lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 kiblnd_cksum (void *ptr, int nob)
  50 {
  51         char  *c  = ptr;
  52         __u32  sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return (sum == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 340         peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
 341         peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 342         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 343
 344         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 345         INIT_LIST_HEAD(&peer->ibp_conns);
 346         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer (kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(peer->ibp_connecting == 0);
 371         LASSERT(peer->ibp_accepting == 0);
 372         LASSERT(list_empty(&peer->ibp_conns));
 373         LASSERT(list_empty(&peer->ibp_tx_queue));
 374
 375         LIBCFS_FREE(peer, sizeof(*peer));
 376
 377         /* NB a peer's connections keep a reference on their peer until
 378          * they are destroyed, so we can be assured that _all_ state to do
 379          * with this peer has been cleaned up when its refcount drops to
 380          * zero. */
 381         atomic_dec(&net->ibn_npeers);
 382 }
 383
 384 kib_peer_t *
 385 kiblnd_find_peer_locked (lnet_nid_t nid)
 386 {
 387         /* the caller is responsible for accounting the additional reference
 388          * that this creates */
 389         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 390         struct list_head        *tmp;
 391         kib_peer_t              *peer;
 392
 393         list_for_each(tmp, peer_list) {
 394
 395                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 396
 397                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 398                         peer->ibp_accepting > 0 ||
 399                         !list_empty(&peer->ibp_conns)); /* active conn */
 400
 401                 if (peer->ibp_nid != nid)
 402                         continue;
 403
 404                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 405                        peer, libcfs_nid2str(nid),
 406                        atomic_read(&peer->ibp_refcount),
 407                        peer->ibp_version);
 408                 return peer;
 409         }
 410         return NULL;
 411 }
 412
 413 void
 414 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 415 {
 416         LASSERT(list_empty(&peer->ibp_conns));
 417
 418         LASSERT (kiblnd_peer_active(peer));
 419         list_del_init(&peer->ibp_list);
 420         /* lose peerlist's ref */
 421         kiblnd_peer_decref(peer);
 422 }
 423
 424 static int
 425 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 426                      lnet_nid_t *nidp, int *count)
 427 {
 428         kib_peer_t              *peer;
 429         struct list_head        *ptmp;
 430         int                      i;
 431         unsigned long            flags;
 432
 433         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 434
 435         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 436
 437                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 438
 439                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 440                         LASSERT(peer->ibp_connecting > 0 ||
 441                                 peer->ibp_accepting > 0 ||
 442                                 !list_empty(&peer->ibp_conns));
 443
 444                         if (peer->ibp_ni != ni)
 445                                 continue;
 446
 447                         if (index-- > 0)
 448                                 continue;
 449
 450                         *nidp = peer->ibp_nid;
 451                         *count = atomic_read(&peer->ibp_refcount);
 452
 453                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 454                                                flags);
 455                         return 0;
 456                 }
 457         }
 458
 459         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 460         return -ENOENT;
 461 }
 462
 463 static void
 464 kiblnd_del_peer_locked (kib_peer_t *peer)
 465 {
 466         struct list_head        *ctmp;
 467         struct list_head        *cnxt;
 468         kib_conn_t              *conn;
 469
 470         if (list_empty(&peer->ibp_conns)) {
 471                 kiblnd_unlink_peer_locked(peer);
 472         } else {
 473                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 474                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 475
 476                         kiblnd_close_conn_locked(conn, 0);
 477                 }
 478                 /* NB closing peer's last conn unlinked it. */
 479         }
 480         /* NB peer now unlinked; might even be freed if the peer table had the
 481          * last ref on it. */
 482 }
 483
 484 static int
 485 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 486 {
 487         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 488         struct list_head        *ptmp;
 489         struct list_head        *pnxt;
 490         kib_peer_t              *peer;
 491         int                     lo;
 492         int                     hi;
 493         int                     i;
 494         unsigned long           flags;
 495         int                     rc = -ENOENT;
 496
 497         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 498
 499         if (nid != LNET_NID_ANY) {
 500                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 501         } else {
 502                 lo = 0;
 503                 hi = kiblnd_data.kib_peer_hash_size - 1;
 504         }
 505
 506         for (i = lo; i <= hi; i++) {
 507                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 508                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 509                         LASSERT(peer->ibp_connecting > 0 ||
 510                                 peer->ibp_accepting > 0 ||
 511                                 !list_empty(&peer->ibp_conns));
 512
 513                         if (peer->ibp_ni != ni)
 514                                 continue;
 515
 516                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 517                                 continue;
 518
 519                         if (!list_empty(&peer->ibp_tx_queue)) {
 520                                 LASSERT(list_empty(&peer->ibp_conns));
 521
 522                                 list_splice_init(&peer->ibp_tx_queue,
 523                                                  &zombies);
 524                         }
 525
 526                         kiblnd_del_peer_locked(peer);
 527                         rc = 0;         /* matched something */
 528                 }
 529         }
 530
 531         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 532
 533         kiblnd_txlist_done(ni, &zombies, -EIO);
 534
 535         return rc;
 536 }
 537
 538 static kib_conn_t *
 539 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 540 {
 541         kib_peer_t              *peer;
 542         struct list_head        *ptmp;
 543         kib_conn_t              *conn;
 544         struct list_head        *ctmp;
 545         int                     i;
 546         unsigned long           flags;
 547
 548         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 549
 550         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 551                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 552
 553                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 554                         LASSERT(peer->ibp_connecting > 0 ||
 555                                 peer->ibp_accepting > 0 ||
 556                                 !list_empty(&peer->ibp_conns));
 557
 558                         if (peer->ibp_ni != ni)
 559                                 continue;
 560
 561                         list_for_each(ctmp, &peer->ibp_conns) {
 562                                 if (index-- > 0)
 563                                         continue;
 564
 565                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 566                                 kiblnd_conn_addref(conn);
 567                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 568                                                        flags);
 569                                 return conn;
 570                         }
 571                 }
 572         }
 573
 574         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 575         return NULL;
 576 }
 577
 578 static void
 579 kiblnd_debug_rx (kib_rx_t *rx)
 580 {
 581         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 582                rx, rx->rx_status, rx->rx_msg->ibm_type,
 583                rx->rx_msg->ibm_credits);
 584 }
 585
 586 static void
 587 kiblnd_debug_tx (kib_tx_t *tx)
 588 {
 589         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 590                "cookie "LPX64" msg %s%s type %x cred %d\n",
 591                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 592                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 593                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 594                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 595                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 596 }
 597
 598 void
 599 kiblnd_debug_conn (kib_conn_t *conn)
 600 {
 601         struct list_head        *tmp;
 602         int                     i;
 603
 604         spin_lock(&conn->ibc_lock);
 605
 606         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 607                atomic_read(&conn->ibc_refcount), conn,
 608                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 609         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 610                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 611                conn->ibc_nsends_posted, conn->ibc_credits,
 612                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 613         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 614
 615         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 616         list_for_each(tmp, &conn->ibc_early_rxs)
 617                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 620         list_for_each(tmp, &conn->ibc_tx_noops)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 628         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 632         list_for_each(tmp, &conn->ibc_tx_queue)
 633                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 634
 635         CDEBUG(D_CONSOLE, "   active_txs:\n");
 636         list_for_each(tmp, &conn->ibc_active_txs)
 637                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 638
 639         CDEBUG(D_CONSOLE, "   rxs:\n");
 640         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 641                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 642
 643         spin_unlock(&conn->ibc_lock);
 644 }
 645
 646 int
 647 kiblnd_translate_mtu(int value)
 648 {
 649         switch (value) {
 650         default:
 651                 return -1;
 652         case 0:
 653                 return 0;
 654         case 256:
 655                 return IB_MTU_256;
 656         case 512:
 657                 return IB_MTU_512;
 658         case 1024:
 659                 return IB_MTU_1024;
 660         case 2048:
 661                 return IB_MTU_2048;
 662         case 4096:
 663                 return IB_MTU_4096;
 664         }
 665 }
 666
 667 static void
 668 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 669 {
 670         int           mtu;
 671
 672         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 673         if (cmid->route.path_rec == NULL)
 674                 return;
 675
 676         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 677         LASSERT (mtu >= 0);
 678         if (mtu != 0)
 679                 cmid->route.path_rec->mtu = mtu;
 680 }
 681
 682 static int
 683 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 684 {
 685         cpumask_t       *mask;
 686         int             vectors;
 687         int             off;
 688         int             i;
 689
 690         vectors = conn->ibc_cmid->device->num_comp_vectors;
 691         if (vectors <= 1)
 692                 return 0;
 693
 694         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 695
 696         /* hash NID to CPU id in this partition... */
 697         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 698         for_each_cpu(i, mask) {
 699                 if (off-- == 0)
 700                         return i % vectors;
 701         }
 702
 703         LBUG();
 704         return 1;
 705 }
 706
 707 kib_conn_t *
 708 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 709                    int state, int version)
 710 {
 711         /* CAVEAT EMPTOR:
 712          * If the new conn is created successfully it takes over the caller's
