pool_destroy(pool);
2033                 return -ENOMEM;
2034         }
2035
2036         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2037         *pp_po = pool;
2038         return 0;
2039 }
2040
2041 static void
2042 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2043 {
2044         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2045         int             i;
2046
2047         LASSERT (pool->po_allocated == 0);
2048
2049         if (tpo->tpo_tx_pages != NULL) {
2050                 kiblnd_unmap_tx_pool(tpo);
2051                 kiblnd_free_pages(tpo->tpo_tx_pages);
2052         }
2053
2054         if (tpo->tpo_tx_descs == NULL)
2055                 goto out;
2056
2057         for (i = 0; i < pool->po_size; i++) {
2058                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2059
2060                 cfs_list_del(&tx->tx_list);
2061                 if (tx->tx_pages != NULL)
2062                         LIBCFS_FREE(tx->tx_pages,
2063                                     LNET_MAX_IOV *
2064                                     sizeof(*tx->tx_pages));
2065                 if (tx->tx_frags != NULL)
2066                         LIBCFS_FREE(tx->tx_frags,
2067                                     IBLND_MAX_RDMA_FRAGS *
2068                                             sizeof(*tx->tx_frags));
2069                 if (tx->tx_wrq != NULL)
2070                         LIBCFS_FREE(tx->tx_wrq,
2071                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2072                                     sizeof(*tx->tx_wrq));
2073                 if (tx->tx_sge != NULL)
2074                         LIBCFS_FREE(tx->tx_sge,
2075                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2076                                     sizeof(*tx->tx_sge));
2077                 if (tx->tx_rd != NULL)
2078                         LIBCFS_FREE(tx->tx_rd,
2079                                     offsetof(kib_rdma_desc_t,
2080                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2081         }
2082
2083         LIBCFS_FREE(tpo->tpo_tx_descs,
2084                     pool->po_size * sizeof(kib_tx_t));
2085 out:
2086         kiblnd_fini_pool(pool);
2087         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2088 }
2089
2090 static int kiblnd_tx_pool_size(int ncpts)
2091 {
2092         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2093
2094         return max(IBLND_TX_POOL, ntx);
2095 }
2096
2097 static int
2098 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2099 {
2100         int            i;
2101         int            npg;
2102         kib_pool_t    *pool;
2103         kib_tx_pool_t *tpo;
2104
2105         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2106         if (tpo == NULL) {
2107                 CERROR("Failed to allocate TX pool\n");
2108                 return -ENOMEM;
2109         }
2110
2111         pool = &tpo->tpo_pool;
2112         kiblnd_init_pool(ps, pool, size);
2113         tpo->tpo_tx_descs = NULL;
2114         tpo->tpo_tx_pages = NULL;
2115
2116         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2117         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2118                 CERROR("Can't allocate tx pages: %d\n", npg);
2119                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2120                 return -ENOMEM;
2121         }
2122
2123         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2124                          size * sizeof(kib_tx_t));
2125         if (tpo->tpo_tx_descs == NULL) {
2126                 CERROR("Can't allocate %d tx descriptors\n", size);
2127                 ps->ps_pool_destroy(pool);
2128                 return -ENOMEM;
2129         }
2130
2131         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2132
2133         for (i = 0; i < size; i++) {
2134                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2135
2136                 tx->tx_pool = tpo;
2137                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2138                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2139                                          lnet_cpt_table(), ps->ps_cpt,
2140                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2141                         if (tx->tx_pages == NULL)
2142                                 break;
2143                 }
2144
2145                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2146                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2147                 if (tx->tx_frags == NULL)
2148                         break;
2149
2150                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2151                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2152                                  sizeof(*tx->tx_wrq));
2153                 if (tx->tx_wrq == NULL)
2154                         break;
2155
2156                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2157                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2158                                  sizeof(*tx->tx_sge));
2159                 if (tx->tx_sge == NULL)
2160                         break;
2161
2162                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2163                                  offsetof(kib_rdma_desc_t,
