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