 713          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 714          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 715          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 716          * to destroy 'cmid' here since I'm called from the CM which still has
 717          * its ref on 'cmid'). */
 718         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 719         kib_net_t              *net = peer->ibp_ni->ni_data;
 720         kib_dev_t              *dev;
 721         struct ib_qp_init_attr *init_qp_attr;
 722         struct kib_sched_info   *sched;
 723 #ifdef HAVE_IB_CQ_INIT_ATTR
 724         struct ib_cq_init_attr  cq_attr = {};
 725 #endif
 726         kib_conn_t              *conn;
 727         struct ib_cq            *cq;
 728         unsigned long           flags;
 729         int                     cpt;
 730         int                     rc;
 731         int                     i;
 732
 733         LASSERT(net != NULL);
 734         LASSERT(!in_interrupt());
 735
 736         dev = net->ibn_dev;
 737
 738         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 739         sched = kiblnd_data.kib_scheds[cpt];
 740
 741         LASSERT(sched->ibs_nthreads > 0);
 742
 743         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 744                          sizeof(*init_qp_attr));
 745         if (init_qp_attr == NULL) {
 746                 CERROR("Can't allocate qp_attr for %s\n",
 747                        libcfs_nid2str(peer->ibp_nid));
 748                 goto failed_0;
 749         }
 750
 751         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 752         if (conn == NULL) {
 753                 CERROR("Can't allocate connection for %s\n",
 754                        libcfs_nid2str(peer->ibp_nid));
 755                 goto failed_1;
 756         }
 757
 758         conn->ibc_state = IBLND_CONN_INIT;
 759         conn->ibc_version = version;
 760         conn->ibc_peer = peer;                  /* I take the caller's ref */
 761         cmid->context = conn;                   /* for future CM callbacks */
 762         conn->ibc_cmid = cmid;
 763         conn->ibc_max_frags = peer->ibp_max_frags;
 764         conn->ibc_queue_depth = peer->ibp_queue_depth;
 765
 766         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 767         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 768         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 769         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 770         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 771         INIT_LIST_HEAD(&conn->ibc_active_txs);
 772         spin_lock_init(&conn->ibc_lock);
 773
 774         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 775                          sizeof(*conn->ibc_connvars));
 776         if (conn->ibc_connvars == NULL) {
 777                 CERROR("Can't allocate in-progress connection state\n");
 778                 goto failed_2;
 779         }
 780
 781         write_lock_irqsave(glock, flags);
 782         if (dev->ibd_failover) {
 783                 write_unlock_irqrestore(glock, flags);
 784                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 785                 goto failed_2;
 786         }
 787
 788         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 789                 /* wakeup failover thread and teardown connection */
 790                 if (kiblnd_dev_can_failover(dev)) {
 791                         list_add_tail(&dev->ibd_fail_list,
 792                                       &kiblnd_data.kib_failed_devs);
 793                         wake_up(&kiblnd_data.kib_failover_waitq);
 794                 }
 795
 796                 write_unlock_irqrestore(glock, flags);
 797                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 798                        cmid->device->name, dev->ibd_ifname);
 799                 goto failed_2;
 800         }
 801
 802         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 803         conn->ibc_hdev = dev->ibd_hdev;
 804
 805         kiblnd_setup_mtu_locked(cmid);
 806
 807         write_unlock_irqrestore(glock, flags);
 808
 809         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 810                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 811         if (conn->ibc_rxs == NULL) {
 812                 CERROR("Cannot allocate RX buffers\n");
 813                 goto failed_2;
 814         }
 815
 816         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 817                                 IBLND_RX_MSG_PAGES(conn));
 818         if (rc != 0)
 819                 goto failed_2;
 820
 821         kiblnd_map_rx_descs(conn);
 822
 823 #ifdef HAVE_IB_CQ_INIT_ATTR
 824         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 825         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 826         cq = ib_create_cq(cmid->device,
 827                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 828                           &cq_attr);
 829 #else
 830         cq = ib_create_cq(cmid->device,
 831                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 832                           IBLND_CQ_ENTRIES(conn),
 833                           kiblnd_get_completion_vector(conn, cpt));
 834 #endif
 835         if (IS_ERR(cq)) {
 836                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 837                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 838                 goto failed_2;
 839         }
 840
 841         conn->ibc_cq = cq;
 842
 843         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 844         if (rc != 0) {
 845                 CERROR("Can't request completion notification: %d\n", rc);
 846                 goto failed_2;
 847         }
 848
 849         init_qp_attr->event_handler = kiblnd_qp_event;
 850         init_qp_attr->qp_context = conn;
 851         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 852         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 853         init_qp_attr->cap.max_send_sge = 1;
 854         init_qp_attr->cap.max_recv_sge = 1;
 855         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 856         init_qp_attr->qp_type = IB_QPT_RC;
 857         init_qp_attr->send_cq = cq;
 858         init_qp_attr->recv_cq = cq;
 859
 860         conn->ibc_sched = sched;
 861
 862         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 863         if (rc != 0) {
 864                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 865                        rc, init_qp_attr->cap.max_send_wr,
 866                        init_qp_attr->cap.max_recv_wr);
 867                 goto failed_2;
 868         }
 869
 870         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 871
 872         /* 1 ref for caller and each rxmsg */
 873         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 874         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 875
 876         /* post receives */
 877         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 878                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 879                                     IBLND_POSTRX_NO_CREDIT);
 880                 if (rc != 0) {
 881                         CERROR("Can't post rxmsg: %d\n", rc);
 882
 883                         /* Make posted receives complete */
 884                         kiblnd_abort_receives(conn);
 885
 886                         /* correct # of posted buffers
 887                          * NB locking needed now I'm racing with completion */
 888                         spin_lock_irqsave(&sched->ibs_lock, flags);
 889                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 890                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 891
 892                         /* cmid will be destroyed by CM(ofed) after cm_callback
 893                          * returned, so we can't refer it anymore
 894                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 895                         rdma_destroy_qp(conn->ibc_cmid);
 896                         conn->ibc_cmid = NULL;
 897
 898                         /* Drop my own and unused rxbuffer refcounts */
 899                         while (i++ <= IBLND_RX_MSGS(conn))
 900                                 kiblnd_conn_decref(conn);
 901
 902                         return NULL;
 903                 }
 904         }
 905
 906         /* Init successful! */
 907         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 908                  state == IBLND_CONN_PASSIVE_WAIT);
 909         conn->ibc_state = state;
 910
 911         /* 1 more conn */
 912         atomic_inc(&net->ibn_nconns);
 913         return conn;
 914
 915  failed_2:
 916         kiblnd_destroy_conn(conn);
 917  failed_1:
 918         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 919  failed_0:
 920         return NULL;
 921 }
 922
 923 void
 924 kiblnd_destroy_conn (kib_conn_t *conn)
 925 {
 926         struct rdma_cm_id *cmid = conn->ibc_cmid;
 927         kib_peer_t        *peer = conn->ibc_peer;
 928         int                rc;
 929
 930         LASSERT (!in_interrupt());
 931         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 932         LASSERT(list_empty(&conn->ibc_early_rxs));
 933         LASSERT(list_empty(&conn->ibc_tx_noops));
 934         LASSERT(list_empty(&conn->ibc_tx_queue));
 935         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 936         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 937         LASSERT(list_empty(&conn->ibc_active_txs));
 938         LASSERT (conn->ibc_noops_posted == 0);
 939         LASSERT (conn->ibc_nsends_posted == 0);
 940
 941         switch (conn->ibc_state) {
 942         default:
 943                 /* conn must be completely disengaged from the network */
 944                 LBUG();
 945
 946         case IBLND_CONN_DISCONNECTED:
 947                 /* connvars should have been freed already */
 948                 LASSERT (conn->ibc_connvars == NULL);
 949                 break;
 950
 951         case IBLND_CONN_INIT:
 952                 break;
 953         }
 954
 955         /* conn->ibc_cmid might be destroyed by CM already */
 956         if (cmid != NULL && cmid->qp != NULL)
 957                 rdma_destroy_qp(cmid);
 958
 959         if (conn->ibc_cq != NULL) {
 960                 rc = ib_destroy_cq(conn->ibc_cq);
 961                 if (rc != 0)
 962                         CWARN("Error destroying CQ: %d\n", rc);
 963         }
 964
 965         if (conn->ibc_rx_pages != NULL)
 966                 kiblnd_unmap_rx_descs(conn);
 967
 968         if (conn->ibc_rxs != NULL) {
 969                 LIBCFS_FREE(conn->ibc_rxs,
 970                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 971         }
 972
 973         if (conn->ibc_connvars != NULL)
 974                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 975
 976         if (conn->ibc_hdev != NULL)
 977                 kiblnd_hdev_decref(conn->ibc_hdev);
 978
 979         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 980         if (conn->ibc_state != IBLND_CONN_INIT) {
 981                 kib_net_t *net = peer->ibp_ni->ni_data;
 982
 983                 kiblnd_peer_decref(peer);
 984                 rdma_destroy_id(cmid);
 985                 atomic_dec(&net->ibn_nconns);
 986         }
 987
 988         LIBCFS_FREE(conn, sizeof(*conn));
 989 }
 990
 991 int
 992 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 993 {
 994         kib_conn_t              *conn;
 995         struct list_head        *ctmp;
 996         struct list_head        *cnxt;
 997         int                     count = 0;
 998
 999         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1000                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1001
1002                 CDEBUG(D_NET, "Closing conn -> %s, "
1003                               "version: %x, reason: %d\n",
1004                        libcfs_nid2str(peer->ibp_nid),
1005                        conn->ibc_version, why);
1006
1007                 kiblnd_close_conn_locked(conn, why);
1008                 count++;
1009         }
1010
1011         return count;
1012 }
1013
1014 int
1015 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1016                                 int version, __u64 incarnation)
1017 {
1018         kib_conn_t              *conn;
1019         struct list_head        *ctmp;
1020         struct list_head        *cnxt;
1021         int                     count = 0;
1022
1023         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1024                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1025
1026                 if (conn->ibc_version     == version &&
1027                     conn->ibc_incarnation == incarnation)
1028                         continue;
1029
1030                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1031                               "incarnation:"LPX64"(%x, "LPX64")\n",
1032                        libcfs_nid2str(peer->ibp_nid),
1033                        conn->ibc_version, conn->ibc_incarnation,
