lnet/klnds/o2iblnd/o2iblnd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  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         unsigned long   flags;
 332
 333         LASSERT (net != NULL);
 334         LASSERT (nid != LNET_NID_ANY);
 335
 336         LIBCFS_ALLOC(peer, sizeof(*peer));
 337         if (peer == NULL) {
 338                 CERROR("Cannot allocate peer\n");
 339                 return -ENOMEM;
 340         }
 341
 342         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 343
 344         peer->ibp_ni = ni;
 345         peer->ibp_nid = nid;
 346         peer->ibp_error = 0;
 347         peer->ibp_last_alive = 0;
 348         cfs_atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 349
 350         CFS_INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 351         CFS_INIT_LIST_HEAD(&peer->ibp_conns);
 352         CFS_INIT_LIST_HEAD(&peer->ibp_tx_queue);
 353
 354         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 355
 356         /* always called with a ref on ni, which prevents ni being shutdown */
 357         LASSERT (net->ibn_shutdown == 0);
 358
 359         /* npeers only grows with the global lock held */
 360         cfs_atomic_inc(&net->ibn_npeers);
 361
 362         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 363
 364         *peerp = peer;
 365         return 0;
 366 }
 367
 368 void
 369 kiblnd_destroy_peer (kib_peer_t *peer)
 370 {
 371         kib_net_t *net = peer->ibp_ni->ni_data;
 372
 373         LASSERT (net != NULL);
 374         LASSERT (cfs_atomic_read(&peer->ibp_refcount) == 0);
 375         LASSERT (!kiblnd_peer_active(peer));
 376         LASSERT (peer->ibp_connecting == 0);
 377         LASSERT (peer->ibp_accepting == 0);
 378         LASSERT (cfs_list_empty(&peer->ibp_conns));
 379         LASSERT (cfs_list_empty(&peer->ibp_tx_queue));
 380
 381         LIBCFS_FREE(peer, sizeof(*peer));
 382
 383         /* NB a peer's connections keep a reference on their peer until
 384          * they are destroyed, so we can be assured that _all_ state to do
 385          * with this peer has been cleaned up when its refcount drops to
 386          * zero. */
 387         cfs_atomic_dec(&net->ibn_npeers);
 388 }
 389
 390 kib_peer_t *
 391 kiblnd_find_peer_locked (lnet_nid_t nid)
 392 {
 393         /* the caller is responsible for accounting the additional reference
 394          * that this creates */
 395         cfs_list_t       *peer_list = kiblnd_nid2peerlist(nid);
 396         cfs_list_t       *tmp;
 397         kib_peer_t       *peer;
 398
 399         cfs_list_for_each (tmp, peer_list) {
 400
 401                 peer = cfs_list_entry(tmp, kib_peer_t, ibp_list);
 402
 403                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 404                          peer->ibp_accepting > 0 ||
 405                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
 406
 407                 if (peer->ibp_nid != nid)
 408                         continue;
 409
 410                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 411                        peer, libcfs_nid2str(nid),
 412                        cfs_atomic_read(&peer->ibp_refcount),
 413                        peer->ibp_version);
 414                 return peer;
 415         }
 416         return NULL;
 417 }
 418
 419 void
 420 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 421 {
 422         LASSERT (cfs_list_empty(&peer->ibp_conns));
 423
 424         LASSERT (kiblnd_peer_active(peer));
 425         cfs_list_del_init(&peer->ibp_list);
 426         /* lose peerlist's ref */
 427         kiblnd_peer_decref(peer);
 428 }
 429
 430 int
 431 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 432                       lnet_nid_t *nidp, int *count)
 433 {
 434         kib_peer_t            *peer;
 435         cfs_list_t            *ptmp;
 436         int                    i;
 437         unsigned long          flags;
 438
 439         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 440
 441         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 442
 443                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 444
 445                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 446                         LASSERT (peer->ibp_connecting > 0 ||
 447                                  peer->ibp_accepting > 0 ||
 448                                  !cfs_list_empty(&peer->ibp_conns));
 449
 450                         if (peer->ibp_ni != ni)
 451                                 continue;
 452
 453                         if (index-- > 0)
 454                                 continue;
 455
 456                         *nidp = peer->ibp_nid;
 457                         *count = cfs_atomic_read(&peer->ibp_refcount);
 458
 459                         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 460                                                    flags);
 461                         return 0;
 462                 }
 463         }
 464
 465         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 466         return -ENOENT;
 467 }
 468
 469 void
 470 kiblnd_del_peer_locked (kib_peer_t *peer)
 471 {
 472         cfs_list_t           *ctmp;
 473         cfs_list_t           *cnxt;
 474         kib_conn_t           *conn;
 475
 476         if (cfs_list_empty(&peer->ibp_conns)) {
 477                 kiblnd_unlink_peer_locked(peer);
 478         } else {
 479                 cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 480                         conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 481
 482                         kiblnd_close_conn_locked(conn, 0);
 483                 }
 484                 /* NB closing peer's last conn unlinked it. */
 485         }
 486         /* NB peer now unlinked; might even be freed if the peer table had the
 487          * last ref on it. */
 488 }
 489
 490 int
 491 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 492 {
 493         CFS_LIST_HEAD         (zombies);
 494         cfs_list_t            *ptmp;
 495         cfs_list_t            *pnxt;
 496         kib_peer_t            *peer;
 497         int                    lo;
 498         int                    hi;
 499         int                    i;
 500         unsigned long          flags;
 501         int                    rc = -ENOENT;
 502
 503         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 504
 505         if (nid != LNET_NID_ANY) {
 506                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 507         } else {
 508                 lo = 0;
 509                 hi = kiblnd_data.kib_peer_hash_size - 1;
 510         }
 511
 512         for (i = lo; i <= hi; i++) {
 513                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 514                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 515                         LASSERT (peer->ibp_connecting > 0 ||
 516                                  peer->ibp_accepting > 0 ||
 517                                  !cfs_list_empty(&peer->ibp_conns));
 518
 519                         if (peer->ibp_ni != ni)
 520                                 continue;
 521
 522                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 523                                 continue;
 524
 525                         if (!cfs_list_empty(&peer->ibp_tx_queue)) {
 526                                 LASSERT (cfs_list_empty(&peer->ibp_conns));
 527
 528                                 cfs_list_splice_init(&peer->ibp_tx_queue,
 529                                                      &zombies);
 530                         }
 531
 532                         kiblnd_del_peer_locked(peer);
 533                         rc = 0;         /* matched something */
 534                 }
 535         }
 536
 537         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 538
 539         kiblnd_txlist_done(ni, &zombies, -EIO);
 540
 541         return rc;
 542 }
 543
 544 kib_conn_t *
 545 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 546 {
 547         kib_peer_t            *peer;
 548         cfs_list_t            *ptmp;
 549         kib_conn_t            *conn;
 550         cfs_list_t            *ctmp;
 551         int                    i;
 552         unsigned long          flags;
 553
 554         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 555
 556         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 557                 cfs_list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 558
 559                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
 560                         LASSERT (peer->ibp_connecting > 0 ||
 561                                  peer->ibp_accepting > 0 ||
 562                                  !cfs_list_empty(&peer->ibp_conns));
 563
 564                         if (peer->ibp_ni != ni)
 565                                 continue;
 566
 567                         cfs_list_for_each (ctmp, &peer->ibp_conns) {
 568                                 if (index-- > 0)
 569                                         continue;
 570
 571                                 conn = cfs_list_entry(ctmp, kib_conn_t,
 572                                                       ibc_list);
 573                                 kiblnd_conn_addref(conn);
 574                                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 575                                                            flags);
 576                                 return conn;
 577                         }
 578                 }
 579         }
 580
 581         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 582         return NULL;
 583 }
 584
 585 void
 586 kiblnd_debug_rx (kib_rx_t *rx)
 587 {
 588         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 589                rx, rx->rx_status, rx->rx_msg->ibm_type,
 590                rx->rx_msg->ibm_credits);
 591 }
 592
 593 void
 594 kiblnd_debug_tx (kib_tx_t *tx)
 595 {
 596         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 597                "cookie "LPX64" msg %s%s type %x cred %d\n",
 598                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 599                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 600                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 601                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 602                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 603 }
 604
 605 void
 606 kiblnd_debug_conn (kib_conn_t *conn)
 607 {
 608         cfs_list_t           *tmp;
 609         int                   i;
 610
 611         cfs_spin_lock(&conn->ibc_lock);
 612
 613         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 614                cfs_atomic_read(&conn->ibc_refcount), conn,
 615                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 616         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 617                conn->ibc_state, conn->ibc_noops_posted,
 618                conn->ibc_nsends_posted, conn->ibc_credits,
 619                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 620         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 621
 622         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 623         cfs_list_for_each(tmp, &conn->ibc_early_rxs)
 624                 kiblnd_debug_rx(cfs_list_entry(tmp, kib_rx_t, rx_list));
 625
 626         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 627         cfs_list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 628                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 629
 630         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 631         cfs_list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 632                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 633
 634         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 635         cfs_list_for_each(tmp, &conn->ibc_tx_queue)
 636                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 637
 638         CDEBUG(D_CONSOLE, "   active_txs:\n");
 639         cfs_list_for_each(tmp, &conn->ibc_active_txs)
 640                 kiblnd_debug_tx(cfs_list_entry(tmp, kib_tx_t, tx_list));
 641
 642         CDEBUG(D_CONSOLE, "   rxs:\n");
 643         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 644                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 645
 646         cfs_spin_unlock(&conn->ibc_lock);
 647 }
 648
 649 int
 650 kiblnd_translate_mtu(int value)
 651 {
 652         switch (value) {
 653         default:
 654                 return -1;
 655         case 0:
 656                 return 0;
 657         case 256:
 658                 return IB_MTU_256;
 659         case 512:
 660                 return IB_MTU_512;