2164                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2165                 if (tx->tx_rd == NULL)
2166                         break;
2167         }
2168
2169         if (i == size) {
2170                 kiblnd_map_tx_pool(tpo);
2171                 *pp_po = pool;
2172                 return 0;
2173         }
2174
2175         ps->ps_pool_destroy(pool);
2176         return -ENOMEM;
2177 }
2178
2179 static void
2180 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2181 {
2182         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2183                                              tps_poolset);
2184         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
2185
2186         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2187 }
2188
2189 void
2190 kiblnd_net_fini_pools(kib_net_t *net)
2191 {
2192         int     i;
2193
2194         cfs_cpt_for_each(i, lnet_cpt_table()) {
2195                 kib_tx_poolset_t        *tps;
2196                 kib_fmr_poolset_t       *fps;
2197                 kib_pmr_poolset_t       *pps;
2198
2199                 if (net->ibn_tx_ps != NULL) {
2200                         tps = net->ibn_tx_ps[i];
2201                         kiblnd_fini_poolset(&tps->tps_poolset);
2202                 }
2203
2204                 if (net->ibn_fmr_ps != NULL) {
2205                         fps = net->ibn_fmr_ps[i];
2206                         kiblnd_fini_fmr_poolset(fps);
2207                 }
2208
2209                 if (net->ibn_pmr_ps != NULL) {
2210                         pps = net->ibn_pmr_ps[i];
2211                         kiblnd_fini_poolset(&pps->pps_poolset);
2212                 }
2213         }
2214
2215         if (net->ibn_tx_ps != NULL) {
2216                 cfs_percpt_free(net->ibn_tx_ps);
2217                 net->ibn_tx_ps = NULL;
2218         }
2219
2220         if (net->ibn_fmr_ps != NULL) {
2221                 cfs_percpt_free(net->ibn_fmr_ps);
2222                 net->ibn_fmr_ps = NULL;
2223         }
2224
2225         if (net->ibn_pmr_ps != NULL) {
2226                 cfs_percpt_free(net->ibn_pmr_ps);
2227                 net->ibn_pmr_ps = NULL;
2228         }
2229 }
2230
2231 int
2232 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2233 {
2234         unsigned long   flags;
2235         int             cpt;
2236         int             rc;
2237         int             i;
2238
2239         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2240                                           sizeof(kib_tx_poolset_t));
2241         if (net->ibn_tx_ps == NULL) {
2242                 CERROR("Failed to allocate tx pool array\n");
2243                 return -ENOMEM;
2244         }
2245
2246         for (i = 0; i < ncpts; i++) {
2247                 cpt = (cpts == NULL) ? i : cpts[i];
2248                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2249                                          cpt, net, "TX",
2250                                          kiblnd_tx_pool_size(ncpts),
2251                                          kiblnd_create_tx_pool,
2252                                          kiblnd_destroy_tx_pool,
2253                                          kiblnd_tx_init, NULL);
2254                 if (rc != 0) {
2255                         CERROR("Failed to initialize TX pool\n");
2256                         goto failed;
2257                 }
2258         }
2259
2260         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2261         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2262             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2263                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2264                                            flags);
2265                 return 0;
2266         }
2267
2268         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2269
2270         if (*kiblnd_tunables.kib_fmr_pool_size <
2271             *kiblnd_tunables.kib_ntx / 4) {
2272                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2273                        *kiblnd_tunables.kib_fmr_pool_size,
2274                        *kiblnd_tunables.kib_ntx / 4);
2275                 goto failed;
2276         }
2277
2278         /* premapping can fail if ibd_nmr > 1, so we always create
2279          * FMR/PMR pool and map-on-demand if premapping failed */
2280
2281         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2282                                            sizeof(kib_fmr_poolset_t));
2283         if (net->ibn_fmr_ps == NULL) {
2284                 CERROR("Failed to allocate FMR pool array\n");
2285                 goto failed;
2286         }
2287
2288         for (i = 0; i < ncpts; i++) {
2289                 cpt = (cpts == NULL) ? i : cpts[i];
2290                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2291                                              kiblnd_fmr_pool_size(ncpts),
2292                                              kiblnd_fmr_flush_trigger(ncpts));
2293                 if (rc == -ENOSYS && i == 0) /* no FMR */
2294                         break; /* create PMR pool */
2295                 if (rc != 0)
2296                         goto failed; /* a real error */
2297         }
2298
2299         cfs_percpt_free(net->ibn_fmr_ps);
2300         net->ibn_fmr_ps = NULL;
2301