1034                        version, incarnation);
1035
1036                 kiblnd_close_conn_locked(conn, -ESTALE);
1037                 count++;
1038         }
1039
1040         return count;
1041 }
1042
1043 static int
1044 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1045 {
1046         kib_peer_t              *peer;
1047         struct list_head        *ptmp;
1048         struct list_head        *pnxt;
1049         int                     lo;
1050         int                     hi;
1051         int                     i;
1052         unsigned long           flags;
1053         int                     count = 0;
1054
1055         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1056
1057         if (nid != LNET_NID_ANY)
1058                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1059         else {
1060                 lo = 0;
1061                 hi = kiblnd_data.kib_peer_hash_size - 1;
1062         }
1063
1064         for (i = lo; i <= hi; i++) {
1065                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1066
1067                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1068                         LASSERT(peer->ibp_connecting > 0 ||
1069                                 peer->ibp_accepting > 0 ||
1070                                 !list_empty(&peer->ibp_conns));
1071
1072                         if (peer->ibp_ni != ni)
1073                                 continue;
1074
1075                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1076                                 continue;
1077
1078                         count += kiblnd_close_peer_conns_locked(peer, 0);
1079                 }
1080         }
1081
1082         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1083
1084         /* wildcards always succeed */
1085         if (nid == LNET_NID_ANY)
1086                 return 0;
1087
1088         return (count == 0) ? -ENOENT : 0;
1089 }
1090
1091 static int
1092 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1093 {
1094         struct libcfs_ioctl_data *data = arg;
1095         int                       rc = -EINVAL;
1096
1097         switch(cmd) {
1098         case IOC_LIBCFS_GET_PEER: {
1099                 lnet_nid_t   nid = 0;
1100                 int          count = 0;
1101
1102                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1103                                           &nid, &count);
1104                 data->ioc_nid    = nid;
1105                 data->ioc_count  = count;
1106                 break;
1107         }
1108
1109         case IOC_LIBCFS_DEL_PEER: {
1110                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1111                 break;
1112         }
1113         case IOC_LIBCFS_GET_CONN: {
1114                 kib_conn_t *conn;
1115
1116                 rc = 0;
1117                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1118                 if (conn == NULL) {
1119                         rc = -ENOENT;
1120                         break;
1121                 }
1122
1123                 LASSERT (conn->ibc_cmid != NULL);
1124                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1125                 if (conn->ibc_cmid->route.path_rec == NULL)
1126                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1127                 else
1128                         data->ioc_u32[0] =
1129                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1130                 kiblnd_conn_decref(conn);
1131                 break;
1132         }
1133         case IOC_LIBCFS_CLOSE_CONNECTION: {
1134                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1135                 break;
1136         }
1137
1138         default:
1139                 break;
1140         }
1141
1142         return rc;
1143 }
1144
1145 static void
1146 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1147 {
1148         cfs_time_t      last_alive = 0;
1149         cfs_time_t      now = cfs_time_current();
1150         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1151         kib_peer_t      *peer;
1152         unsigned long   flags;
1153
1154         read_lock_irqsave(glock, flags);
1155
1156         peer = kiblnd_find_peer_locked(nid);
1157         if (peer != NULL) {
1158                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1159                         peer->ibp_accepting > 0 ||
1160                         !list_empty(&peer->ibp_conns));  /* active conn */
1161                 last_alive = peer->ibp_last_alive;
1162         }
1163
1164         read_unlock_irqrestore(glock, flags);
1165
1166         if (last_alive != 0)
1167                 *when = last_alive;
1168
1169         /* peer is not persistent in hash, trigger peer creation
1170          * and connection establishment with a NULL tx */
1171         if (peer == NULL)
1172                 kiblnd_launch_tx(ni, NULL, nid);
1173
1174         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1175                libcfs_nid2str(nid), peer,
1176                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1177         return;
1178 }
1179
1180 static void
1181 kiblnd_free_pages(kib_pages_t *p)
1182 {
1183         int     npages = p->ibp_npages;
1184         int     i;
1185
1186         for (i = 0; i < npages; i++) {
1187                 if (p->ibp_pages[i] != NULL)
1188                         __free_page(p->ibp_pages[i]);
1189         }
1190
1191         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1192 }
1193
1194 int
1195 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1196 {
1197         kib_pages_t     *p;
1198         int             i;
1199
1200         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1201                          offsetof(kib_pages_t, ibp_pages[npages]));
1202         if (p == NULL) {
1203                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1204                 return -ENOMEM;
1205         }
1206
1207         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1208         p->ibp_npages = npages;
1209
1210         for (i = 0; i < npages; i++) {
1211                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1212                                                      GFP_NOFS);
1213                 if (p->ibp_pages[i] == NULL) {
1214                         CERROR("Can't allocate page %d of %d\n", i, npages);
1215                         kiblnd_free_pages(p);
1216                         return -ENOMEM;
1217                 }
1218         }
1219
1220         *pp = p;
1221         return 0;
1222 }
1223
1224 void
1225 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1226 {
1227         kib_rx_t *rx;
1228         int       i;
1229
1230         LASSERT (conn->ibc_rxs != NULL);
1231         LASSERT (conn->ibc_hdev != NULL);
1232
1233         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1234                 rx = &conn->ibc_rxs[i];
1235
1236                 LASSERT(rx->rx_nob >= 0); /* not posted */
1237
1238                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1239                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1240                                                           rx->rx_msgaddr),
1241                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1242         }
1243
1244         kiblnd_free_pages(conn->ibc_rx_pages);
1245
1246         conn->ibc_rx_pages = NULL;
1247 }
1248
1249 void
1250 kiblnd_map_rx_descs(kib_conn_t *conn)
1251 {
1252         kib_rx_t       *rx;
1253         struct page    *pg;
1254         int             pg_off;
1255         int             ipg;
1256         int             i;
1257
1258         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1259                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1260                 rx = &conn->ibc_rxs[i];
1261
1262                 rx->rx_conn = conn;
1263                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1264
1265                 rx->rx_msgaddr =
1266                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1267                                               rx->rx_msg, IBLND_MSG_SIZE,
1268                                               DMA_FROM_DEVICE);
1269                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1270                                                   rx->rx_msgaddr));
1271                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1272
1273                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1274                        i, rx->rx_msg, rx->rx_msgaddr,
1275                        (__u64)(page_to_phys(pg) + pg_off));
1276
1277                 pg_off += IBLND_MSG_SIZE;
1278                 LASSERT(pg_off <= PAGE_SIZE);
1279
1280                 if (pg_off == PAGE_SIZE) {
1281                         pg_off = 0;
1282                         ipg++;
1283                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1284                 }
1285         }
1286 }
1287
1288 static void
1289 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1290 {
1291         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1292         kib_tx_t       *tx;
1293         int             i;
1294
1295         LASSERT (tpo->tpo_pool.po_allocated == 0);
1296
1297         if (hdev == NULL)
1298                 return;
1299
1300         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1301                 tx = &tpo->tpo_tx_descs[i];
1302                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1303                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1304                                                           tx->tx_msgaddr),
1305                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1306         }
1307
1308         kiblnd_hdev_decref(hdev);
1309         tpo->tpo_hdev = NULL;
1310 }
1311
1312 static kib_hca_dev_t *
1313 kiblnd_current_hdev(kib_dev_t *dev)
1314 {
1315         kib_hca_dev_t *hdev;
1316         unsigned long  flags;
1317         int            i = 0;
1318
1319         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1320         while (dev->ibd_failover) {
1321                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1322                 if (i++ % 50 == 0)
1323                         CDEBUG(D_NET, "%s: Wait for failover\n",
1324                                dev->ibd_ifname);
1325                 schedule_timeout(cfs_time_seconds(1) / 100);
1326
1327                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1328         }
1329
1330         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331         hdev = dev->ibd_hdev;
1332
1333         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1334
1335         return hdev;
1336 }
1337
1338 static void
1339 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1340 {
1341         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1342         kib_pool_t     *pool  = &tpo->tpo_pool;
1343         kib_net_t      *net   = pool->po_owner->ps_net;
1344         kib_dev_t      *dev;
1345         struct page    *page;
1346         kib_tx_t       *tx;
1347         int             page_offset;
1348         int             ipage;
1349         int             i;
1350
1351         LASSERT (net != NULL);
1352
1353         dev = net->ibn_dev;
1354
1355         /* pre-mapped messages are not bigger than 1 page */
1356         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1357
1358         /* No fancy arithmetic when we do the buffer calculations */
1359         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1360
1361         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1362
1363         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1364                 page = txpgs->ibp_pages[ipage];
1365                 tx = &tpo->tpo_tx_descs[i];
1366
1367                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1368                                            page_offset);
1369
1370                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1371                                                        tx->tx_msg,
1372                                                        IBLND_MSG_SIZE,
1373                                                        DMA_TO_DEVICE);
1374                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1375                                                   tx->tx_msgaddr));
1376                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1377
1378                 list_add(&tx->tx_list, &pool->po_free_list);
1379
1380                 page_offset += IBLND_MSG_SIZE;
1381                 LASSERT(page_offset <= PAGE_SIZE);
1382
1383                 if (page_offset == PAGE_SIZE) {
1384                         page_offset = 0;
1385                         ipage++;
1386                         LASSERT(ipage <= txpgs->ibp_npages);
1387                 }
1388         }
1389 }
1390
1391 struct ib_mr *
1392 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1393                       int negotiated_nfrags)
1394 {
1395         __u16   nfrags = (negotiated_nfrags != -1) ?