 661         case 1024:
 662                 return IB_MTU_1024;
 663         case 2048:
 664                 return IB_MTU_2048;
 665         case 4096:
 666                 return IB_MTU_4096;
 667         }
 668 }
 669
 670 static void
 671 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 672 {
 673         int           mtu;
 674
 675         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 676         if (cmid->route.path_rec == NULL)
 677                 return;
 678
 679         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 680         LASSERT (mtu >= 0);
 681         if (mtu != 0)
 682                 cmid->route.path_rec->mtu = mtu;
 683 }
 684
 685 kib_conn_t *
 686 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 687                    int state, int version)
 688 {
 689         /* CAVEAT EMPTOR:
 690          * If the new conn is created successfully it takes over the caller's
 691          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 692          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 693          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 694          * to destroy 'cmid' here since I'm called from the CM which still has
 695          * its ref on 'cmid'). */
 696         cfs_rwlock_t           *glock = &kiblnd_data.kib_global_lock;
 697         kib_net_t              *net = peer->ibp_ni->ni_data;
 698         kib_dev_t              *dev = net->ibn_dev;
 699         struct ib_qp_init_attr *init_qp_attr;
 700         kib_conn_t             *conn;
 701         struct ib_cq           *cq;
 702         unsigned long           flags;
 703         int                     rc;
 704         int                     i;
 705
 706         LASSERT (net != NULL);
 707         LASSERT (!cfs_in_interrupt());
 708
 709         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 710         if (init_qp_attr == NULL) {
 711                 CERROR("Can't allocate qp_attr for %s\n",
 712                        libcfs_nid2str(peer->ibp_nid));
 713                 goto failed_0;
 714         }
 715
 716         LIBCFS_ALLOC(conn, sizeof(*conn));
 717         if (conn == NULL) {
 718                 CERROR("Can't allocate connection for %s\n",
 719                        libcfs_nid2str(peer->ibp_nid));
 720                 goto failed_1;
 721         }
 722
 723         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 724
 725         conn->ibc_state = IBLND_CONN_INIT;
 726         conn->ibc_version = version;
 727         conn->ibc_peer = peer;                  /* I take the caller's ref */
 728         cmid->context = conn;                   /* for future CM callbacks */
 729         conn->ibc_cmid = cmid;
 730
 731         CFS_INIT_LIST_HEAD(&conn->ibc_early_rxs);
 732         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue);
 733         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 734         CFS_INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 735         CFS_INIT_LIST_HEAD(&conn->ibc_active_txs);
 736         cfs_spin_lock_init(&conn->ibc_lock);
 737
 738         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 739         if (conn->ibc_connvars == NULL) {
 740                 CERROR("Can't allocate in-progress connection state\n");
 741                 goto failed_2;
 742         }
 743         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 744
 745         cfs_write_lock_irqsave(glock, flags);
 746         if (dev->ibd_failover) {
 747                 cfs_write_unlock_irqrestore(glock, flags);
 748                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 749                 goto failed_2;
 750         }
 751
 752         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 753                 /* wakeup failover thread and teardown connection */
 754                 if (kiblnd_dev_can_failover(dev)) {
 755                         cfs_list_add_tail(&dev->ibd_fail_list,
 756                                       &kiblnd_data.kib_failed_devs);
 757                         cfs_waitq_signal(&kiblnd_data.kib_failover_waitq);
 758                 }
 759
 760                 cfs_write_unlock_irqrestore(glock, flags);
 761                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 762                        cmid->device->name, dev->ibd_ifname);
 763                 goto failed_2;
 764         }
 765
 766         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 767         conn->ibc_hdev = dev->ibd_hdev;
 768
 769         kiblnd_setup_mtu_locked(cmid);
 770
 771         cfs_write_unlock_irqrestore(glock, flags);
 772
 773         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 774         if (conn->ibc_rxs == NULL) {
 775                 CERROR("Cannot allocate RX buffers\n");
 776                 goto failed_2;
 777         }
 778         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 779
 780         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages,
 781                                 IBLND_RX_MSG_PAGES(version));
 782         if (rc != 0)
 783                 goto failed_2;
 784
 785         kiblnd_map_rx_descs(conn);
 786
 787 #ifdef HAVE_OFED_IB_COMP_VECTOR
 788         cq = ib_create_cq(cmid->device,
 789                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 790                           IBLND_CQ_ENTRIES(version), 0);
 791 #else
 792         cq = ib_create_cq(cmid->device,
 793                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 794                           IBLND_CQ_ENTRIES(version));
 795 #endif
 796         if (IS_ERR(cq)) {
 797                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 798                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 799                 goto failed_2;
 800         }
 801
 802         conn->ibc_cq = cq;
 803
 804         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 805         if (rc != 0) {
 806                 CERROR("Can't request completion notificiation: %d\n", rc);
 807                 goto failed_2;
 808         }
 809
 810         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 811         init_qp_attr->event_handler = kiblnd_qp_event;
 812         init_qp_attr->qp_context = conn;
 813         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 814         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 815         init_qp_attr->cap.max_send_sge = 1;
 816         init_qp_attr->cap.max_recv_sge = 1;
 817         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 818         init_qp_attr->qp_type = IB_QPT_RC;
 819         init_qp_attr->send_cq = cq;
 820         init_qp_attr->recv_cq = cq;
 821
 822         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 823         if (rc != 0) {
 824                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 825                        rc, init_qp_attr->cap.max_send_wr,
 826                        init_qp_attr->cap.max_recv_wr);
 827                 goto failed_2;
 828         }
 829
 830         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 831
 832         /* 1 ref for caller and each rxmsg */
 833         cfs_atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 834         conn->ibc_nrx = IBLND_RX_MSGS(version);
 835
 836         /* post receives */
 837         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 838                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 839                                     IBLND_POSTRX_NO_CREDIT);
 840                 if (rc != 0) {
 841                         CERROR("Can't post rxmsg: %d\n", rc);
 842
 843                         /* Make posted receives complete */
 844                         kiblnd_abort_receives(conn);
 845
 846                         /* correct # of posted buffers
 847                          * NB locking needed now I'm racing with completion */
 848                         cfs_spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
 849                                               flags);
 850                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 851                         cfs_spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 852                                                    flags);
 853
 854                         /* cmid will be destroyed by CM(ofed) after cm_callback
 855                          * returned, so we can't refer it anymore
 856                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 857                         rdma_destroy_qp(conn->ibc_cmid);
 858                         conn->ibc_cmid = NULL;
 859
 860                         /* Drop my own and unused rxbuffer refcounts */
 861                         while (i++ <= IBLND_RX_MSGS(version))
 862                                 kiblnd_conn_decref(conn);
 863
 864                         return NULL;
 865                 }
 866         }
 867
 868         /* Init successful! */
 869         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 870                  state == IBLND_CONN_PASSIVE_WAIT);
 871         conn->ibc_state = state;
 872
 873         /* 1 more conn */
 874         cfs_atomic_inc(&net->ibn_nconns);
 875         return conn;
 876
 877  failed_2:
 878         kiblnd_destroy_conn(conn);
 879  failed_1:
 880         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 881  failed_0:
 882         return NULL;
 883 }
 884
 885 void
 886 kiblnd_destroy_conn (kib_conn_t *conn)
 887 {
 888         struct rdma_cm_id *cmid = conn->ibc_cmid;
 889         kib_peer_t        *peer = conn->ibc_peer;
 890         int                rc;
 891
 892         LASSERT (!cfs_in_interrupt());
 893         LASSERT (cfs_atomic_read(&conn->ibc_refcount) == 0);
 894         LASSERT (cfs_list_empty(&conn->ibc_early_rxs));
 895         LASSERT (cfs_list_empty(&conn->ibc_tx_queue));
 896         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_rsrvd));
 897         LASSERT (cfs_list_empty(&conn->ibc_tx_queue_nocred));
 898         LASSERT (cfs_list_empty(&conn->ibc_active_txs));
 899         LASSERT (conn->ibc_noops_posted == 0);
 900         LASSERT (conn->ibc_nsends_posted == 0);
 901
 902         switch (conn->ibc_state) {
 903         default:
 904                 /* conn must be completely disengaged from the network */
 905                 LBUG();
 906
 907         case IBLND_CONN_DISCONNECTED:
 908                 /* connvars should have been freed already */
 909                 LASSERT (conn->ibc_connvars == NULL);
 910                 break;
 911
 912         case IBLND_CONN_INIT:
 913                 break;
 914         }
 915
 916         /* conn->ibc_cmid might be destroyed by CM already */
 917         if (cmid != NULL && cmid->qp != NULL)
 918                 rdma_destroy_qp(cmid);
 919
 920         if (conn->ibc_cq != NULL) {
 921                 rc = ib_destroy_cq(conn->ibc_cq);
 922                 if (rc != 0)
 923                         CWARN("Error destroying CQ: %d\n", rc);
 924         }
 925
 926         if (conn->ibc_rx_pages != NULL)
 927                 kiblnd_unmap_rx_descs(conn);
 928
 929         if (conn->ibc_rxs != NULL) {
 930                 LIBCFS_FREE(conn->ibc_rxs,
 931                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 932         }
 933
 934         if (conn->ibc_connvars != NULL)
 935                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 936
 937         if (conn->ibc_hdev != NULL)
 938                 kiblnd_hdev_decref(conn->ibc_hdev);
 939
 940         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 941         if (conn->ibc_state != IBLND_CONN_INIT) {
 942                 kib_net_t *net = peer->ibp_ni->ni_data;
 943
 944                 kiblnd_peer_decref(peer);
 945                 rdma_destroy_id(cmid);
 946                 cfs_atomic_dec(&net->ibn_nconns);
 947         }
 948
 949         LIBCFS_FREE(conn, sizeof(*conn));
 950 }
 951
 952 int
 953 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 954 {
 955         kib_conn_t             *conn;
 956         cfs_list_t             *ctmp;
 957         cfs_list_t             *cnxt;
 958         int                     count = 0;
 959
 960         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 961                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 962
 963                 CDEBUG(D_NET, "Closing conn -> %s, "
 964                               "version: %x, reason: %d\n",
 965                        libcfs_nid2str(peer->ibp_nid),
 966                        conn->ibc_version, why);
 967
 968                 kiblnd_close_conn_locked(conn, why);
 969                 count++;
 970         }
 971
 972         return count;
 973 }
 974
 975 int
 976 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 977                                  int version, __u64 incarnation)
 978 {