2302         if (*kiblnd_tunables.kib_pmr_pool_size <
2303             *kiblnd_tunables.kib_ntx / 4) {
2304                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2305                        *kiblnd_tunables.kib_pmr_pool_size,
2306                        *kiblnd_tunables.kib_ntx / 4);
2307                 goto failed;
2308         }
2309
2310         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2311                                            sizeof(kib_pmr_poolset_t));
2312         if (net->ibn_pmr_ps == NULL) {
2313                 CERROR("Failed to allocate PMR pool array\n");
2314                 goto failed;
2315         }
2316
2317         for (i = 0; i < ncpts; i++) {
2318                 cpt = (cpts == NULL) ? i : cpts[i];
2319                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2320                                          cpt, net, "PMR",
2321                                          kiblnd_pmr_pool_size(ncpts),
2322                                          kiblnd_create_pmr_pool,
2323                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2324                 if (rc != 0)
2325                         goto failed;
2326         }
2327
2328         return 0;
2329
2330  failed:
2331         kiblnd_net_fini_pools(net);
2332         return rc;
2333 }
2334
2335 static int
2336 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2337 {
2338         struct ib_device_attr *attr;
2339         int                    rc;
2340
2341         /* It's safe to assume a HCA can handle a page size
2342          * matching that of the native system */
2343         hdev->ibh_page_shift = PAGE_SHIFT;
2344         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2345         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2346
2347         LIBCFS_ALLOC(attr, sizeof(*attr));
2348         if (attr == NULL) {
2349                 CERROR("Out of memory\n");
2350                 return -ENOMEM;
2351         }
2352
2353         rc = ib_query_device(hdev->ibh_ibdev, attr);
2354         if (rc == 0)
2355                 hdev->ibh_mr_size = attr->max_mr_size;
2356
2357         LIBCFS_FREE(attr, sizeof(*attr));
2358
2359         if (rc != 0) {
2360                 CERROR("Failed to query IB device: %d\n", rc);
2361                 return rc;
2362         }
2363
2364         if (hdev->ibh_mr_size == ~0ULL) {
2365                 hdev->ibh_mr_shift = 64;
2366                 return 0;
2367         }
2368
2369         for (hdev->ibh_mr_shift = 0;
2370              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2371                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2372                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2373                         return 0;
2374         }
2375
2376         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2377         return -EINVAL;
2378 }
2379
2380 void
2381 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2382 {
2383         int     i;
2384
2385         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2386                 return;
2387
2388         for (i = 0; i < hdev->ibh_nmrs; i++) {
2389                 if (hdev->ibh_mrs[i] == NULL)
2390                         break;
2391
2392                 ib_dereg_mr(hdev->ibh_mrs[i]);
2393         }
2394
2395         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2396         hdev->ibh_mrs  = NULL;
2397         hdev->ibh_nmrs = 0;
2398 }
2399
2400 void
2401 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2402 {
2403         kiblnd_hdev_cleanup_mrs(hdev);
2404
2405         if (hdev->ibh_pd != NULL)
2406                 ib_dealloc_pd(hdev->ibh_pd);
2407
2408         if (hdev->ibh_cmid != NULL)
2409                 rdma_destroy_id(hdev->ibh_cmid);
2410
2411         LIBCFS_FREE(hdev, sizeof(*hdev));
2412 }
2413
2414 int
2415 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2416 {
2417         struct ib_mr *mr;
2418         int           i;
2419         int           rc;
2420         __u64         mm_size;
2421         __u64         mr_size;
2422         int           acflags = IB_ACCESS_LOCAL_WRITE |
2423                                 IB_ACCESS_REMOTE_WRITE;
2424
2425         rc = kiblnd_hdev_get_attr(hdev);
2426         if (rc != 0)
2427                 return rc;
2428
2429         if (hdev->ibh_mr_shift == 64) {
2430                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2431                 if (hdev->ibh_mrs == NULL) {
2432                         CERROR("Failed to allocate MRs table\n");
2433                         return -ENOMEM;
2434                 }
2435
2436                 hdev->ibh_mrs[0] = NULL;
2437                 hdev->ibh_nmrs   = 1;
2438
2439                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2440                 if (IS_ERR(mr)) {
2441                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2442                         kiblnd_hdev_cleanup_mrs(hdev);
2443                         return PTR_ERR(mr);
2444                 }
2445
2446                 hdev->ibh_mrs[0] = mr;
2447
2448                 goto out;
2449         }
2450
2451         mr_size = (1ULL << hdev->ibh_mr_shift);
2452         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2453
2454         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2455