1396           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1397
1398         LASSERT(hdev->ibh_mrs != NULL);
1399
1400         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1401             nfrags <= rd->rd_nfrags)
1402                 return NULL;
1403
1404         return hdev->ibh_mrs;
1405 }
1406
1407 static void
1408 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1409 {
1410         LASSERT (pool->fpo_map_count == 0);
1411
1412         if (pool->fpo_fmr_pool != NULL)
1413                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1414
1415         if (pool->fpo_hdev != NULL)
1416                 kiblnd_hdev_decref(pool->fpo_hdev);
1417
1418         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1419 }
1420
1421 static void
1422 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1423 {
1424         kib_fmr_pool_t *pool;
1425
1426         while (!list_empty(head)) {
1427                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1428                 list_del(&pool->fpo_list);
1429                 kiblnd_destroy_fmr_pool(pool);
1430         }
1431 }
1432
1433 static int kiblnd_fmr_pool_size(int ncpts)
1434 {
1435         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1436
1437         return max(IBLND_FMR_POOL, size);
1438 }
1439
1440 static int kiblnd_fmr_flush_trigger(int ncpts)
1441 {
1442         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1443
1444         return max(IBLND_FMR_POOL_FLUSH, size);
1445 }
1446
1447 static int
1448 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1449 {
1450         /* FMR pool for RDMA */
1451         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1452         kib_fmr_pool_t          *fpo;
1453         struct ib_fmr_pool_param param = {
1454                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1455                 .page_shift        = PAGE_SHIFT,
1456                 .access            = (IB_ACCESS_LOCAL_WRITE |
1457                                       IB_ACCESS_REMOTE_WRITE),
1458                 .pool_size         = fps->fps_pool_size,
1459                 .dirty_watermark   = fps->fps_flush_trigger,
1460                 .flush_function    = NULL,
1461                 .flush_arg         = NULL,
1462                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1463         int rc;
1464
1465         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1466         if (fpo == NULL)
1467                 return -ENOMEM;
1468
1469         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1470
1471         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1472         if (IS_ERR(fpo->fpo_fmr_pool)) {
1473                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1474                 CERROR("Failed to create FMR pool: %d\n", rc);
1475
1476                 kiblnd_hdev_decref(fpo->fpo_hdev);
1477                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1478                 return rc;
1479         }
1480
1481         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1482         fpo->fpo_owner    = fps;
1483         *pp_fpo = fpo;
1484
1485         return 0;
1486 }
1487
1488 static void
1489 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1490 {
1491         if (fps->fps_net == NULL) /* intialized? */
1492                 return;
1493
1494         spin_lock(&fps->fps_lock);
1495
1496         while (!list_empty(&fps->fps_pool_list)) {
1497                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1498                                                  kib_fmr_pool_t, fpo_list);
1499                 fpo->fpo_failed = 1;
1500                 list_del(&fpo->fpo_list);
1501                 if (fpo->fpo_map_count == 0)
1502                         list_add(&fpo->fpo_list, zombies);
1503                 else
1504                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1505         }
1506
1507         spin_unlock(&fps->fps_lock);
1508 }
1509
1510 static void
1511 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1512 {
1513         if (fps->fps_net != NULL) { /* initialized? */
1514                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1515                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1516         }
1517 }
1518
1519 static int
1520 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1521                         int pool_size, int flush_trigger)
1522 {
1523         kib_fmr_pool_t *fpo;
1524         int             rc;
1525
1526         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1527
1528         fps->fps_net = net;
1529         fps->fps_cpt = cpt;
1530         fps->fps_pool_size = pool_size;
1531         fps->fps_flush_trigger = flush_trigger;
1532         spin_lock_init(&fps->fps_lock);
1533         INIT_LIST_HEAD(&fps->fps_pool_list);
1534         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1535
1536         rc = kiblnd_create_fmr_pool(fps, &fpo);
1537         if (rc == 0)
1538                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1539
1540         return rc;
1541 }
1542
1543 static int
1544 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1545 {
1546         if (fpo->fpo_map_count != 0) /* still in use */
1547                 return 0;
1548         if (fpo->fpo_failed)
1549                 return 1;
1550         return cfs_time_aftereq(now, fpo->fpo_deadline);
1551 }
1552
1553 void
1554 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1555 {
1556         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1557         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1558         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1559         cfs_time_t         now = cfs_time_current();
1560         kib_fmr_pool_t    *tmp;
1561         int                rc;
1562
1563         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1564         LASSERT(rc == 0);
1565
1566         if (status != 0) {
1567                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1568                 LASSERT(rc == 0);
1569         }
1570
1571         fmr->fmr_pool = NULL;
1572         fmr->fmr_pfmr = NULL;
1573
1574         spin_lock(&fps->fps_lock);
1575         fpo->fpo_map_count--;   /* decref the pool */
1576
1577         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1578                 /* the first pool is persistent */
1579                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1580                         continue;
1581
1582                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1583                         list_move(&fpo->fpo_list, &zombies);
1584                         fps->fps_version++;
1585                 }
1586         }
1587         spin_unlock(&fps->fps_lock);
1588
1589         if (!list_empty(&zombies))
1590                 kiblnd_destroy_fmr_pool_list(&zombies);
1591 }
1592
1593 int
1594 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1595                     __u64 iov, kib_fmr_t *fmr)
1596 {
1597         struct ib_pool_fmr *pfmr;
1598         kib_fmr_pool_t     *fpo;
1599         __u64               version;
1600         int                 rc;
1601
1602 again:
1603         spin_lock(&fps->fps_lock);
1604         version = fps->fps_version;
1605         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1606                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1607                 fpo->fpo_map_count++;
1608                 spin_unlock(&fps->fps_lock);
1609
1610                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1611                                             pages, npages, iov);
1612                 if (likely(!IS_ERR(pfmr))) {
1613                         fmr->fmr_pool = fpo;
1614                         fmr->fmr_pfmr = pfmr;
1615                         return 0;
1616                 }
1617
1618                 spin_lock(&fps->fps_lock);
1619                 fpo->fpo_map_count--;
1620                 if (PTR_ERR(pfmr) != -EAGAIN) {
1621                         spin_unlock(&fps->fps_lock);
1622                         return PTR_ERR(pfmr);
1623                 }
1624
1625                 /* EAGAIN and ... */
1626                 if (version != fps->fps_version) {
1627                         spin_unlock(&fps->fps_lock);
1628                         goto again;
1629                 }
1630         }
1631
1632         if (fps->fps_increasing) {
1633                 spin_unlock(&fps->fps_lock);
1634                 CDEBUG(D_NET, "Another thread is allocating new "
1635                        "FMR pool, waiting for her to complete\n");
1636                 schedule();
1637                 goto again;
1638
1639         }
1640
1641         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1642                 /* someone failed recently */
1643                 spin_unlock(&fps->fps_lock);
1644                 return -EAGAIN;
1645         }
1646
1647         fps->fps_increasing = 1;
1648         spin_unlock(&fps->fps_lock);
1649
1650         CDEBUG(D_NET, "Allocate new FMR pool\n");
1651         rc = kiblnd_create_fmr_pool(fps, &fpo);
1652         spin_lock(&fps->fps_lock);
1653         fps->fps_increasing = 0;
1654         if (rc == 0) {
1655                 fps->fps_version++;
1656                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1657         } else {
1658                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1659         }
1660         spin_unlock(&fps->fps_lock);
1661
1662         goto again;
1663 }
1664
1665 static void
1666 kiblnd_fini_pool(kib_pool_t *pool)
1667 {
1668         LASSERT(list_empty(&pool->po_free_list));
1669         LASSERT(pool->po_allocated == 0);
1670
1671         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1672 }
1673
1674 static void
1675 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1676 {
1677         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1678
1679         memset(pool, 0, sizeof(kib_pool_t));
1680         INIT_LIST_HEAD(&pool->po_free_list);
1681         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1682         pool->po_owner    = ps;
1683         pool->po_size     = size;
1684 }
1685
1686 static void
1687 kiblnd_destroy_pool_list(struct list_head *head)