 979         kib_conn_t             *conn;
 980         cfs_list_t             *ctmp;
 981         cfs_list_t             *cnxt;
 982         int                     count = 0;
 983
 984         cfs_list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 985                 conn = cfs_list_entry(ctmp, kib_conn_t, ibc_list);
 986
 987                 if (conn->ibc_version     == version &&
 988                     conn->ibc_incarnation == incarnation)
 989                         continue;
 990
 991                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
 992                               "incarnation:"LPX64"(%x, "LPX64")\n",
 993                        libcfs_nid2str(peer->ibp_nid),
 994                        conn->ibc_version, conn->ibc_incarnation,
 995                        version, incarnation);
 996
 997                 kiblnd_close_conn_locked(conn, -ESTALE);
 998                 count++;
 999         }
1000
1001         return count;
1002 }
1003
1004 int
1005 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1006 {
1007         kib_peer_t             *peer;
1008         cfs_list_t             *ptmp;
1009         cfs_list_t             *pnxt;
1010         int                     lo;
1011         int                     hi;
1012         int                     i;
1013         unsigned long           flags;
1014         int                     count = 0;
1015
1016         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1017
1018         if (nid != LNET_NID_ANY)
1019                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1020         else {
1021                 lo = 0;
1022                 hi = kiblnd_data.kib_peer_hash_size - 1;
1023         }
1024
1025         for (i = lo; i <= hi; i++) {
1026                 cfs_list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1027
1028                         peer = cfs_list_entry(ptmp, kib_peer_t, ibp_list);
1029                         LASSERT (peer->ibp_connecting > 0 ||
1030                                  peer->ibp_accepting > 0 ||
1031                                  !cfs_list_empty(&peer->ibp_conns));
1032
1033                         if (peer->ibp_ni != ni)
1034                                 continue;
1035
1036                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1037                                 continue;
1038
1039                         count += kiblnd_close_peer_conns_locked(peer, 0);
1040                 }
1041         }
1042
1043         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1044
1045         /* wildcards always succeed */
1046         if (nid == LNET_NID_ANY)
1047                 return 0;
1048
1049         return (count == 0) ? -ENOENT : 0;
1050 }
1051
1052 int
1053 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1054 {
1055         struct libcfs_ioctl_data *data = arg;
1056         int                       rc = -EINVAL;
1057
1058         switch(cmd) {
1059         case IOC_LIBCFS_GET_PEER: {
1060                 lnet_nid_t   nid = 0;
1061                 int          count = 0;
1062
1063                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1064                                           &nid, &count);
1065                 data->ioc_nid    = nid;
1066                 data->ioc_count  = count;
1067                 break;
1068         }
1069
1070         case IOC_LIBCFS_DEL_PEER: {
1071                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1072                 break;
1073         }
1074         case IOC_LIBCFS_GET_CONN: {
1075                 kib_conn_t *conn;
1076
1077                 rc = 0;
1078                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1079                 if (conn == NULL) {
1080                         rc = -ENOENT;
1081                         break;
1082                 }
1083
1084                 LASSERT (conn->ibc_cmid != NULL);
1085                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1086                 if (conn->ibc_cmid->route.path_rec == NULL)
1087                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1088                 else
1089                         data->ioc_u32[0] =
1090                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1091                 kiblnd_conn_decref(conn);
1092                 break;
1093         }
1094         case IOC_LIBCFS_CLOSE_CONNECTION: {
1095                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1096                 break;
1097         }
1098
1099         default:
1100                 break;
1101         }
1102
1103         return rc;
1104 }
1105
1106 void
1107 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1108 {
1109         cfs_time_t         last_alive = 0;
1110         cfs_time_t         now = cfs_time_current();
1111         cfs_rwlock_t      *glock = &kiblnd_data.kib_global_lock;
1112         kib_peer_t        *peer;
1113         unsigned long      flags;
1114
1115         cfs_read_lock_irqsave(glock, flags);
1116
1117         peer = kiblnd_find_peer_locked(nid);
1118         if (peer != NULL) {
1119                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1120                          peer->ibp_accepting > 0 ||
1121                          !cfs_list_empty(&peer->ibp_conns));  /* active conn */
1122                 last_alive = peer->ibp_last_alive;
1123         }
1124
1125         cfs_read_unlock_irqrestore(glock, flags);
1126
1127         if (last_alive != 0)
1128                 *when = last_alive;
1129
1130         /* peer is not persistent in hash, trigger peer creation
1131          * and connection establishment with a NULL tx */
1132         if (peer == NULL)
1133                 kiblnd_launch_tx(ni, NULL, nid);
1134
1135         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1136                libcfs_nid2str(nid), peer,
1137                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1138         return;
1139 }
1140
1141 void
1142 kiblnd_free_pages (kib_pages_t *p)
1143 {
1144         int         npages = p->ibp_npages;
1145         int         i;
1146
1147         for (i = 0; i < npages; i++)
1148                 if (p->ibp_pages[i] != NULL)
1149                         __free_page(p->ibp_pages[i]);
1150
1151         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1152 }
1153
1154 int
1155 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1156 {
1157         kib_pages_t   *p;
1158         int            i;
1159
1160         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1161         if (p == NULL) {
1162                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1163                 return -ENOMEM;
1164         }
1165
1166         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1167         p->ibp_npages = npages;
1168
1169         for (i = 0; i < npages; i++) {
1170                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1171                 if (p->ibp_pages[i] == NULL) {
1172                         CERROR("Can't allocate page %d of %d\n", i, npages);
1173                         kiblnd_free_pages(p);
1174                         return -ENOMEM;
1175                 }
1176         }
1177
1178         *pp = p;
1179         return 0;
1180 }
1181
1182 void
1183 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1184 {
1185         kib_rx_t *rx;
1186         int       i;
1187
1188         LASSERT (conn->ibc_rxs != NULL);
1189         LASSERT (conn->ibc_hdev != NULL);
1190
1191         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1192                 rx = &conn->ibc_rxs[i];
1193
1194                 LASSERT (rx->rx_nob >= 0); /* not posted */
1195
1196                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1197                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1198                                                           rx->rx_msgaddr),
1199                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1200         }
1201
1202         kiblnd_free_pages(conn->ibc_rx_pages);
1203
1204         conn->ibc_rx_pages = NULL;
1205 }
1206
1207 void
1208 kiblnd_map_rx_descs(kib_conn_t *conn)
1209 {
1210         kib_rx_t       *rx;
1211         struct page    *pg;
1212         int             pg_off;
1213         int             ipg;
1214         int             i;
1215
1216         for (pg_off = ipg = i = 0;
1217              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1218                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1219                 rx = &conn->ibc_rxs[i];
1220
1221                 rx->rx_conn = conn;
1222                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1223
1224                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1225                                                        rx->rx_msg, IBLND_MSG_SIZE,
1226                                                        DMA_FROM_DEVICE);
1227                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1228                                                    rx->rx_msgaddr));
1229                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1230
1231                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1232                        i, rx->rx_msg, rx->rx_msgaddr,
1233                        lnet_page2phys(pg) + pg_off);
1234
1235                 pg_off += IBLND_MSG_SIZE;
1236                 LASSERT (pg_off <= PAGE_SIZE);
1237
1238                 if (pg_off == PAGE_SIZE) {
1239                         pg_off = 0;
1240                         ipg++;
1241                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1242                 }
1243         }
1244 }
1245
1246 static void
1247 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1248 {
1249         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1250         kib_tx_t       *tx;
1251         int             i;
1252
1253         LASSERT (tpo->tpo_pool.po_allocated == 0);
1254
1255         if (hdev == NULL)
1256                 return;
1257
1258         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1259                 tx = &tpo->tpo_tx_descs[i];
1260                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1261                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1262                                                           tx->tx_msgaddr),
1263                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1264         }
1265
1266         kiblnd_hdev_decref(hdev);
1267         tpo->tpo_hdev = NULL;
1268 }
1269
1270 static kib_hca_dev_t *
1271 kiblnd_current_hdev(kib_dev_t *dev)
1272 {
1273         kib_hca_dev_t *hdev;
1274         unsigned long  flags;
1275         int            i = 0;
1276
1277         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1278         while (dev->ibd_failover) {
1279                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1280                 if (i++ % 50 == 0)
1281                         CDEBUG(D_NET, "Wait for dev(%s) failover\n", dev->ibd_ifname);
1282                 cfs_schedule_timeout(cfs_time_seconds(1) / 100);
1283
1284                 cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1285         }
1286
1287         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1288         hdev = dev->ibd_hdev;
1289
1290         cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1291
1292         return hdev;
1293 }
1294
1295 static void
1296 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1297 {
1298         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1299         kib_pool_t     *pool  = &tpo->tpo_pool;
1300         kib_net_t      *net   = pool->po_owner->ps_net;
1301         kib_dev_t      *dev   = net->ibn_dev;
1302         struct page    *page;
1303         kib_tx_t       *tx;
1304         int             page_offset;
1305         int             ipage;
1306         int             i;
1307
1308         LASSERT (net != NULL);
1309
1310         /* pre-mapped messages are not bigger than 1 page */
1311         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1312
1313         /* No fancy arithmetic when we do the buffer calculations */
1314         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1315
1316         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1317
1318         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1319                 page = txpgs->ibp_pages[ipage];
1320                 tx = &tpo->tpo_tx_descs[i];
1321
1322                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1323                                            page_offset);
1324
1325                 tx->tx_msgaddr = kiblnd_dma_map_single(
1326                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1327                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1328                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1329                                                    tx->tx_msgaddr));