2456         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2457                 /* it's 4T..., assume we will re-code at that time */
2458                 CERROR("Can't support memory size: x"LPX64
2459                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2460                 return -EINVAL;
2461         }
2462
2463         /* create an array of MRs to cover all memory */
2464         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2465         if (hdev->ibh_mrs == NULL) {
2466                 CERROR("Failed to allocate MRs' table\n");
2467                 return -ENOMEM;
2468         }
2469
2470         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2471
2472         for (i = 0; i < hdev->ibh_nmrs; i++) {
2473                 struct ib_phys_buf ipb;
2474                 __u64              iova;
2475
2476                 ipb.size = hdev->ibh_mr_size;
2477                 ipb.addr = i * mr_size;
2478                 iova     = ipb.addr;
2479
2480                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2481                 if (IS_ERR(mr)) {
2482                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2483                                " size "LPX64" : %ld\n",
2484                                ipb.addr, ipb.size, PTR_ERR(mr));
2485                         kiblnd_hdev_cleanup_mrs(hdev);
2486                         return PTR_ERR(mr);
2487                 }
2488
2489                 LASSERT (iova == ipb.addr);
2490
2491                 hdev->ibh_mrs[i] = mr;
2492         }
2493
2494 out:
2495         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2496                 LCONSOLE_INFO("Register global MR array, MR size: "
2497                               LPX64", array size: %d\n",
2498                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2499         return 0;
2500 }
2501
2502 static int
2503 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2504 {       /* DUMMY */
2505         return 0;
2506 }
2507
2508 static int
2509 kiblnd_dev_need_failover(kib_dev_t *dev)
2510 {
2511         struct rdma_cm_id  *cmid;
2512         struct sockaddr_in  srcaddr;
2513         struct sockaddr_in  dstaddr;
2514         int                 rc;
2515
2516         if (dev->ibd_hdev == NULL || /* initializing */
2517             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2518             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2519                 return 1;
2520
2521         /* XXX: it's UGLY, but I don't have better way to find
2522          * ib-bonding HCA failover because:
2523          *
2524          * a. no reliable CM event for HCA failover...
2525          * b. no OFED API to get ib_device for current net_device...
2526          *
2527          * We have only two choices at this point:
2528          *
2529          * a. rdma_bind_addr(), it will conflict with listener cmid
2530          * b. rdma_resolve_addr() to zero addr */
2531         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2532                                      IB_QPT_RC);
2533         if (IS_ERR(cmid)) {
2534                 rc = PTR_ERR(cmid);
2535                 CERROR("Failed to create cmid for failover: %d\n", rc);
2536                 return rc;
2537         }
2538
2539         memset(&srcaddr, 0, sizeof(srcaddr));
2540         srcaddr.sin_family      = AF_INET;
2541         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2542
2543         memset(&dstaddr, 0, sizeof(dstaddr));
2544         dstaddr.sin_family = AF_INET;
2545         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2546                                (struct sockaddr *)&dstaddr, 1);
2547         if (rc != 0 || cmid->device == NULL) {
2548                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2549                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2550                        cmid->device, rc);
2551                 rdma_destroy_id(cmid);
2552                 return rc;
2553         }
2554
2555         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2556                 /* don't need device failover */
2557                 rdma_destroy_id(cmid);
2558                 return 0;
2559         }
2560
2561         return 1;
2562 }
2563
2564 int
2565 kiblnd_dev_failover(kib_dev_t *dev)
2566 {
2567         CFS_LIST_HEAD      (zombie_tpo);
2568         CFS_LIST_HEAD      (zombie_ppo);
2569         CFS_LIST_HEAD      (zombie_fpo);
2570         struct rdma_cm_id  *cmid  = NULL;
2571         kib_hca_dev_t      *hdev  = NULL;
2572         kib_hca_dev_t      *old;
2573         struct ib_pd       *pd;
2574         kib_net_t          *net;
2575         struct sockaddr_in  addr;
2576         unsigned long       flags;
2577         int                 rc = 0;
2578         int                 i;
2579
2580         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2581                  dev->ibd_can_failover ||
2582                  dev->ibd_hdev == NULL);
2583
2584         rc = kiblnd_dev_need_failover(dev);
2585         if (rc <= 0)
2586                 goto out;
2587
2588         if (dev->ibd_hdev != NULL &&
2589             dev->ibd_hdev->ibh_cmid != NULL) {
2590                 /* XXX it's not good to close old listener at here,
2591                  * because we can fail to create new listener.