1688 {
1689         kib_pool_t *pool;
1690
1691         while (!list_empty(head)) {
1692                 pool = list_entry(head->next, kib_pool_t, po_list);
1693                 list_del(&pool->po_list);
1694
1695                 LASSERT(pool->po_owner != NULL);
1696                 pool->po_owner->ps_pool_destroy(pool);
1697         }
1698 }
1699
1700 static void
1701 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1702 {
1703         if (ps->ps_net == NULL) /* intialized? */
1704                 return;
1705
1706         spin_lock(&ps->ps_lock);
1707         while (!list_empty(&ps->ps_pool_list)) {
1708                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1709                                             kib_pool_t, po_list);
1710                 po->po_failed = 1;
1711                 list_del(&po->po_list);
1712                 if (po->po_allocated == 0)
1713                         list_add(&po->po_list, zombies);
1714                 else
1715                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1716         }
1717         spin_unlock(&ps->ps_lock);
1718 }
1719
1720 static void
1721 kiblnd_fini_poolset(kib_poolset_t *ps)
1722 {
1723         if (ps->ps_net != NULL) { /* initialized? */
1724                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1725                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1726         }
1727 }
1728
1729 static int
1730 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1731                     kib_net_t *net, char *name, int size,
1732                     kib_ps_pool_create_t po_create,
1733                     kib_ps_pool_destroy_t po_destroy,
1734                     kib_ps_node_init_t nd_init,
1735                     kib_ps_node_fini_t nd_fini)
1736 {
1737         kib_pool_t      *pool;
1738         int             rc;
1739
1740         memset(ps, 0, sizeof(kib_poolset_t));
1741
1742         ps->ps_cpt          = cpt;
1743         ps->ps_net          = net;
1744         ps->ps_pool_create  = po_create;
1745         ps->ps_pool_destroy = po_destroy;
1746         ps->ps_node_init    = nd_init;
1747         ps->ps_node_fini    = nd_fini;
1748         ps->ps_pool_size    = size;
1749         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1750             >= sizeof(ps->ps_name))
1751                 return -E2BIG;
1752         spin_lock_init(&ps->ps_lock);
1753         INIT_LIST_HEAD(&ps->ps_pool_list);
1754         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1755
1756         rc = ps->ps_pool_create(ps, size, &pool);
1757         if (rc == 0)
1758                 list_add(&pool->po_list, &ps->ps_pool_list);
1759         else
1760                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1761
1762         return rc;
1763 }
1764
1765 static int
1766 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1767 {
1768         if (pool->po_allocated != 0) /* still in use */
1769                 return 0;
1770         if (pool->po_failed)
1771                 return 1;
1772         return cfs_time_aftereq(now, pool->po_deadline);
1773 }
1774
1775 void
1776 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1777 {
1778         struct list_head zombies = LIST_HEAD_INIT(zombies);
1779         kib_poolset_t   *ps = pool->po_owner;
1780         kib_pool_t      *tmp;
1781         cfs_time_t       now = cfs_time_current();
1782
1783         spin_lock(&ps->ps_lock);
1784
1785         if (ps->ps_node_fini != NULL)
1786                 ps->ps_node_fini(pool, node);
1787
1788         LASSERT(pool->po_allocated > 0);
1789         list_add(node, &pool->po_free_list);
1790         pool->po_allocated--;
1791
1792         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1793                 /* the first pool is persistent */
1794                 if (ps->ps_pool_list.next == &pool->po_list)
1795                         continue;
1796
1797                 if (kiblnd_pool_is_idle(pool, now))
1798                         list_move(&pool->po_list, &zombies);
1799         }
1800         spin_unlock(&ps->ps_lock);
1801
1802         if (!list_empty(&zombies))
1803                 kiblnd_destroy_pool_list(&zombies);
1804 }
1805
1806 struct list_head *
1807 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1808 {
1809         struct list_head        *node;
1810         kib_pool_t              *pool;
1811         int                     rc;
1812
1813 again:
1814         spin_lock(&ps->ps_lock);
1815         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1816                 if (list_empty(&pool->po_free_list))
1817                         continue;
1818
1819                 pool->po_allocated++;
1820                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1821                 node = pool->po_free_list.next;
1822                 list_del(node);
1823
1824                 if (ps->ps_node_init != NULL) {
1825                         /* still hold the lock */
1826                         ps->ps_node_init(pool, node);
1827                 }
1828                 spin_unlock(&ps->ps_lock);
1829                 return node;
1830         }
1831
1832         /* no available tx pool and ... */
1833         if (ps->ps_increasing) {
1834                 /* another thread is allocating a new pool */
1835                 spin_unlock(&ps->ps_lock);
1836                 CDEBUG(D_NET, "Another thread is allocating new "
1837                        "%s pool, waiting for her to complete\n",
1838                        ps->ps_name);
1839                 schedule();
1840                 goto again;
1841         }
1842
1843         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1844                 /* someone failed recently */
1845                 spin_unlock(&ps->ps_lock);
1846                 return NULL;
1847         }
1848
1849         ps->ps_increasing = 1;
1850         spin_unlock(&ps->ps_lock);
1851
1852         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1853
1854         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1855
1856         spin_lock(&ps->ps_lock);
1857         ps->ps_increasing = 0;
1858         if (rc == 0) {
1859                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1860         } else {
1861                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1862                 CERROR("Can't allocate new %s pool because out of memory\n",
1863                        ps->ps_name);
1864         }
1865         spin_unlock(&ps->ps_lock);
1866
1867         goto again;
1868 }
1869
1870 static void
1871 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1872 {
1873         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1874         int             i;
1875
1876         LASSERT (pool->po_allocated == 0);
1877
1878         if (tpo->tpo_tx_pages != NULL) {
1879                 kiblnd_unmap_tx_pool(tpo);
1880                 kiblnd_free_pages(tpo->tpo_tx_pages);
1881         }
1882
1883         if (tpo->tpo_tx_descs == NULL)
1884                 goto out;
1885
1886         for (i = 0; i < pool->po_size; i++) {
1887                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1888
1889                 list_del(&tx->tx_list);
1890                 if (tx->tx_pages != NULL)
1891                         LIBCFS_FREE(tx->tx_pages,
1892                                     LNET_MAX_IOV *
1893                                     sizeof(*tx->tx_pages));
1894                 if (tx->tx_frags != NULL)
1895                         LIBCFS_FREE(tx->tx_frags,
1896                                     IBLND_MAX_RDMA_FRAGS *
1897                                             sizeof(*tx->tx_frags));
1898                 if (tx->tx_wrq != NULL)
1899                         LIBCFS_FREE(tx->tx_wrq,
1900                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1901                                     sizeof(*tx->tx_wrq));
1902                 if (tx->tx_sge != NULL)
1903                         LIBCFS_FREE(tx->tx_sge,
1904                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1905                                     sizeof(*tx->tx_sge));
1906                 if (tx->tx_rd != NULL)
1907                         LIBCFS_FREE(tx->tx_rd,
1908                                     offsetof(kib_rdma_desc_t,
1909                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1910         }
1911
1912         LIBCFS_FREE(tpo->tpo_tx_descs,
1913                     pool->po_size * sizeof(kib_tx_t));
1914 out:
1915         kiblnd_fini_pool(pool);
1916         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1917 }
1918
1919 static int kiblnd_tx_pool_size(int ncpts)
1920 {
1921         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1922
1923         return max(IBLND_TX_POOL, ntx);
1924 }
1925
1926 static int
1927 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1928 {
1929         int            i;
1930         int            npg;
1931         kib_pool_t    *pool;
1932         kib_tx_pool_t *tpo;
1933
1934         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1935         if (tpo == NULL) {
1936                 CERROR("Failed to allocate TX pool\n");
1937                 return -ENOMEM;
1938         }
1939
1940         pool = &tpo->tpo_pool;
1941         kiblnd_init_pool(ps, pool, size);
1942         tpo->tpo_tx_descs = NULL;
1943         tpo->tpo_tx_pages = NULL;
1944
1945         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1946         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1947                 CERROR("Can't allocate tx pages: %d\n", npg);
1948                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1949                 return -ENOMEM;
1950         }
1951
1952         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1953                          size * sizeof(kib_tx_t));
1954         if (tpo->tpo_tx_descs == NULL) {
1955                 CERROR("Can't allocate %d tx descriptors\n", size);
1956                 ps->ps_pool_destroy(pool);
1957                 return -ENOMEM;
1958         }
1959
1960         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1961
1962         for (i = 0; i < size; i++) {