1330                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1331
1332                 cfs_list_add(&tx->tx_list, &pool->po_free_list);
1333
1334                 page_offset += IBLND_MSG_SIZE;
1335                 LASSERT (page_offset <= PAGE_SIZE);
1336
1337                 if (page_offset == PAGE_SIZE) {
1338                         page_offset = 0;
1339                         ipage++;
1340                         LASSERT (ipage <= txpgs->ibp_npages);
1341                 }
1342         }
1343 }
1344
1345 struct ib_mr *
1346 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1347 {
1348         __u64   index;
1349
1350         LASSERT (hdev->ibh_mrs[0] != NULL);
1351
1352         if (hdev->ibh_nmrs == 1)
1353                 return hdev->ibh_mrs[0];
1354
1355         index = addr >> hdev->ibh_mr_shift;
1356
1357         if (index <  hdev->ibh_nmrs &&
1358             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1359                 return hdev->ibh_mrs[index];
1360
1361         return NULL;
1362 }
1363
1364 struct ib_mr *
1365 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1366 {
1367         struct ib_mr *prev_mr;
1368         struct ib_mr *mr;
1369         int           i;
1370
1371         LASSERT (hdev->ibh_mrs[0] != NULL);
1372
1373         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1374             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1375                 return NULL;
1376
1377         if (hdev->ibh_nmrs == 1)
1378                 return hdev->ibh_mrs[0];
1379
1380         for (i = 0, mr = prev_mr = NULL;
1381              i < rd->rd_nfrags; i++) {
1382                 mr = kiblnd_find_dma_mr(hdev,
1383                                         rd->rd_frags[i].rf_addr,
1384                                         rd->rd_frags[i].rf_nob);
1385                 if (prev_mr == NULL)
1386                         prev_mr = mr;
1387
1388                 if (mr == NULL || prev_mr != mr) {
1389                         /* Can't covered by one single MR */
1390                         mr = NULL;
1391                         break;
1392                 }
1393         }
1394
1395         return mr;
1396 }
1397
1398 void
1399 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1400 {
1401         LASSERT (pool->fpo_map_count == 0);
1402
1403         if (pool->fpo_fmr_pool != NULL)
1404                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1405
1406         if (pool->fpo_hdev != NULL)
1407                 kiblnd_hdev_decref(pool->fpo_hdev);
1408
1409         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1410 }
1411
1412 void
1413 kiblnd_destroy_fmr_pool_list(cfs_list_t *head)
1414 {
1415         kib_fmr_pool_t *pool;
1416
1417         while (!cfs_list_empty(head)) {
1418                 pool = cfs_list_entry(head->next, kib_fmr_pool_t, fpo_list);
1419                 cfs_list_del(&pool->fpo_list);
1420                 kiblnd_destroy_fmr_pool(pool);
1421         }
1422 }
1423
1424 int
1425 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1426 {
1427         /* FMR pool for RDMA */
1428         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1429         kib_fmr_pool_t          *fpo;
1430         struct ib_fmr_pool_param param = {
1431                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1432                 .page_shift        = PAGE_SHIFT,
1433                 .access            = (IB_ACCESS_LOCAL_WRITE |
1434                                       IB_ACCESS_REMOTE_WRITE),
1435                 .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1436                 .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1437                 .flush_function    = NULL,
1438                 .flush_arg         = NULL,
1439                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1440         int rc;
1441
1442         LIBCFS_ALLOC(fpo, sizeof(kib_fmr_pool_t));
1443         if (fpo == NULL)
1444                 return -ENOMEM;
1445
1446         memset(fpo, 0, sizeof(kib_fmr_pool_t));
1447         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1448         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1449         if (IS_ERR(fpo->fpo_fmr_pool)) {
1450                 CERROR("Failed to create FMR pool: %ld\n",
1451                        PTR_ERR(fpo->fpo_fmr_pool));
1452                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1453                 kiblnd_hdev_decref(fpo->fpo_hdev);
1454                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1455                 return rc;
1456         }
1457
1458         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1459         fpo->fpo_owner    = fps;
1460         *pp_fpo = fpo;
1461
1462         return 0;
1463 }
1464
1465 static void
1466 kiblnd_fail_fmr_pool_set(kib_fmr_poolset_t *fps, cfs_list_t *zombies)
1467 {
1468         cfs_spin_lock(&fps->fps_lock);
1469
1470         while (!cfs_list_empty(&fps->fps_pool_list)) {
1471                 kib_fmr_pool_t *fpo = cfs_list_entry(fps->fps_pool_list.next,
1472                                                  kib_fmr_pool_t, fpo_list);
1473                 fpo->fpo_failed = 1;
1474                 cfs_list_del(&fpo->fpo_list);
1475                 if (fpo->fpo_map_count == 0)
1476                         cfs_list_add(&fpo->fpo_list, zombies);
1477                 else
1478                         cfs_list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1479         }
1480
1481         cfs_spin_unlock(&fps->fps_lock);
1482 }
1483
1484 static void
1485 kiblnd_fini_fmr_pool_set(kib_fmr_poolset_t *fps)
1486 {
1487         kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1488         kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1489 }
1490
1491 static int
1492 kiblnd_init_fmr_pool_set(kib_fmr_poolset_t *fps, kib_net_t *net)
1493 {
1494         kib_fmr_pool_t *fpo;
1495         int             rc;
1496
1497         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1498
1499         fps->fps_net = net;
1500         cfs_spin_lock_init(&fps->fps_lock);
1501         CFS_INIT_LIST_HEAD(&fps->fps_pool_list);
1502         CFS_INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1503         rc = kiblnd_create_fmr_pool(fps, &fpo);
1504         if (rc == 0)
1505                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1506
1507         return rc;
1508 }
1509
1510 static int
1511 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1512 {
1513         if (fpo->fpo_map_count != 0) /* still in use */
1514                 return 0;
1515         if (fpo->fpo_failed)
1516                 return 1;
1517         return cfs_time_aftereq(now, fpo->fpo_deadline);
1518 }
1519
1520 void
1521 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1522 {
1523         CFS_LIST_HEAD     (zombies);
1524         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1525         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1526         cfs_time_t         now = cfs_time_current();
1527         kib_fmr_pool_t    *tmp;
1528         int                rc;
1529
1530         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1531         LASSERT (rc == 0);
1532
1533         if (status != 0) {
1534                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1535                 LASSERT (rc == 0);
1536         }
1537
1538         fmr->fmr_pool = NULL;
1539         fmr->fmr_pfmr = NULL;
1540
1541         cfs_spin_lock(&fps->fps_lock);
1542         fpo->fpo_map_count --;  /* decref the pool */
1543
1544         cfs_list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1545                 /* the first pool is persistent */
1546                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1547                         continue;
1548
1549                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1550                         cfs_list_move(&fpo->fpo_list, &zombies);
1551                         fps->fps_version ++;
1552                 }
1553         }
1554         cfs_spin_unlock(&fps->fps_lock);
1555
1556         if (!cfs_list_empty(&zombies))
1557                 kiblnd_destroy_fmr_pool_list(&zombies);
1558 }
1559
1560 int
1561 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1562                     __u64 iov, kib_fmr_t *fmr)
1563 {
1564         struct ib_pool_fmr *pfmr;
1565         kib_fmr_pool_t     *fpo;
1566         __u64               version;
1567         int                 rc;
1568
1569         LASSERT (fps->fps_net->ibn_with_fmr);
1570  again:
1571         cfs_spin_lock(&fps->fps_lock);
1572         version = fps->fps_version;
1573         cfs_list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1574                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1575                 fpo->fpo_map_count ++;
1576                 cfs_spin_unlock(&fps->fps_lock);
1577
1578                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1579                                             pages, npages, iov);
1580                 if (likely(!IS_ERR(pfmr))) {
1581                         fmr->fmr_pool = fpo;
1582                         fmr->fmr_pfmr = pfmr;
1583                         return 0;
1584                 }
1585
1586                 cfs_spin_lock(&fps->fps_lock);
1587                 fpo->fpo_map_count --;
1588                 if (PTR_ERR(pfmr) != -EAGAIN) {
1589                         cfs_spin_unlock(&fps->fps_lock);
1590                         return PTR_ERR(pfmr);
1591                 }
1592
1593                 /* EAGAIN and ... */
1594                 if (version != fps->fps_version) {
1595                         cfs_spin_unlock(&fps->fps_lock);
1596                         goto again;
1597                 }
1598         }
1599
1600         if (fps->fps_increasing) {
1601                 cfs_spin_unlock(&fps->fps_lock);
1602                 CDEBUG(D_NET, "Another thread is allocating new "
1603                               "FMR pool, waiting for her to complete\n");
1604                 cfs_schedule();
1605                 goto again;
1606
1607         }
1608
1609         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1610                 /* someone failed recently */
1611                 cfs_spin_unlock(&fps->fps_lock);
1612                 return -EAGAIN;
1613         }
1614
1615         fps->fps_increasing = 1;
1616         cfs_spin_unlock(&fps->fps_lock);
1617
1618         CDEBUG(D_NET, "Allocate new FMR pool\n");
1619         rc = kiblnd_create_fmr_pool(fps, &fpo);
1620         cfs_spin_lock(&fps->fps_lock);
1621         fps->fps_increasing = 0;
1622         if (rc == 0) {
1623                 fps->fps_version ++;
1624                 cfs_list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1625         } else {
1626                 fps->fps_next_retry = cfs_time_shift(10);
1627         }
1628         cfs_spin_unlock(&fps->fps_lock);
1629
1630         goto again;
1631 }
1632
1633 static void
1634 kiblnd_fini_pool(kib_pool_t *pool)
1635 {
1636         LASSERT (cfs_list_empty(&pool->po_free_list));
1637         LASSERT (pool->po_allocated == 0);
1638
1639         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1640 }
1641
1642 static void
1643 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1644 {
1645         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1646
1647         memset(pool, 0, sizeof(kib_pool_t));
1648         CFS_INIT_LIST_HEAD(&pool->po_free_list);
1649         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1650         pool->po_owner    = ps;
1651         pool->po_size     = size;
1652 }
1653
1654 void
1655 kiblnd_destroy_pool_list(cfs_list_t *head)
1656 {
1657         kib_pool_t *pool;
1658
1659         while (!cfs_list_empty(head)) {
1660                 pool = cfs_list_entry(head->next, kib_pool_t, po_list);
1661                 cfs_list_del(&pool->po_list);
1662
1663                 LASSERT (pool->po_owner != NULL);
1664                 pool->po_owner->ps_pool_destroy(pool);
1665         }
1666 }
1667
1668 static void
1669 kiblnd_fail_pool_set(kib_poolset_t *ps, cfs_list_t *zombies)
1670 {
1671         cfs_spin_lock(&ps->ps_lock);
1672         while (!cfs_list_empty(&ps->ps_pool_list)) {
1673                 kib_pool_t *po = cfs_list_entry(ps->ps_pool_list.next,
1674                                             kib_pool_t, po_list);
1675                 po->po_failed = 1;
1676                 cfs_list_del(&po->po_list);
1677                 if (po->po_allocated == 0)
1678                         cfs_list_add(&po->po_list, zombies);
1679                 else
1680                         cfs_list_add(&po->po_list, &ps->ps_failed_pool_list);
1681         }
1682         cfs_spin_unlock(&ps->ps_lock);