2592                  * But we have to close it now, otherwise rdma_bind_addr
2593                  * will return EADDRINUSE... How crap! */
2594                 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2595
2596                 cmid = dev->ibd_hdev->ibh_cmid;
2597                 /* make next schedule of kiblnd_dev_need_failover
2598                  * will return 1 for me */
2599                 dev->ibd_hdev->ibh_cmid  = NULL;
2600                 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2601
2602                 rdma_destroy_id(cmid);
2603         }
2604
2605         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2606                                      IB_QPT_RC);
2607         if (IS_ERR(cmid)) {
2608                 rc = PTR_ERR(cmid);
2609                 CERROR("Failed to create cmid for failover: %d\n", rc);
2610                 goto out;
2611         }
2612
2613         memset(&addr, 0, sizeof(addr));
2614         addr.sin_family      = AF_INET;
2615         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2616         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2617
2618         /* Bind to failover device or port */
2619         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2620         if (rc != 0 || cmid->device == NULL) {
2621                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2622                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2623                        cmid->device, rc);
2624                 rdma_destroy_id(cmid);
2625                 goto out;
2626         }
2627
2628         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2629         if (hdev == NULL) {
2630                 CERROR("Failed to allocate kib_hca_dev\n");
2631                 rdma_destroy_id(cmid);
2632                 rc = -ENOMEM;
2633                 goto out;
2634         }
2635
2636         atomic_set(&hdev->ibh_ref, 1);
2637         hdev->ibh_dev   = dev;
2638         hdev->ibh_cmid  = cmid;
2639         hdev->ibh_ibdev = cmid->device;
2640
2641         pd = ib_alloc_pd(cmid->device);
2642         if (IS_ERR(pd)) {
2643                 rc = PTR_ERR(pd);
2644                 CERROR("Can't allocate PD: %d\n", rc);
2645                 goto out;
2646         }
2647
2648         hdev->ibh_pd = pd;
2649
2650         rc = rdma_listen(cmid, 0);
2651         if (rc != 0) {
2652                 CERROR("Can't start new listener: %d\n", rc);
2653                 goto out;
2654         }
2655
2656         rc = kiblnd_hdev_setup_mrs(hdev);
2657         if (rc != 0) {
2658                 CERROR("Can't setup device: %d\n", rc);
2659                 goto out;
2660         }
2661
2662         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2663
2664         old = dev->ibd_hdev;
2665         dev->ibd_hdev = hdev; /* take over the refcount */
2666         hdev = old;
2667
2668         cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2669                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2670                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2671                                             &zombie_tpo);
2672
2673                         if (net->ibn_fmr_ps != NULL) {
2674                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2675                                                         &zombie_fpo);
2676
2677                         } else if (net->ibn_pmr_ps != NULL) {
2678                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2679                                                     pps_poolset, &zombie_ppo);
2680                         }
2681                 }
2682         }
2683
2684         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2685  out:
2686         if (!cfs_list_empty(&zombie_tpo))
2687                 kiblnd_destroy_pool_list(&zombie_tpo);
2688         if (!cfs_list_empty(&zombie_ppo))
2689                 kiblnd_destroy_pool_list(&zombie_ppo);
2690         if (!cfs_list_empty(&zombie_fpo))
2691                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2692         if (hdev != NULL)
2693                 kiblnd_hdev_decref(hdev);
2694
2695         if (rc != 0)
2696                 dev->ibd_failed_failover++;
2697         else
2698                 dev->ibd_failed_failover = 0;
2699
2700         return rc;
2701 }
2702
2703 void
2704 kiblnd_destroy_dev (kib_dev_t *dev)
2705 {
2706         LASSERT (dev->ibd_nnets == 0);
2707         LASSERT (cfs_list_empty(&dev->ibd_nets));
2708
2709         cfs_list_del(&dev->ibd_fail_list);
2710         cfs_list_del(&dev->ibd_list);
2711
2712         if (dev->ibd_hdev != NULL)
2713                 kiblnd_hdev_decref(dev->ibd_hdev);
2714
2715         LIBCFS_FREE(dev, sizeof(*dev));
2716 }
2717
2718 kib_dev_t *
2719 kiblnd_create_dev(char *ifname)
2720 {
2721         struct net_device *netdev;
2722         kib_dev_t         *dev;
2723         __u32              netmask;
2724         __u32              ip;
2725         int                up;
2726         int                rc;
2727
2728         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2729         if (rc != 0) {
2730                 CERROR("Can't query IPoIB interface %s: %d\n",
2731                        ifname, rc);
2732                 return NULL;
2733         }
2734
2735         if (!up) {
2736                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2737                 return NULL;
2738         }
2739
2740         LIBCFS_ALLOC(dev, sizeof(*dev));
2741         if (dev == NULL)
2742                 return NULL;
2743
2744         memset(dev, 0, sizeof(*dev));
2745 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2746         netdev = dev_get_by_name(&init_net, ifname);
2747 #else
2748         netdev = dev_get_by_name(ifname);
2749 #endif
2750         if (netdev == NULL) {