1963                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1964
1965                 tx->tx_pool = tpo;
1966                 if (ps->ps_net->ibn_fmr_ps != NULL) {
1967                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1968                                          lnet_cpt_table(), ps->ps_cpt,
1969                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1970                         if (tx->tx_pages == NULL)
1971                                 break;
1972                 }
1973
1974                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1975                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1976                 if (tx->tx_frags == NULL)
1977                         break;
1978
1979                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1980
1981                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1982                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1983                                  sizeof(*tx->tx_wrq));
1984                 if (tx->tx_wrq == NULL)
1985                         break;
1986
1987                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
1988                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1989                                  sizeof(*tx->tx_sge));
1990                 if (tx->tx_sge == NULL)
1991                         break;
1992
1993                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
1994                                  offsetof(kib_rdma_desc_t,
1995                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
1996                 if (tx->tx_rd == NULL)
1997                         break;
1998         }
1999
2000         if (i == size) {
2001                 kiblnd_map_tx_pool(tpo);
2002                 *pp_po = pool;
2003                 return 0;
2004         }
2005
2006         ps->ps_pool_destroy(pool);
2007         return -ENOMEM;
2008 }
2009
2010 static void
2011 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2012 {
2013         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2014                                              tps_poolset);
2015         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2016
2017         tx->tx_cookie = tps->tps_next_tx_cookie++;
2018 }
2019
2020 static void
2021 kiblnd_net_fini_pools(kib_net_t *net)
2022 {
2023         int     i;
2024
2025         cfs_cpt_for_each(i, lnet_cpt_table()) {
2026                 kib_tx_poolset_t        *tps;
2027                 kib_fmr_poolset_t       *fps;
2028
2029                 if (net->ibn_tx_ps != NULL) {
2030                         tps = net->ibn_tx_ps[i];
2031                         kiblnd_fini_poolset(&tps->tps_poolset);
2032                 }
2033
2034                 if (net->ibn_fmr_ps != NULL) {
2035                         fps = net->ibn_fmr_ps[i];
2036                         kiblnd_fini_fmr_poolset(fps);
2037                 }
2038         }
2039
2040         if (net->ibn_tx_ps != NULL) {
2041                 cfs_percpt_free(net->ibn_tx_ps);
2042                 net->ibn_tx_ps = NULL;
2043         }
2044
2045         if (net->ibn_fmr_ps != NULL) {
2046                 cfs_percpt_free(net->ibn_fmr_ps);
2047                 net->ibn_fmr_ps = NULL;
2048         }
2049 }
2050
2051 static int
2052 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2053 {
2054         unsigned long   flags;
2055         int             cpt;
2056         int             rc;
2057         int             i;
2058
2059         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2060         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2061                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2062                                            flags);
2063                 goto create_tx_pool;
2064         }
2065
2066         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2067
2068         if (*kiblnd_tunables.kib_fmr_pool_size <
2069             *kiblnd_tunables.kib_ntx / 4) {
2070                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2071                        *kiblnd_tunables.kib_fmr_pool_size,
2072                        *kiblnd_tunables.kib_ntx / 4);
2073                 rc = -EINVAL;
2074                 goto failed;
2075         }
2076
2077         /* TX pool must be created later than FMR, see LU-2268
2078          * for details */
2079         LASSERT(net->ibn_tx_ps == NULL);
2080
2081         /* premapping can fail if ibd_nmr > 1, so we always create
2082          * FMR pool and map-on-demand if premapping failed */
2083
2084         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2085                                            sizeof(kib_fmr_poolset_t));
2086         if (net->ibn_fmr_ps == NULL) {
2087                 CERROR("Failed to allocate FMR pool array\n");
2088                 rc = -ENOMEM;
2089                 goto failed;
2090         }
2091
2092         for (i = 0; i < ncpts; i++) {
2093                 cpt = (cpts == NULL) ? i : cpts[i];
2094                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2095                                              kiblnd_fmr_pool_size(ncpts),
2096                                              kiblnd_fmr_flush_trigger(ncpts));
2097                 if (rc != 0) {
2098                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2099                                cpt, rc);
2100                         goto failed;
2101                 }
2102         }
2103
2104         if (i > 0)
2105                 LASSERT(i == ncpts);
2106
2107  create_tx_pool:
2108         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2109                                           sizeof(kib_tx_poolset_t));
2110         if (net->ibn_tx_ps == NULL) {
2111                 CERROR("Failed to allocate tx pool array\n");
2112                 rc = -ENOMEM;
2113                 goto failed;
2114         }
2115
2116         for (i = 0; i < ncpts; i++) {
2117                 cpt = (cpts == NULL) ? i : cpts[i];
2118                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2119                                          cpt, net, "TX",
2120                                          kiblnd_tx_pool_size(ncpts),
2121                                          kiblnd_create_tx_pool,
2122                                          kiblnd_destroy_tx_pool,
2123                                          kiblnd_tx_init, NULL);
2124                 if (rc != 0) {
2125                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2126                                cpt, rc);
2127                         goto failed;
2128                 }
2129         }
2130
2131         return 0;
2132  failed:
2133         kiblnd_net_fini_pools(net);
2134         LASSERT(rc != 0);
2135         return rc;
2136 }
2137
2138 static int
2139 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2140 {
2141         struct ib_device_attr *attr;
2142         int                    rc;
2143
2144         /* It's safe to assume a HCA can handle a page size
2145          * matching that of the native system */
2146         hdev->ibh_page_shift = PAGE_SHIFT;
2147         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2148         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2149
2150         LIBCFS_ALLOC(attr, sizeof(*attr));
2151         if (attr == NULL) {
2152                 CERROR("Out of memory\n");
2153                 return -ENOMEM;
2154         }
2155
2156         rc = ib_query_device(hdev->ibh_ibdev, attr);
2157         if (rc == 0)
2158                 hdev->ibh_mr_size = attr->max_mr_size;
2159
2160         LIBCFS_FREE(attr, sizeof(*attr));
2161
2162         if (rc != 0) {
2163                 CERROR("Failed to query IB device: %d\n", rc);
2164                 return rc;
2165         }
2166
2167         if (hdev->ibh_mr_size == ~0ULL) {
2168                 hdev->ibh_mr_shift = 64;
2169                 return 0;
2170         }
2171
2172         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2173         return -EINVAL;
2174 }
2175
2176 static void
2177 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2178 {
2179         if (hdev->ibh_mrs == NULL)
2180                 return;
2181
2182         ib_dereg_mr(hdev->ibh_mrs);
2183
2184         hdev->ibh_mrs = NULL;
2185 }
2186
2187 void
2188 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2189 {
2190         kiblnd_hdev_cleanup_mrs(hdev);
2191
2192         if (hdev->ibh_pd != NULL)
2193                 ib_dealloc_pd(hdev->ibh_pd);
2194
2195         if (hdev->ibh_cmid != NULL)
2196                 rdma_destroy_id(hdev->ibh_cmid);
2197
2198         LIBCFS_FREE(hdev, sizeof(*hdev));
2199 }
2200
2201 static int
2202 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2203 {
2204         struct ib_mr *mr;
2205         int           rc;
2206         int           acflags = IB_ACCESS_LOCAL_WRITE |
2207                                 IB_ACCESS_REMOTE_WRITE;
2208
2209         rc = kiblnd_hdev_get_attr(hdev);
2210         if (rc != 0)
2211                 return rc;
2212
2213         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2214         if (IS_ERR(mr)) {
2215                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2216                 kiblnd_hdev_cleanup_mrs(hdev);
2217                 return PTR_ERR(mr);
2218         }
2219
2220         hdev->ibh_mrs = mr;
2221
2222         return 0;
2223 }
2224
2225 static int
2226 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2227 {       /* DUMMY */
2228         return 0;
2229 }
2230
2231 static int
2232 kiblnd_dev_need_failover(kib_dev_t *dev)
2233 {
2234         struct rdma_cm_id  *cmid;
2235         struct sockaddr_in  srcaddr;
2236         struct sockaddr_in  dstaddr;
2237         int                 rc;
2238
2239         if (dev->ibd_hdev == NULL || /* initializing */
2240             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2241             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2242                 return 1;
2243
2244         /* XXX: it's UGLY, but I don't have better way to find
2245          * ib-bonding HCA failover because:
2246          *
2247          * a. no reliable CM event for HCA failover...