1683 }
1684
1685 static void
1686 kiblnd_fini_pool_set(kib_poolset_t *ps)
1687 {
1688         kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1689         kiblnd_destroy_pool_list(&ps->ps_pool_list);
1690 }
1691
1692 static int
1693 kiblnd_init_pool_set(kib_poolset_t *ps, kib_net_t *net,
1694                      char *name, int size,
1695                      kib_ps_pool_create_t po_create,
1696                      kib_ps_pool_destroy_t po_destroy,
1697                      kib_ps_node_init_t nd_init,
1698                      kib_ps_node_fini_t nd_fini)
1699 {
1700         kib_pool_t    *pool;
1701         int            rc;
1702
1703         memset(ps, 0, sizeof(kib_poolset_t));
1704
1705         ps->ps_net          = net;
1706         ps->ps_pool_create  = po_create;
1707         ps->ps_pool_destroy = po_destroy;
1708         ps->ps_node_init    = nd_init;
1709         ps->ps_node_fini    = nd_fini;
1710         ps->ps_pool_size    = size;
1711         strncpy(ps->ps_name, name, IBLND_POOL_NAME_LEN);
1712         cfs_spin_lock_init(&ps->ps_lock);
1713         CFS_INIT_LIST_HEAD(&ps->ps_pool_list);
1714         CFS_INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1715
1716         rc = ps->ps_pool_create(ps, size, &pool);
1717         if (rc == 0)
1718                 cfs_list_add(&pool->po_list, &ps->ps_pool_list);
1719         else
1720                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1721
1722         return rc;
1723 }
1724
1725 static int
1726 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1727 {
1728         if (pool->po_allocated != 0) /* still in use */
1729                 return 0;
1730         if (pool->po_failed)
1731                 return 1;
1732         return cfs_time_aftereq(now, pool->po_deadline);
1733 }
1734
1735 void
1736 kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node)
1737 {
1738         CFS_LIST_HEAD  (zombies);
1739         kib_poolset_t  *ps = pool->po_owner;
1740         kib_pool_t     *tmp;
1741         cfs_time_t      now = cfs_time_current();
1742
1743         cfs_spin_lock(&ps->ps_lock);
1744
1745         if (ps->ps_node_fini != NULL)
1746                 ps->ps_node_fini(pool, node);
1747
1748         LASSERT (pool->po_allocated > 0);
1749         cfs_list_add(node, &pool->po_free_list);
1750         pool->po_allocated --;
1751
1752         cfs_list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1753                 /* the first pool is persistent */
1754                 if (ps->ps_pool_list.next == &pool->po_list)
1755                         continue;
1756
1757                 if (kiblnd_pool_is_idle(pool, now))
1758                         cfs_list_move(&pool->po_list, &zombies);
1759         }
1760         cfs_spin_unlock(&ps->ps_lock);
1761
1762         if (!cfs_list_empty(&zombies))
1763                 kiblnd_destroy_pool_list(&zombies);
1764 }
1765
1766 cfs_list_t *
1767 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1768 {
1769         cfs_list_t            *node;
1770         kib_pool_t            *pool;
1771         int                    rc;
1772
1773  again:
1774         cfs_spin_lock(&ps->ps_lock);
1775         cfs_list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1776                 if (cfs_list_empty(&pool->po_free_list))
1777                         continue;
1778
1779                 pool->po_allocated ++;
1780                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1781                 node = pool->po_free_list.next;
1782                 cfs_list_del(node);
1783
1784                 if (ps->ps_node_init != NULL) {
1785                         /* still hold the lock */
1786                         ps->ps_node_init(pool, node);
1787                 }
1788                 cfs_spin_unlock(&ps->ps_lock);
1789                 return node;
1790         }
1791
1792         /* no available tx pool and ... */
1793         if (ps->ps_increasing) {
1794                 /* another thread is allocating a new pool */
1795                 cfs_spin_unlock(&ps->ps_lock);
1796                 CDEBUG(D_NET, "Another thread is allocating new "
1797                        "%s pool, waiting for her to complete\n",
1798                        ps->ps_name);
1799                 cfs_schedule();
1800                 goto again;
1801         }
1802
1803         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1804                 /* someone failed recently */
1805                 cfs_spin_unlock(&ps->ps_lock);
1806                 return NULL;
1807         }
1808
1809         ps->ps_increasing = 1;
1810         cfs_spin_unlock(&ps->ps_lock);
1811
1812         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1813
1814         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1815
1816         cfs_spin_lock(&ps->ps_lock);
1817         ps->ps_increasing = 0;
1818         if (rc == 0) {
1819                 cfs_list_add_tail(&pool->po_list, &ps->ps_pool_list);
1820         } else {
1821                 /* retry 10 seconds later */
1822                 ps->ps_next_retry = cfs_time_shift(10);
1823                 CERROR("Can't allocate new %s pool because out of memory\n",
1824                        ps->ps_name);
1825         }
1826         cfs_spin_unlock(&ps->ps_lock);
1827
1828         goto again;
1829 }
1830
1831 void
1832 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1833 {
1834         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1835         struct ib_mr        *mr  = pmr->pmr_mr;
1836
1837         pmr->pmr_mr = NULL;
1838         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1839         if (mr != NULL)
1840                 ib_dereg_mr(mr);
1841 }
1842
1843 int
1844 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1845                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1846 {
1847         kib_phys_mr_t *pmr;
1848         cfs_list_t    *node;
1849         int            rc;
1850         int            i;
1851
1852         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1853         if (node == NULL) {
1854                 CERROR("Failed to allocate PMR descriptor\n");
1855                 return -ENOMEM;
1856         }
1857
1858         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1859         if (pmr->pmr_pool->ppo_hdev != hdev) {
1860                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1861                 return -EAGAIN;
1862         }
1863
1864         for (i = 0; i < rd->rd_nfrags; i ++) {
1865                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1866                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1867         }
1868
1869         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1870                                      pmr->pmr_ipb, rd->rd_nfrags,
1871                                      IB_ACCESS_LOCAL_WRITE |
1872                                      IB_ACCESS_REMOTE_WRITE,
1873                                      iova);
1874         if (!IS_ERR(pmr->pmr_mr)) {
1875                 pmr->pmr_iova = *iova;
1876                 *pp_pmr = pmr;
1877                 return 0;
1878         }
1879
1880         rc = PTR_ERR(pmr->pmr_mr);
1881         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1882
1883         pmr->pmr_mr = NULL;
1884         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1885
1886         return rc;
1887 }
1888
1889 static void
1890 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1891 {
1892         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1893         kib_phys_mr_t  *pmr;
1894
1895         LASSERT (pool->po_allocated == 0);
1896
1897         while (!cfs_list_empty(&pool->po_free_list)) {
1898                 pmr = cfs_list_entry(pool->po_free_list.next,
1899                                      kib_phys_mr_t, pmr_list);
1900
1901                 LASSERT (pmr->pmr_mr == NULL);
1902                 cfs_list_del(&pmr->pmr_list);
1903
1904                 if (pmr->pmr_ipb != NULL) {
1905                         LIBCFS_FREE(pmr->pmr_ipb,
1906                                     IBLND_MAX_RDMA_FRAGS *
1907                                     sizeof(struct ib_phys_buf));
1908                 }
1909
1910                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1911         }
1912
1913         kiblnd_fini_pool(pool);
1914         if (ppo->ppo_hdev != NULL)
1915                 kiblnd_hdev_decref(ppo->ppo_hdev);
1916
1917         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1918 }
1919
1920 static int
1921 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1922 {
1923         kib_pmr_pool_t      *ppo;
1924         kib_pool_t          *pool;
1925         kib_phys_mr_t       *pmr;
1926         int                  i;
1927
1928         LIBCFS_ALLOC(ppo, sizeof(kib_pmr_pool_t));
1929         if (ppo == NULL) {
1930                 CERROR("Failed to allocate PMR pool\n");
1931                 return -ENOMEM;
1932         }
1933
1934         pool = &ppo->ppo_pool;
1935         kiblnd_init_pool(ps, pool, size);
1936
1937         for (i = 0; i < size; i++) {
1938                 LIBCFS_ALLOC(pmr, sizeof(kib_phys_mr_t));
1939                 if (pmr == NULL)
1940                         break;
1941
1942                 memset(pmr, 0, sizeof(kib_phys_mr_t));
1943                 pmr->pmr_pool = ppo;
1944                 LIBCFS_ALLOC(pmr->pmr_ipb,
1945                              IBLND_MAX_RDMA_FRAGS *
1946                              sizeof(struct ib_phys_buf));
1947                 if (pmr->pmr_ipb == NULL)
1948                         break;
1949
1950                 cfs_list_add(&pmr->pmr_list, &pool->po_free_list);
1951         }
1952
1953         if (i < size) {
1954                 ps->ps_pool_destroy(pool);
1955                 return -ENOMEM;
1956         }
1957
1958         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1959         *pp_po = pool;
1960         return 0;
1961 }
1962
1963 static void
1964 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1965 {
1966         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1967         int             i;
1968
1969         LASSERT (pool->po_allocated == 0);
1970
1971         if (tpo->tpo_tx_pages != NULL) {
1972                 kiblnd_unmap_tx_pool(tpo);
1973                 kiblnd_free_pages(tpo->tpo_tx_pages);
1974         }
1975
1976         if (tpo->tpo_tx_descs == NULL)
1977                 goto out;
1978
1979         for (i = 0; i < pool->po_size; i++) {
1980                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1981
1982                 cfs_list_del(&tx->tx_list);
1983                 if (tx->tx_pages != NULL)
1984                         LIBCFS_FREE(tx->tx_pages,
1985                                     LNET_MAX_IOV *
1986                                     sizeof(*tx->tx_pages));
1987                 if (tx->tx_frags != NULL)
1988                         LIBCFS_FREE(tx->tx_frags,
1989                                     IBLND_MAX_RDMA_FRAGS *
1990                                             sizeof(*tx->tx_frags));
1991                 if (tx->tx_wrq != NULL)
1992                         LIBCFS_FREE(tx->tx_wrq,
1993                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1994                                     sizeof(*tx->tx_wrq));
1995                 if (tx->tx_sge != NULL)
1996                         LIBCFS_FREE(tx->tx_sge,
1997                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1998                                     sizeof(*tx->tx_sge));
1999                 if (tx->tx_rd != NULL)
2000                         LIBCFS_FREE(tx->tx_rd,
2001                                     offsetof(kib_rdma_desc_t,
2002                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2003         }
2004
2005         LIBCFS_FREE(tpo->tpo_tx_descs,
2006                     pool->po_size * sizeof(kib_tx_t));
2007 out:
2008         kiblnd_fini_pool(pool);
2009         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2010 }
2011
2012 static int
2013 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2014 {
2015         int            i;
2016         int            npg;
2017         kib_pool_t    *pool;
2018         kib_tx_pool_t *tpo;
2019
2020         LIBCFS_ALLOC(tpo, sizeof(kib_tx_pool_t));
2021         if (tpo == NULL) {
2022                 CERROR("Failed to allocate TX pool\n");
2023                 return -ENOMEM;
2024         }
2025
2026         pool = &tpo->tpo_pool;
2027         kiblnd_init_pool(ps, pool, size);
2028         tpo->tpo_tx_descs = NULL;
2029         tpo->tpo_tx_pages = NULL;
2030
2031         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2032         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, npg) != 0) {
2033                 CERROR("Can't allocate tx pages: %d\n", npg);