2751                 dev->ibd_can_failover = 0;
2752         } else {
2753                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2754                 dev_put(netdev);
2755         }
2756
2757         CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2758         CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2759         CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2760         dev->ibd_ifip = ip;
2761         strcpy(&dev->ibd_ifname[0], ifname);
2762
2763         /* initialize the device */
2764         rc = kiblnd_dev_failover(dev);
2765         if (rc != 0) {
2766                 CERROR("Can't initialize device: %d\n", rc);
2767                 LIBCFS_FREE(dev, sizeof(*dev));
2768                 return NULL;
2769         }
2770
2771         cfs_list_add_tail(&dev->ibd_list,
2772                           &kiblnd_data.kib_devs);
2773         return dev;
2774 }
2775
2776 void
2777 kiblnd_base_shutdown(void)
2778 {
2779         struct kib_sched_info   *sched;
2780         int                     i;
2781
2782         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2783
2784         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2785                cfs_atomic_read(&libcfs_kmemory));
2786
2787         switch (kiblnd_data.kib_init) {
2788         default:
2789                 LBUG();
2790
2791         case IBLND_INIT_ALL:
2792         case IBLND_INIT_DATA:
2793                 LASSERT (kiblnd_data.kib_peers != NULL);
2794                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2795                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2796                 }
2797                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2798                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2799
2800                 /* flag threads to terminate; wake and wait for them to die */
2801                 kiblnd_data.kib_shutdown = 1;
2802
2803                 /* NB: we really want to stop scheduler threads net by net
2804                  * instead of the whole module, this should be improved
2805                  * with dynamic configuration LNet */
2806                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2807                         cfs_waitq_broadcast(&sched->ibs_waitq);
2808
2809                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2810                 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2811
2812                 i = 2;
2813                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2814                         i++;
2815                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2816                                "Waiting for %d threads to terminate\n",
2817                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2818                         cfs_pause(cfs_time_seconds(1));
2819                 }
2820
2821                 /* fall through */
2822
2823         case IBLND_INIT_NOTHING:
2824                 break;
2825         }
2826
2827         if (kiblnd_data.kib_peers != NULL) {
2828                 LIBCFS_FREE(kiblnd_data.kib_peers,
2829                             sizeof(cfs_list_t) *
2830                             kiblnd_data.kib_peer_hash_size);
2831         }
2832
2833         if (kiblnd_data.kib_scheds != NULL)
2834                 cfs_percpt_free(kiblnd_data.kib_scheds);
2835
2836         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2837                cfs_atomic_read(&libcfs_kmemory));
2838
2839         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2840         PORTAL_MODULE_UNUSE;
2841 }
2842
2843 void
2844 kiblnd_shutdown (lnet_ni_t *ni)
2845 {
2846         kib_net_t        *net = ni->ni_data;
2847         cfs_rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2848         int               i;
2849         unsigned long     flags;
2850
2851         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2852
2853         if (net == NULL)
2854                 goto out;
2855
2856         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2857                cfs_atomic_read(&libcfs_kmemory));
2858
2859         cfs_write_lock_irqsave(g_lock, flags);
2860         net->ibn_shutdown = 1;
2861         cfs_write_unlock_irqrestore(g_lock, flags);
2862
2863         switch (net->ibn_init) {
2864         default:
2865                 LBUG();
2866
2867         case IBLND_INIT_ALL:
2868                 /* nuke all existing peers within this net */
2869                 kiblnd_del_peer(ni, LNET_NID_ANY);
2870
2871                 /* Wait for all peer state to clean up */
2872                 i = 2;
2873                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2874                         i++;
2875                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2876                                "%s: waiting for %d peers to disconnect\n",
2877                                libcfs_nid2str(ni->ni_nid),
2878                                cfs_atomic_read(&net->ibn_npeers));
2879                         cfs_pause(cfs_time_seconds(1));
2880                 }
2881
2882                 kiblnd_net_fini_pools(net);
2883
2884                 cfs_write_lock_irqsave(g_lock, flags);
2885                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2886                 net->ibn_dev->ibd_nnets--;
2887                 cfs_list_del(&net->ibn_list);
2888                 cfs_write_unlock_irqrestore(g_lock, flags);
2889
2890                 /* fall through */
2891
2892         case IBLND_INIT_NOTHING:
2893                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2894
2895                 if (net->ibn_dev != NULL &&
2896                     net->ibn_dev->ibd_nnets == 0)
2897                         kiblnd_destroy_dev(net->ibn_dev);
2898
2899                 break;
2900         }
2901