2248          * b. no OFED API to get ib_device for current net_device...
2249          *
2250          * We have only two choices at this point:
2251          *
2252          * a. rdma_bind_addr(), it will conflict with listener cmid
2253          * b. rdma_resolve_addr() to zero addr */
2254         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2255                                      IB_QPT_RC);
2256         if (IS_ERR(cmid)) {
2257                 rc = PTR_ERR(cmid);
2258                 CERROR("Failed to create cmid for failover: %d\n", rc);
2259                 return rc;
2260         }
2261
2262         memset(&srcaddr, 0, sizeof(srcaddr));
2263         srcaddr.sin_family      = AF_INET;
2264         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2265
2266         memset(&dstaddr, 0, sizeof(dstaddr));
2267         dstaddr.sin_family = AF_INET;
2268         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2269                                (struct sockaddr *)&dstaddr, 1);
2270         if (rc != 0 || cmid->device == NULL) {
2271                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2272                        dev->ibd_ifname, &dev->ibd_ifip,
2273                        cmid->device, rc);
2274                 rdma_destroy_id(cmid);
2275                 return rc;
2276         }
2277
2278         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2279         rdma_destroy_id(cmid);
2280         return rc;
2281 }
2282
2283 int
2284 kiblnd_dev_failover(kib_dev_t *dev)
2285 {
2286         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2287         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2288         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2289         struct rdma_cm_id  *cmid  = NULL;
2290         kib_hca_dev_t      *hdev  = NULL;
2291         kib_hca_dev_t      *old;
2292         struct ib_pd       *pd;
2293         kib_net_t          *net;
2294         struct sockaddr_in  addr;
2295         unsigned long       flags;
2296         int                 rc = 0;
2297         int                 i;
2298
2299         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2300                  dev->ibd_can_failover ||
2301                  dev->ibd_hdev == NULL);
2302
2303         rc = kiblnd_dev_need_failover(dev);
2304         if (rc <= 0)
2305                 goto out;
2306
2307         if (dev->ibd_hdev != NULL &&
2308             dev->ibd_hdev->ibh_cmid != NULL) {
2309                 /* XXX it's not good to close old listener at here,
2310                  * because we can fail to create new listener.
2311                  * But we have to close it now, otherwise rdma_bind_addr
2312                  * will return EADDRINUSE... How crap! */
2313                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2314
2315                 cmid = dev->ibd_hdev->ibh_cmid;
2316                 /* make next schedule of kiblnd_dev_need_failover()
2317                  * return 1 for me */
2318                 dev->ibd_hdev->ibh_cmid  = NULL;
2319                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2320
2321                 rdma_destroy_id(cmid);
2322         }
2323
2324         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2325                                      IB_QPT_RC);
2326         if (IS_ERR(cmid)) {
2327                 rc = PTR_ERR(cmid);
2328                 CERROR("Failed to create cmid for failover: %d\n", rc);
2329                 goto out;
2330         }
2331
2332         memset(&addr, 0, sizeof(addr));
2333         addr.sin_family      = AF_INET;
2334         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2335         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2336
2337         /* Bind to failover device or port */
2338         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2339         if (rc != 0 || cmid->device == NULL) {
2340                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2341                        dev->ibd_ifname, &dev->ibd_ifip,
2342                        cmid->device, rc);
2343                 rdma_destroy_id(cmid);
2344                 goto out;
2345         }
2346
2347         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2348         if (hdev == NULL) {
2349                 CERROR("Failed to allocate kib_hca_dev\n");
2350                 rdma_destroy_id(cmid);
2351                 rc = -ENOMEM;
2352                 goto out;
2353         }
2354
2355         atomic_set(&hdev->ibh_ref, 1);
2356         hdev->ibh_dev   = dev;
2357         hdev->ibh_cmid  = cmid;
2358         hdev->ibh_ibdev = cmid->device;
2359
2360         pd = ib_alloc_pd(cmid->device);
2361         if (IS_ERR(pd)) {
2362                 rc = PTR_ERR(pd);
2363                 CERROR("Can't allocate PD: %d\n", rc);
2364                 goto out;
2365         }
2366
2367         hdev->ibh_pd = pd;
2368
2369         rc = rdma_listen(cmid, 0);
2370         if (rc != 0) {
2371                 CERROR("Can't start new listener: %d\n", rc);
2372                 goto out;
2373         }
2374
2375         rc = kiblnd_hdev_setup_mrs(hdev);
2376         if (rc != 0) {
2377                 CERROR("Can't setup device: %d\n", rc);
2378                 goto out;
2379         }
2380
2381         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2382
2383         old = dev->ibd_hdev;
2384         dev->ibd_hdev = hdev;   /* take over the refcount */
2385         hdev = old;
2386
2387         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2388                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2389                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2390                                             &zombie_tpo);
2391
2392                         if (net->ibn_fmr_ps != NULL)
2393                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2394                                                         &zombie_fpo);
2395                 }
2396         }
2397
2398         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2399  out:
2400         if (!list_empty(&zombie_tpo))
2401                 kiblnd_destroy_pool_list(&zombie_tpo);
2402         if (!list_empty(&zombie_ppo))
2403                 kiblnd_destroy_pool_list(&zombie_ppo);
2404         if (!list_empty(&zombie_fpo))
2405                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2406         if (hdev != NULL)
2407                 kiblnd_hdev_decref(hdev);
2408
2409         if (rc != 0)
2410                 dev->ibd_failed_failover++;
2411         else
2412                 dev->ibd_failed_failover = 0;
2413
2414         return rc;
2415 }
2416
2417 void
2418 kiblnd_destroy_dev (kib_dev_t *dev)
2419 {
2420         LASSERT (dev->ibd_nnets == 0);
2421         LASSERT(list_empty(&dev->ibd_nets));
2422
2423         list_del(&dev->ibd_fail_list);
2424         list_del(&dev->ibd_list);
2425
2426         if (dev->ibd_hdev != NULL)
2427                 kiblnd_hdev_decref(dev->ibd_hdev);
2428
2429         LIBCFS_FREE(dev, sizeof(*dev));
2430 }
2431
2432 static kib_dev_t *
2433 kiblnd_create_dev(char *ifname)
2434 {
2435         struct net_device *netdev;
2436         kib_dev_t         *dev;
2437         __u32              netmask;
2438         __u32              ip;
2439         int                up;
2440         int                rc;
2441
2442         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2443         if (rc != 0) {
2444                 CERROR("Can't query IPoIB interface %s: %d\n",
2445                        ifname, rc);
2446                 return NULL;
2447         }
2448
2449         if (!up) {
2450                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2451                 return NULL;
2452         }
2453
2454         LIBCFS_ALLOC(dev, sizeof(*dev));
2455         if (dev == NULL)
2456                 return NULL;
2457
2458         memset(dev, 0, sizeof(*dev));
2459         netdev = dev_get_by_name(&init_net, ifname);
2460         if (netdev == NULL) {
2461                 dev->ibd_can_failover = 0;
2462         } else {
2463                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2464                 dev_put(netdev);
2465         }
2466
2467         INIT_LIST_HEAD(&dev->ibd_nets);
2468         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2469         INIT_LIST_HEAD(&dev->ibd_fail_list);
2470         dev->ibd_ifip = ip;
2471         strcpy(&dev->ibd_ifname[0], ifname);
2472
2473         /* initialize the device */
2474         rc = kiblnd_dev_failover(dev);
2475         if (rc != 0) {
2476                 CERROR("Can't initialize device: %d\n", rc);
2477                 LIBCFS_FREE(dev, sizeof(*dev));
2478                 return NULL;
2479         }
2480
2481         list_add_tail(&dev->ibd_list,
2482                           &kiblnd_data.kib_devs);
2483         return dev;
2484 }
2485
2486 static void
2487 kiblnd_base_shutdown(void)
2488 {
2489         struct kib_sched_info   *sched;
2490         int                     i;
2491
2492         LASSERT(list_empty(&kiblnd_data.kib_devs));
2493
2494         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2495                atomic_read(&libcfs_kmemory));
2496
2497         switch (kiblnd_data.kib_init) {
2498         default:
2499                 LBUG();
2500
2501         case IBLND_INIT_ALL:
2502         case IBLND_INIT_DATA:
2503                 LASSERT (kiblnd_data.kib_peers != NULL);
2504                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2505                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2506                 }
2507                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2508                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2509
2510                 /* flag threads to terminate; wake and wait for them to die */
2511                 kiblnd_data.kib_shutdown = 1;
2512
2513                 /* NB: we really want to stop scheduler threads net by net
2514                  * instead of the whole module, this should be improved
2515                  * with dynamic configuration LNet */
2516                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2517                         wake_up_all(&sched->ibs_waitq);
2518
2519                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2520                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2521
2522                 i = 2;
2523                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2524                         i++;
2525                         /* power of 2? */
2526                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2527                                "Waiting for %d threads to terminate\n",
2528                                atomic_read(&kiblnd_data.kib_nthreads));