2034                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2035                 return -ENOMEM;
2036         }
2037
2038         LIBCFS_ALLOC (tpo->tpo_tx_descs, size * sizeof(kib_tx_t));
2039         if (tpo->tpo_tx_descs == NULL) {
2040                 CERROR("Can't allocate %d tx descriptors\n", size);
2041                 ps->ps_pool_destroy(pool);
2042                 return -ENOMEM;
2043         }
2044
2045         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2046
2047         for (i = 0; i < size; i++) {
2048                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2049
2050                 tx->tx_pool = tpo;
2051                 if (ps->ps_net->ibn_with_fmr){
2052                         LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
2053                                      sizeof(*tx->tx_pages));
2054                         if (tx->tx_pages == NULL)
2055                                 break;
2056                 }
2057
2058                 LIBCFS_ALLOC(tx->tx_frags,
2059                              IBLND_MAX_RDMA_FRAGS *
2060                              sizeof(*tx->tx_frags));
2061                 if (tx->tx_frags == NULL)
2062                         break;
2063
2064                 LIBCFS_ALLOC(tx->tx_wrq,
2065                              (1 + IBLND_MAX_RDMA_FRAGS) *
2066                              sizeof(*tx->tx_wrq));
2067                 if (tx->tx_wrq == NULL)
2068                         break;
2069
2070                 LIBCFS_ALLOC(tx->tx_sge,
2071                              (1 + IBLND_MAX_RDMA_FRAGS) *
2072                              sizeof(*tx->tx_sge));
2073                 if (tx->tx_sge == NULL)
2074                         break;
2075
2076                 LIBCFS_ALLOC(tx->tx_rd,
2077                              offsetof(kib_rdma_desc_t,
2078                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
2079                 if (tx->tx_rd == NULL)
2080                         break;
2081         }
2082
2083         if (i == size) {
2084                 kiblnd_map_tx_pool(tpo);
2085                 *pp_po = pool;
2086                 return 0;
2087         }
2088
2089         ps->ps_pool_destroy(pool);
2090         return -ENOMEM;
2091 }
2092
2093 static void
2094 kiblnd_tx_init(kib_pool_t *pool, cfs_list_t *node)
2095 {
2096         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2097                                              tps_poolset);
2098         kib_tx_t         *tx  = cfs_list_entry(node, kib_tx_t, tx_list);
2099
2100         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2101 }
2102
2103 void
2104 kiblnd_ni_fini_pools(kib_net_t *net)
2105 {
2106         kiblnd_fini_pool_set(&net->ibn_tx_ps.tps_poolset);
2107         if (net->ibn_with_fmr)
2108                 kiblnd_fini_fmr_pool_set(&net->ibn_fmr_ps);
2109         else if (net->ibn_with_pmr)
2110                 kiblnd_fini_pool_set(&net->ibn_pmr_ps.pps_poolset);
2111 }
2112
2113 int
2114 kiblnd_net_init_pools(kib_net_t *net)
2115 {
2116         kib_fmr_poolset_t *fps = &net->ibn_fmr_ps;
2117         kib_pmr_poolset_t *pps = &net->ibn_pmr_ps;
2118         kib_tx_poolset_t  *tps = &net->ibn_tx_ps;
2119         unsigned long      flags;
2120         int                rc;
2121
2122         if (*kiblnd_tunables.kib_fmr_pool_size <
2123             *kiblnd_tunables.kib_ntx / 4) {
2124                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2125                        *kiblnd_tunables.kib_fmr_pool_size,
2126                        *kiblnd_tunables.kib_ntx / 4);
2127                 return -EINVAL;
2128         }
2129
2130         if (*kiblnd_tunables.kib_pmr_pool_size <
2131             *kiblnd_tunables.kib_ntx / 4) {
2132                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2133                        *kiblnd_tunables.kib_pmr_pool_size,
2134                        *kiblnd_tunables.kib_ntx / 4);
2135                 return -EINVAL;
2136         }
2137
2138         cfs_read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2139         if (*kiblnd_tunables.kib_map_on_demand > 0 ||
2140             net->ibn_dev->ibd_hdev->ibh_nmrs > 1) {
2141                 /* premapping can fail if ibd_nmr > 1, so we always create
2142                  * FMR/PMR pool and map-on-demand if premapping failed */
2143                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2144                 rc = kiblnd_init_fmr_pool_set(fps, net);
2145                 if (rc == 0) {
2146                         net->ibn_with_fmr = 1;
2147                 } else if (rc == -ENOSYS) {
2148                         rc = kiblnd_init_pool_set(&pps->pps_poolset, net, "PMR",
2149                                                   *kiblnd_tunables.kib_pmr_pool_size,
2150                                                   kiblnd_create_pmr_pool,
2151                                                   kiblnd_destroy_pmr_pool,
2152                                                   NULL, NULL);
2153                         if (rc == 0)
2154                                 net->ibn_with_pmr = 1;
2155                 }
2156                 if (rc != 0)
2157                         return rc;
2158         } else {
2159                 cfs_read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2160         }
2161
2162         rc = kiblnd_init_pool_set(&tps->tps_poolset, net, "TX", IBLND_TX_MSGS(),
2163                                   kiblnd_create_tx_pool, kiblnd_destroy_tx_pool,
2164                                   kiblnd_tx_init, NULL);
2165         if (rc == 0)
2166                 return 0;
2167
2168         if (net->ibn_with_fmr)
2169                 kiblnd_fini_fmr_pool_set(fps);
2170         else if (net->ibn_with_pmr)
2171                 kiblnd_fini_pool_set(&pps->pps_poolset);
2172
2173         return rc;
2174 }
2175
2176 static int
2177 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2178 {
2179         struct ib_device_attr *attr;
2180         int                    rc;
2181
2182         /* It's safe to assume a HCA can handle a page size
2183          * matching that of the native system */
2184         hdev->ibh_page_shift = PAGE_SHIFT;
2185         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2186         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2187
2188         LIBCFS_ALLOC(attr, sizeof(*attr));
2189         if (attr == NULL) {
2190                 CERROR("Out of memory\n");
2191                 return -ENOMEM;
2192         }
2193
2194         rc = ib_query_device(hdev->ibh_ibdev, attr);
2195         if (rc == 0)
2196                 hdev->ibh_mr_size = attr->max_mr_size;
2197
2198         LIBCFS_FREE(attr, sizeof(*attr));
2199
2200         if (rc != 0) {
2201                 CERROR("Failed to query IB device: %d\n", rc);
2202                 return rc;
2203         }
2204
2205 #ifdef HAVE_OFED_TRANSPORT_IWARP
2206         /* XXX We can't trust this value returned by Chelsio driver, it's wrong
2207          * and we have reported the bug, remove these in the future when Chelsio
2208          * bug got fixed. */
2209         if (rdma_node_get_transport(hdev->ibh_ibdev->node_type) ==
2210             RDMA_TRANSPORT_IWARP)
2211                 hdev->ibh_mr_size = (1ULL << 32) - 1;
2212 #endif
2213
2214         if (hdev->ibh_mr_size == ~0ULL) {
2215                 hdev->ibh_mr_shift = 64;
2216                 return 0;
2217         }
2218
2219         for (hdev->ibh_mr_shift = 0;
2220              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2221                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2222                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2223                         return 0;
2224         }
2225
2226         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2227         return -EINVAL;
2228 }
2229
2230 void
2231 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2232 {
2233         int     i;
2234
2235         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2236                 return;
2237
2238         for (i = 0; i < hdev->ibh_nmrs; i++) {
2239                 if (hdev->ibh_mrs[i] == NULL)
2240                         break;
2241
2242                 ib_dereg_mr(hdev->ibh_mrs[i]);
2243         }
2244
2245         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2246         hdev->ibh_mrs  = NULL;
2247         hdev->ibh_nmrs = 0;
2248 }
2249
2250 void
2251 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2252 {
2253         kiblnd_hdev_cleanup_mrs(hdev);
2254
2255         if (hdev->ibh_pd != NULL)
2256                 ib_dealloc_pd(hdev->ibh_pd);
2257
2258         if (hdev->ibh_cmid != NULL)
2259                 rdma_destroy_id(hdev->ibh_cmid);
2260
2261         LIBCFS_FREE(hdev, sizeof(*hdev));
2262 }
2263
2264 int
2265 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2266 {
2267         struct ib_mr *mr;
2268         int           i;
2269         int           rc;
2270         __u64         mm_size;
2271         __u64         mr_size;
2272         int           acflags = IB_ACCESS_LOCAL_WRITE |
2273                                 IB_ACCESS_REMOTE_WRITE;
2274
2275         rc = kiblnd_hdev_get_attr(hdev);
2276         if (rc != 0)
2277                 return rc;
2278
2279         if (hdev->ibh_mr_shift == 64) {
2280                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2281                 if (hdev->ibh_mrs == NULL) {
2282                         CERROR("Failed to allocate MRs table\n");
2283                         return -ENOMEM;
2284                 }
2285
2286                 hdev->ibh_mrs[0] = NULL;
2287                 hdev->ibh_nmrs   = 1;
2288
2289                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2290                 if (IS_ERR(mr)) {
2291                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2292                         kiblnd_hdev_cleanup_mrs(hdev);
2293                         return PTR_ERR(mr);
2294                 }
2295
2296                 hdev->ibh_mrs[0] = mr;
2297
2298                 goto out;
2299         }
2300
2301         mr_size = (1ULL << hdev->ibh_mr_shift);
2302         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2303
2304         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2305
2306         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2307                 /* it's 4T..., assume we will re-code at that time */
2308                 CERROR("Can't support memory size: x"LPX64
2309                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2310                 return -EINVAL;
2311         }
2312
2313         /* create an array of MRs to cover all memory */
2314         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2315         if (hdev->ibh_mrs == NULL) {
2316                 CERROR("Failed to allocate MRs' table\n");
2317                 return -ENOMEM;
2318         }
2319
2320         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2321
2322         for (i = 0; i < hdev->ibh_nmrs; i++) {
2323                 struct ib_phys_buf ipb;
2324                 __u64              iova;
2325
2326                 ipb.size = hdev->ibh_mr_size;
2327                 ipb.addr = i * mr_size;
2328                 iova     = ipb.addr;
2329
2330                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2331                 if (IS_ERR(mr)) {
2332                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2333                                " size "LPX64" : %ld\n",
2334                                ipb.addr, ipb.size, PTR_ERR(mr));
2335                         kiblnd_hdev_cleanup_mrs(hdev);
2336                         return PTR_ERR(mr);
2337                 }
2338
2339                 LASSERT (iova == ipb.addr);
2340
2341                 hdev->ibh_mrs[i] = mr;
2342         }
2343
2344 out:
2345         CDEBUG(D_CONSOLE, "Register global MR array, MR size: "
2346                           LPX64", array size: %d\n",
2347                           hdev->ibh_mr_size, hdev->ibh_nmrs);
2348         return 0;
2349 }
2350
2351 static int
2352 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2353 {       /* DUMMY */
2354         return 0;
2355 }
2356
2357 static int
2358 kiblnd_dev_need_failover(kib_dev_t *dev)
2359 {
2360         struct rdma_cm_id  *cmid;
2361         struct sockaddr_in  srcaddr;
2362         struct sockaddr_in  dstaddr;
2363         int                 rc;
2364
2365         if (dev->ibd_hdev == NULL || /* initializing */
2366             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2367             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2368                 return 1;
2369
2370         /* XXX: it's UGLY, but I don't have better way to find
2371          * ib-bonding HCA failover because:
2372          *
2373          * a. no reliable CM event for HCA failover...