2902         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2903                cfs_atomic_read(&libcfs_kmemory));
2904
2905         net->ibn_init = IBLND_INIT_NOTHING;
2906         ni->ni_data = NULL;
2907
2908         LIBCFS_FREE(net, sizeof(*net));
2909
2910 out:
2911         if (cfs_list_empty(&kiblnd_data.kib_devs))
2912                 kiblnd_base_shutdown();
2913         return;
2914 }
2915
2916 int
2917 kiblnd_base_startup(void)
2918 {
2919         struct kib_sched_info   *sched;
2920         int                     rc;
2921         int                     i;
2922
2923         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2924
2925         PORTAL_MODULE_USE;
2926         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2927
2928         cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2929
2930         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2931         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2932
2933         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2934         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2935                      sizeof(cfs_list_t) *
2936                             kiblnd_data.kib_peer_hash_size);
2937         if (kiblnd_data.kib_peers == NULL) {
2938                 goto failed;
2939         }
2940         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2941                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2942
2943         cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2944         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2945         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2946         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2947         cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2948
2949         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2950                                                   sizeof(*sched));
2951         if (kiblnd_data.kib_scheds == NULL)
2952                 goto failed;
2953
2954         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2955                 int     nthrs;
2956
2957                 cfs_spin_lock_init(&sched->ibs_lock);
2958                 CFS_INIT_LIST_HEAD(&sched->ibs_conns);
2959                 cfs_waitq_init(&sched->ibs_waitq);
2960
2961                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2962                 if (*kiblnd_tunables.kib_nscheds > 0) {
2963                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2964                 } else {
2965                         /* max to half of CPUs, another half is reserved for
2966                          * upper layer modules */
2967                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2968                 }
2969
2970                 sched->ibs_nthreads_max = nthrs;
2971                 sched->ibs_cpt = i;
2972         }
2973
2974         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2975
2976         /* lists/ptrs/locks initialised */
2977         kiblnd_data.kib_init = IBLND_INIT_DATA;
2978         /*****************************************************/
2979
2980         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2981         if (rc != 0) {
2982                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2983                 goto failed;
2984         }
2985
2986         if (*kiblnd_tunables.kib_dev_failover != 0)
2987                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
2988
2989         if (rc != 0) {
2990                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2991                 goto failed;
2992         }
2993
2994         /* flag everything initialised */
2995         kiblnd_data.kib_init = IBLND_INIT_ALL;
2996         /*****************************************************/
2997
2998         return 0;
2999
3000  failed:
3001         kiblnd_base_shutdown();
3002         return -ENETDOWN;
3003 }
3004
3005 int
3006 kiblnd_start_schedulers(struct kib_sched_info *sched)
3007 {
3008         int     rc = 0;
3009         int     nthrs;
3010         int     i;
3011
3012         if (sched->ibs_nthreads == 0) {
3013                 if (*kiblnd_tunables.kib_nscheds > 0) {
3014                         nthrs = sched->ibs_nthreads_max;
3015                 } else {
3016                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3017                                                sched->ibs_cpt);
3018                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3019                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3020                 }
3021         } else {
3022                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3023                 /* increase one thread if there is new interface */
3024                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3025         }
3026
3027         for (i = 0; i < nthrs; i++) {
3028                 long    id;
3029
3030                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3031                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id);
3032                 if (rc == 0)
3033                         continue;
3034
3035                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3036                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3037                 break;
3038         }
3039
3040         sched->ibs_nthreads += i;
3041         return rc;
3042 }
3043
3044 int
3045 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3046 {
3047         int     cpt;
3048         int     rc;
3049         int     i;
3050
3051         for (i = 0; i < ncpts; i++) {
3052                 struct kib_sched_info *sched;
3053
3054                 cpt = (cpts == NULL) ? i : cpts[i];
3055                 sched = kiblnd_data.kib_scheds[cpt];
3056
3057                 if (!newdev && sched->ibs_nthreads > 0)