2374          * b. no OFED API to get ib_device for current net_device...
2375          *
2376          * We have only two choices at this point:
2377          *
2378          * a. rdma_bind_addr(), it will conflict with listener cmid
2379          * b. rdma_resolve_addr() to zero addr */
2380         cmid = rdma_create_id(kiblnd_dummy_callback,
2381                               dev, RDMA_PS_TCP);
2382         if (IS_ERR(cmid)) {
2383                 rc = PTR_ERR(cmid);
2384                 CERROR("Failed to create cmid for failover: %d\n", rc);
2385                 return rc;
2386         }
2387
2388         memset(&srcaddr, 0, sizeof(srcaddr));
2389         srcaddr.sin_family      = AF_INET;
2390         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2391
2392         memset(&dstaddr, 0, sizeof(dstaddr));
2393         dstaddr.sin_family = AF_INET;
2394         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2395                                (struct sockaddr *)&dstaddr, 1);
2396         if (rc != 0) {
2397                 CERROR("Failed to bind %s to device: %d\n",
2398                        dev->ibd_ifname, rc);
2399                 rdma_destroy_id(cmid);
2400                 return rc;
2401         }
2402
2403         LASSERT (cmid->device != NULL);
2404         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2405                 /* don't need device failover */
2406                 rdma_destroy_id(cmid);
2407                 return 0;
2408         }
2409
2410         return 1;
2411 }
2412
2413 int
2414 kiblnd_dev_failover(kib_dev_t *dev)
2415 {
2416         CFS_LIST_HEAD      (zombie_tpo);
2417         CFS_LIST_HEAD      (zombie_ppo);
2418         CFS_LIST_HEAD      (zombie_fpo);
2419         struct rdma_cm_id  *cmid  = NULL;
2420         kib_hca_dev_t      *hdev  = NULL;
2421         kib_hca_dev_t      *old;
2422         struct ib_pd       *pd;
2423         kib_net_t          *net;
2424         struct sockaddr_in  addr;
2425         unsigned long       flags;
2426         int                 rc = 0;
2427
2428         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2429                  dev->ibd_can_failover ||
2430                  dev->ibd_hdev == NULL);
2431
2432         rc = kiblnd_dev_need_failover(dev);
2433         if (rc <= 0)
2434                 goto out;
2435
2436         if (dev->ibd_hdev != NULL &&
2437             dev->ibd_hdev->ibh_cmid != NULL) {
2438                 /* XXX it's not good to close old listener at here,
2439                  * because we can fail to create new listener.
2440                  * But we have to close it now, otherwise rdma_bind_addr
2441                  * will return EADDRINUSE... How crap! */
2442                 cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2443
2444                 cmid = dev->ibd_hdev->ibh_cmid;
2445                 /* make next schedule of kiblnd_dev_need_failover
2446                  * will return 1 for me */
2447                 dev->ibd_hdev->ibh_cmid  = NULL;
2448                 cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2449
2450                 rdma_destroy_id(cmid);
2451         }
2452
2453         cmid = rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP);
2454         if (IS_ERR(cmid)) {
2455                 rc = PTR_ERR(cmid);
2456                 CERROR("Failed to create cmid for failover: %d\n", rc);
2457                 goto out;
2458         }
2459
2460         memset(&addr, 0, sizeof(addr));
2461         addr.sin_family      = AF_INET;
2462         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2463         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2464
2465         /* Bind to failover device or port */
2466         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2467         if (rc != 0) {
2468                 CERROR("Failed to bind %s to device: %d\n",
2469                        dev->ibd_ifname, rc);
2470                 rdma_destroy_id(cmid);
2471                 goto out;
2472         }
2473
2474         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2475         if (hdev == NULL) {
2476                 CERROR("Failed to allocate kib_hca_dev\n");
2477                 rdma_destroy_id(cmid);
2478                 rc = -ENOMEM;
2479                 goto out;
2480         }
2481         memset(hdev, 0, sizeof(*hdev));
2482         atomic_set(&hdev->ibh_ref, 1);
2483         hdev->ibh_dev   = dev;
2484         hdev->ibh_cmid  = cmid;
2485         hdev->ibh_ibdev = cmid->device;
2486
2487         pd = ib_alloc_pd(cmid->device);
2488         if (IS_ERR(pd)) {
2489                 rc = PTR_ERR(pd);
2490                 CERROR("Can't allocate PD: %d\n", rc);
2491                 goto out;
2492         }
2493
2494         hdev->ibh_pd = pd;
2495
2496         rc = rdma_listen(cmid, 0);
2497         if (rc != 0) {
2498                 CERROR("Can't start new listener: %d\n", rc);
2499                 goto out;
2500         }
2501
2502         rc = kiblnd_hdev_setup_mrs(hdev);
2503         if (rc != 0) {
2504                 CERROR("Can't setup device: %d\n", rc);
2505                 goto out;
2506         }
2507
2508         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2509
2510         old = dev->ibd_hdev;
2511         dev->ibd_hdev = hdev; /* take over the refcount */
2512         hdev = old;
2513
2514         cfs_list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2515                 kiblnd_fail_pool_set(&net->ibn_tx_ps.tps_poolset, &zombie_tpo);
2516                 if (net->ibn_with_pmr)
2517                         kiblnd_fail_pool_set(&net->ibn_pmr_ps.pps_poolset, &zombie_ppo);
2518                 if (net->ibn_with_fmr)
2519                         kiblnd_fail_fmr_pool_set(&net->ibn_fmr_ps, &zombie_fpo);
2520         }
2521
2522         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2523  out:
2524         if (!cfs_list_empty(&zombie_tpo))
2525                 kiblnd_destroy_pool_list(&zombie_tpo);
2526         if (!cfs_list_empty(&zombie_ppo))
2527                 kiblnd_destroy_pool_list(&zombie_ppo);
2528         if (!cfs_list_empty(&zombie_fpo))
2529                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2530         if (hdev != NULL)
2531                 kiblnd_hdev_decref(hdev);
2532
2533         if (rc != 0)
2534                 dev->ibd_failed_failover++;
2535         else
2536                 dev->ibd_failed_failover = 0;
2537
2538         return rc;
2539 }
2540
2541 void
2542 kiblnd_destroy_dev (kib_dev_t *dev)
2543 {
2544         LASSERT (dev->ibd_nnets == 0);
2545         LASSERT (cfs_list_empty(&dev->ibd_nets));
2546
2547         cfs_list_del(&dev->ibd_fail_list);
2548         cfs_list_del(&dev->ibd_list);
2549
2550         if (dev->ibd_hdev != NULL)
2551                 kiblnd_hdev_decref(dev->ibd_hdev);
2552
2553         LIBCFS_FREE(dev, sizeof(*dev));
2554 }
2555
2556 kib_dev_t *
2557 kiblnd_create_dev(char *ifname)
2558 {
2559         struct net_device *netdev;
2560         kib_dev_t         *dev;
2561         __u32              netmask;
2562         __u32              ip;
2563         int                up;
2564         int                rc;
2565
2566         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2567         if (rc != 0) {
2568                 CERROR("Can't query IPoIB interface %s: %d\n",
2569                        ifname, rc);
2570                 return NULL;
2571         }
2572
2573         if (!up) {
2574                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2575                 return NULL;
2576         }
2577
2578         LIBCFS_ALLOC(dev, sizeof(*dev));
2579         if (dev == NULL)
2580                 return NULL;
2581
2582         memset(dev, 0, sizeof(*dev));
2583 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2584         if ((netdev = dev_get_by_name(&init_net, ifname)) == NULL) {
2585 #else
2586         if ((netdev = dev_get_by_name(ifname)) == NULL) {
2587 #endif
2588                 CERROR("Can't find netdev of IF: %s\n", ifname);
2589                 LIBCFS_FREE(dev, sizeof(*dev));
2590                 return NULL;
2591         }
2592
2593         dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2594         dev_put(netdev);
2595
2596         CFS_INIT_LIST_HEAD(&dev->ibd_nets);
2597         CFS_INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2598         CFS_INIT_LIST_HEAD(&dev->ibd_fail_list);
2599         dev->ibd_ifip = ip;
2600         strcpy(&dev->ibd_ifname[0], ifname);
2601
2602         /* initialize the device */
2603         rc = kiblnd_dev_failover(dev);
2604         if (rc != 0) {
2605                 CERROR("Can't initialize device: %d\n", rc);
2606                 LIBCFS_FREE(dev, sizeof(*dev));
2607                 return NULL;
2608         }
2609
2610         cfs_list_add_tail(&dev->ibd_list,
2611                           &kiblnd_data.kib_devs);
2612         return dev;
2613 }
2614
2615 void
2616 kiblnd_base_shutdown (void)
2617 {
2618         int i;
2619
2620         LASSERT (cfs_list_empty(&kiblnd_data.kib_devs));
2621
2622         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2623                cfs_atomic_read(&libcfs_kmemory));
2624
2625         switch (kiblnd_data.kib_init) {
2626         default:
2627                 LBUG();
2628
2629         case IBLND_INIT_ALL:
2630         case IBLND_INIT_DATA:
2631                 LASSERT (kiblnd_data.kib_peers != NULL);
2632                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2633                         LASSERT (cfs_list_empty(&kiblnd_data.kib_peers[i]));
2634                 }
2635                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_zombies));
2636                 LASSERT (cfs_list_empty(&kiblnd_data.kib_connd_conns));
2637
2638                 /* flag threads to terminate; wake and wait for them to die */
2639                 kiblnd_data.kib_shutdown = 1;
2640                 cfs_waitq_broadcast(&kiblnd_data.kib_sched_waitq);
2641                 cfs_waitq_broadcast(&kiblnd_data.kib_connd_waitq);
2642                 cfs_waitq_broadcast(&kiblnd_data.kib_failover_waitq);
2643
2644                 i = 2;
2645                 while (cfs_atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2646                         i++;
2647                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2648                                "Waiting for %d threads to terminate\n",
2649                                cfs_atomic_read(&kiblnd_data.kib_nthreads));