3058                         continue;
3059
3060                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3061                 if (rc != 0) {
3062                         CERROR("Failed to start scheduler threads for %s\n",
3063                                dev->ibd_ifname);
3064                         return rc;
3065                 }
3066         }
3067         return 0;
3068 }
3069
3070 kib_dev_t *
3071 kiblnd_dev_search(char *ifname)
3072 {
3073         kib_dev_t       *alias = NULL;
3074         kib_dev_t       *dev;
3075         char            *colon;
3076         char            *colon2;
3077
3078         colon = strchr(ifname, ':');
3079         cfs_list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3080                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3081                         return dev;
3082
3083                 if (alias != NULL)
3084                         continue;
3085
3086                 colon2 = strchr(dev->ibd_ifname, ':');
3087                 if (colon != NULL)
3088                         *colon = 0;
3089                 if (colon2 != NULL)
3090                         *colon2 = 0;
3091
3092                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3093                         alias = dev;
3094
3095                 if (colon != NULL)
3096                         *colon = ':';
3097                 if (colon2 != NULL)
3098                         *colon2 = ':';
3099         }
3100         return alias;
3101 }
3102
3103 int
3104 kiblnd_startup (lnet_ni_t *ni)
3105 {
3106         char                     *ifname;
3107         kib_dev_t                *ibdev = NULL;
3108         kib_net_t                *net;
3109         struct timeval            tv;
3110         unsigned long             flags;
3111         int                       rc;
3112         int                       newdev;
3113
3114         LASSERT (ni->ni_lnd == &the_o2iblnd);
3115
3116         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3117                 rc = kiblnd_base_startup();
3118                 if (rc != 0)
3119                         return rc;
3120         }
3121
3122         LIBCFS_ALLOC(net, sizeof(*net));
3123         ni->ni_data = net;
3124         if (net == NULL)
3125                 goto failed;
3126
3127         memset(net, 0, sizeof(*net));
3128
3129         cfs_gettimeofday(&tv);
3130         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3131
3132         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3133         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3134         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3135         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3136
3137         if (ni->ni_interfaces[0] != NULL) {
3138                 /* Use the IPoIB interface specified in 'networks=' */
3139
3140                 CLASSERT (LNET_MAX_INTERFACES > 1);
3141                 if (ni->ni_interfaces[1] != NULL) {
3142                         CERROR("Multiple interfaces not supported\n");
3143                         goto failed;
3144                 }
3145
3146                 ifname = ni->ni_interfaces[0];
3147         } else {
3148                 ifname = *kiblnd_tunables.kib_default_ipif;
3149         }
3150
3151         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3152                 CERROR("IPoIB interface name too long: %s\n", ifname);
3153                 goto failed;
3154         }
3155
3156         ibdev = kiblnd_dev_search(ifname);
3157
3158         newdev = ibdev == NULL;
3159         /* hmm...create kib_dev even for alias */
3160         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3161                 ibdev = kiblnd_create_dev(ifname);
3162
3163         if (ibdev == NULL)
3164                 goto failed;
3165
3166         net->ibn_dev = ibdev;
3167         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3168
3169         rc = kiblnd_dev_start_threads(ibdev, newdev,
3170                                       ni->ni_cpts, ni->ni_ncpts);
3171         if (rc != 0)
3172                 goto failed;
3173
3174         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3175         if (rc != 0) {
3176                 CERROR("Failed to initialize NI pools: %d\n", rc);
3177                 goto failed;
3178         }
3179
3180         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3181         ibdev->ibd_nnets++;
3182         cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3183         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3184
3185         net->ibn_init = IBLND_INIT_ALL;
3186
3187         return 0;
3188
3189 failed:
3190         if (net->ibn_dev == NULL && ibdev != NULL)
3191                 kiblnd_destroy_dev(ibdev);
3192
3193         kiblnd_shutdown(ni);
3194
3195         CDEBUG(D_NET, "kiblnd_startup failed\n");
3196         return -ENETDOWN;
3197 }
3198
3199 void __exit
3200 kiblnd_module_fini (void)
3201 {
3202         lnet_unregister_lnd(&the_o2iblnd);
3203         kiblnd_tunables_fini();
3204 }
3205
3206 int __init
3207 kiblnd_module_init (void)
3208 {
3209         int    rc;
3210
3211         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3212         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3213                   <= IBLND_MSG_SIZE);
3214         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3215                   <= IBLND_MSG_SIZE);
3216
3217         rc = kiblnd_tunables_init();
3218         if (rc != 0)
3219                 return rc;
3220
3221         lnet_register_lnd(&the_o2iblnd);
3222
3223         return 0;
3224 }
3225
3226 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3227 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3228 MODULE_LICENSE("GPL");
3229
3230 module_init(kiblnd_module_init);
3231 module_exit(kiblnd_module_fini);