2650                         cfs_pause(cfs_time_seconds(1));
2651                 }
2652
2653                 /* fall through */
2654
2655         case IBLND_INIT_NOTHING:
2656                 break;
2657         }
2658
2659         if (kiblnd_data.kib_peers != NULL)
2660                 LIBCFS_FREE(kiblnd_data.kib_peers,
2661                             sizeof(cfs_list_t) *
2662                             kiblnd_data.kib_peer_hash_size);
2663
2664         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2665                cfs_atomic_read(&libcfs_kmemory));
2666
2667         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2668         PORTAL_MODULE_UNUSE;
2669 }
2670
2671 void
2672 kiblnd_shutdown (lnet_ni_t *ni)
2673 {
2674         kib_net_t        *net = ni->ni_data;
2675         cfs_rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2676         int               i;
2677         unsigned long     flags;
2678
2679         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2680
2681         if (net == NULL)
2682                 goto out;
2683
2684         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2685                cfs_atomic_read(&libcfs_kmemory));
2686
2687         cfs_write_lock_irqsave(g_lock, flags);
2688         net->ibn_shutdown = 1;
2689         cfs_write_unlock_irqrestore(g_lock, flags);
2690
2691         switch (net->ibn_init) {
2692         default:
2693                 LBUG();
2694
2695         case IBLND_INIT_ALL:
2696                 /* nuke all existing peers within this net */
2697                 kiblnd_del_peer(ni, LNET_NID_ANY);
2698
2699                 /* Wait for all peer state to clean up */
2700                 i = 2;
2701                 while (cfs_atomic_read(&net->ibn_npeers) != 0) {
2702                         i++;
2703                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2704                                "%s: waiting for %d peers to disconnect\n",
2705                                libcfs_nid2str(ni->ni_nid),
2706                                cfs_atomic_read(&net->ibn_npeers));
2707                         cfs_pause(cfs_time_seconds(1));
2708                 }
2709
2710                 kiblnd_ni_fini_pools(net);
2711
2712                 cfs_write_lock_irqsave(g_lock, flags);
2713                 LASSERT (net->ibn_dev->ibd_nnets > 0);
2714                 net->ibn_dev->ibd_nnets--;
2715                 cfs_list_del(&net->ibn_list);
2716                 cfs_write_unlock_irqrestore(g_lock, flags);
2717
2718                 /* fall through */
2719
2720         case IBLND_INIT_NOTHING:
2721                 LASSERT (cfs_atomic_read(&net->ibn_nconns) == 0);
2722
2723                 if (net->ibn_dev != NULL &&
2724                     net->ibn_dev->ibd_nnets == 0)
2725                         kiblnd_destroy_dev(net->ibn_dev);
2726
2727                 break;
2728         }
2729
2730         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2731                cfs_atomic_read(&libcfs_kmemory));
2732
2733         net->ibn_init = IBLND_INIT_NOTHING;
2734         ni->ni_data = NULL;
2735
2736         LIBCFS_FREE(net, sizeof(*net));
2737
2738 out:
2739         if (cfs_list_empty(&kiblnd_data.kib_devs))
2740                 kiblnd_base_shutdown();
2741         return;
2742 }
2743
2744 int
2745 kiblnd_base_startup (void)
2746 {
2747         int i;
2748         int rc;
2749
2750         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2751
2752         PORTAL_MODULE_USE;
2753         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2754
2755         cfs_rwlock_init(&kiblnd_data.kib_global_lock);
2756
2757         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2758         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2759
2760         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2761         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2762                      sizeof(cfs_list_t) *
2763                             kiblnd_data.kib_peer_hash_size);
2764         if (kiblnd_data.kib_peers == NULL) {
2765                 goto failed;
2766         }
2767         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2768                 CFS_INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2769
2770         cfs_spin_lock_init(&kiblnd_data.kib_connd_lock);
2771         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2772         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2773         cfs_waitq_init(&kiblnd_data.kib_connd_waitq);
2774
2775         cfs_spin_lock_init(&kiblnd_data.kib_sched_lock);
2776         CFS_INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
2777         cfs_waitq_init(&kiblnd_data.kib_sched_waitq);
2778         cfs_waitq_init(&kiblnd_data.kib_failover_waitq);
2779
2780         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2781
2782         /* lists/ptrs/locks initialised */
2783         kiblnd_data.kib_init = IBLND_INIT_DATA;
2784         /*****************************************************/
2785
2786         for (i = 0; i < IBLND_N_SCHED; i++) {
2787                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
2788                 if (rc != 0) {
2789                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
2790                                i, rc);
2791                         goto failed;
2792                 }
2793         }
2794
2795         rc = kiblnd_thread_start(kiblnd_connd, NULL);
2796         if (rc != 0) {
2797                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2798                 goto failed;
2799         }
2800
2801         if (*kiblnd_tunables.kib_dev_failover != 0)
2802                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL);
2803
2804         if (rc != 0) {
2805                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2806                 goto failed;
2807         }
2808
2809         /* flag everything initialised */
2810         kiblnd_data.kib_init = IBLND_INIT_ALL;
2811         /*****************************************************/
2812
2813         return 0;
2814
2815  failed:
2816         kiblnd_base_shutdown();
2817         return -ENETDOWN;
2818 }
2819
2820 int
2821 kiblnd_startup (lnet_ni_t *ni)
2822 {
2823         char                     *ifname;
2824         kib_dev_t                *ibdev = NULL;
2825         kib_net_t                *net;
2826         cfs_list_t               *tmp;
2827         struct timeval            tv;
2828         unsigned long             flags;
2829         int                       rc;
2830
2831         LASSERT (ni->ni_lnd == &the_o2iblnd);
2832
2833         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2834                 rc = kiblnd_base_startup();
2835                 if (rc != 0)
2836                         return rc;
2837         }
2838
2839         LIBCFS_ALLOC(net, sizeof(*net));
2840         ni->ni_data = net;
2841         if (net == NULL)
2842                 goto failed;
2843
2844         memset(net, 0, sizeof(*net));
2845
2846         cfs_gettimeofday(&tv);
2847         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2848
2849         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2850         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2851         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2852         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2853
2854         if (ni->ni_interfaces[0] != NULL) {
2855                 /* Use the IPoIB interface specified in 'networks=' */
2856
2857                 CLASSERT (LNET_MAX_INTERFACES > 1);
2858                 if (ni->ni_interfaces[1] != NULL) {
2859                         CERROR("Multiple interfaces not supported\n");
2860                         goto failed;
2861                 }
2862
2863                 ifname = ni->ni_interfaces[0];
2864         } else {
2865                 ifname = *kiblnd_tunables.kib_default_ipif;
2866         }
2867
2868         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2869                 CERROR("IPoIB interface name too long: %s\n", ifname);
2870                 goto failed;
2871         }
2872
2873         cfs_list_for_each (tmp, &kiblnd_data.kib_devs) {
2874                 ibdev = cfs_list_entry(tmp, kib_dev_t, ibd_list);
2875
2876                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
2877                         break;
2878
2879                 ibdev = NULL;
2880         }
2881
2882         if (ibdev == NULL)
2883                 ibdev = kiblnd_create_dev(ifname);
2884
2885         if (ibdev == NULL)
2886                 goto failed;
2887
2888         net->ibn_dev = ibdev;
2889         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2890
2891         rc = kiblnd_net_init_pools(net);
2892         if (rc != 0) {
2893                 CERROR("Failed to initialize NI pools: %d\n", rc);
2894                 goto failed;
2895         }
2896
2897         cfs_write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2898         ibdev->ibd_nnets++;
2899         cfs_list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2900         cfs_write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2901
2902         net->ibn_init = IBLND_INIT_ALL;
2903
2904         return 0;
2905
2906 failed:
2907         if (net->ibn_dev == NULL && ibdev != NULL)
2908                 kiblnd_destroy_dev(ibdev);
2909
2910         kiblnd_shutdown(ni);
2911
2912         CDEBUG(D_NET, "kiblnd_startup failed\n");
2913         return -ENETDOWN;
2914 }
2915
2916 void __exit
2917 kiblnd_module_fini (void)
2918 {
2919         lnet_unregister_lnd(&the_o2iblnd);
2920         kiblnd_tunables_fini();
2921 }
2922
2923 int __init
2924 kiblnd_module_init (void)
2925 {
2926         int    rc;
2927
2928         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2929         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2930                   <= IBLND_MSG_SIZE);
2931         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2932                   <= IBLND_MSG_SIZE);
2933
2934         rc = kiblnd_tunables_init();
2935         if (rc != 0)
2936                 return rc;
2937
2938         lnet_register_lnd(&the_o2iblnd);
2939
2940         return 0;
2941 }
2942
2943 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2944 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2945 MODULE_LICENSE("GPL");
2946
2947 module_init(kiblnd_module_init);
2948 module_exit(kiblnd_module_fini);