2529                         set_current_state(TASK_UNINTERRUPTIBLE);
2530                         schedule_timeout(cfs_time_seconds(1));
2531                 }
2532
2533                 /* fall through */
2534
2535         case IBLND_INIT_NOTHING:
2536                 break;
2537         }
2538
2539         if (kiblnd_data.kib_peers != NULL) {
2540                 LIBCFS_FREE(kiblnd_data.kib_peers,
2541                             sizeof(struct list_head) *
2542                             kiblnd_data.kib_peer_hash_size);
2543         }
2544
2545         if (kiblnd_data.kib_scheds != NULL)
2546                 cfs_percpt_free(kiblnd_data.kib_scheds);
2547
2548         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2549                atomic_read(&libcfs_kmemory));
2550
2551         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2552         module_put(THIS_MODULE);
2553 }
2554
2555 static void
2556 kiblnd_shutdown (lnet_ni_t *ni)
2557 {
2558         kib_net_t        *net = ni->ni_data;
2559         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2560         int               i;
2561         unsigned long     flags;
2562
2563         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2564
2565         if (net == NULL)
2566                 goto out;
2567
2568         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2569                atomic_read(&libcfs_kmemory));
2570
2571         write_lock_irqsave(g_lock, flags);
2572         net->ibn_shutdown = 1;
2573         write_unlock_irqrestore(g_lock, flags);
2574
2575         switch (net->ibn_init) {
2576         default:
2577                 LBUG();
2578
2579         case IBLND_INIT_ALL:
2580                 /* nuke all existing peers within this net */
2581                 kiblnd_del_peer(ni, LNET_NID_ANY);
2582
2583                 /* Wait for all peer state to clean up */
2584                 i = 2;
2585                 while (atomic_read(&net->ibn_npeers) != 0) {
2586                         i++;
2587                         /* power of 2? */
2588                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2589                                "%s: waiting for %d peers to disconnect\n",
2590                                libcfs_nid2str(ni->ni_nid),
2591                                atomic_read(&net->ibn_npeers));
2592                         set_current_state(TASK_UNINTERRUPTIBLE);
2593                         schedule_timeout(cfs_time_seconds(1));
2594                 }
2595
2596                 kiblnd_net_fini_pools(net);
2597
2598                 write_lock_irqsave(g_lock, flags);
2599                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2600                 net->ibn_dev->ibd_nnets--;
2601                 list_del(&net->ibn_list);
2602                 write_unlock_irqrestore(g_lock, flags);
2603
2604                 /* fall through */
2605
2606         case IBLND_INIT_NOTHING:
2607                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2608
2609                 if (net->ibn_dev != NULL &&
2610                     net->ibn_dev->ibd_nnets == 0)
2611                         kiblnd_destroy_dev(net->ibn_dev);
2612
2613                 break;
2614         }
2615
2616         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2617                atomic_read(&libcfs_kmemory));
2618
2619         net->ibn_init = IBLND_INIT_NOTHING;
2620         ni->ni_data = NULL;
2621
2622         LIBCFS_FREE(net, sizeof(*net));
2623
2624 out:
2625         if (list_empty(&kiblnd_data.kib_devs))
2626                 kiblnd_base_shutdown();
2627         return;
2628 }
2629
2630 static int
2631 kiblnd_base_startup(void)
2632 {
2633         struct kib_sched_info   *sched;
2634         int                     rc;
2635         int                     i;
2636
2637         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2638
2639         try_module_get(THIS_MODULE);
2640         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2641
2642         rwlock_init(&kiblnd_data.kib_global_lock);
2643
2644         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2645         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2646
2647         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2648         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2649                      sizeof(struct list_head) *
2650                      kiblnd_data.kib_peer_hash_size);
2651         if (kiblnd_data.kib_peers == NULL)
2652                 goto failed;
2653
2654         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2655                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2656
2657         spin_lock_init(&kiblnd_data.kib_connd_lock);
2658         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2659         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2660         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2661         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2662
2663         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2664                                                   sizeof(*sched));
2665         if (kiblnd_data.kib_scheds == NULL)
2666                 goto failed;
2667
2668         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2669                 int     nthrs;
2670
2671                 spin_lock_init(&sched->ibs_lock);
2672                 INIT_LIST_HEAD(&sched->ibs_conns);
2673                 init_waitqueue_head(&sched->ibs_waitq);
2674
2675                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2676                 if (*kiblnd_tunables.kib_nscheds > 0) {
2677                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2678                 } else {
2679                         /* max to half of CPUs, another half is reserved for
2680                          * upper layer modules */
2681                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2682                 }
2683
2684                 sched->ibs_nthreads_max = nthrs;
2685                 sched->ibs_cpt = i;
2686         }
2687
2688         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2689
2690         /* lists/ptrs/locks initialised */
2691         kiblnd_data.kib_init = IBLND_INIT_DATA;
2692         /*****************************************************/
2693
2694         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2695         if (rc != 0) {
2696                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2697                 goto failed;
2698         }
2699
2700         if (*kiblnd_tunables.kib_dev_failover != 0)
2701                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2702                                          "kiblnd_failover");
2703
2704         if (rc != 0) {
2705                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2706                 goto failed;
2707         }
2708
2709         /* flag everything initialised */
2710         kiblnd_data.kib_init = IBLND_INIT_ALL;
2711         /*****************************************************/
2712
2713         return 0;
2714
2715  failed:
2716         kiblnd_base_shutdown();
2717         return -ENETDOWN;
2718 }
2719
2720 static int
2721 kiblnd_start_schedulers(struct kib_sched_info *sched)
2722 {
2723         int     rc = 0;
2724         int     nthrs;
2725         int     i;
2726
2727         if (sched->ibs_nthreads == 0) {
2728                 if (*kiblnd_tunables.kib_nscheds > 0) {
2729                         nthrs = sched->ibs_nthreads_max;
2730                 } else {
2731                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2732                                                sched->ibs_cpt);
2733                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2734                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2735                 }
2736         } else {
2737                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2738                 /* increase one thread if there is new interface */
2739                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2740         }
2741
2742         for (i = 0; i < nthrs; i++) {
2743                 long    id;
2744                 char    name[20];
2745                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2746                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2747                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2748                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2749                 if (rc == 0)
2750                         continue;
2751
2752                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2753                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2754                 break;
2755         }
2756
2757         sched->ibs_nthreads += i;
2758         return rc;
2759 }
2760
2761 static int
2762 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2763 {
2764         int     cpt;
2765         int     rc;
2766         int     i;
2767
2768         for (i = 0; i < ncpts; i++) {
2769                 struct kib_sched_info *sched;
2770
2771                 cpt = (cpts == NULL) ? i : cpts[i];
2772                 sched = kiblnd_data.kib_scheds[cpt];
2773
2774                 if (!newdev && sched->ibs_nthreads > 0)
2775                         continue;
2776
2777                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2778                 if (rc != 0) {
2779                         CERROR("Failed to start scheduler threads for %s\n",
2780                                dev->ibd_ifname);
2781                         return rc;
2782                 }
2783         }
2784         return 0;
2785 }
2786
2787 static kib_dev_t *
2788 kiblnd_dev_search(char *ifname)
2789 {
2790         kib_dev_t       *alias = NULL;
2791         kib_dev_t       *dev;
2792         char            *colon;
2793         char            *colon2;
2794
2795         colon = strchr(ifname, ':');
2796         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2797                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2798                         return dev;
2799
2800                 if (alias != NULL)
2801                         continue;
2802
2803                 colon2 = strchr(dev->ibd_ifname, ':');
2804                 if (colon != NULL)
2805                         *colon = 0;
2806                 if (colon2 != NULL)
2807                         *colon2 = 0;
2808
2809                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2810                         alias = dev;
2811
2812                 if (colon != NULL)
2813                         *colon = ':';
2814                 if (colon2 != NULL)
2815                         *colon2 = ':';
2816         }
2817         return alias;
2818 }
2819
2820 static int
2821 kiblnd_startup (lnet_ni_t *ni)
2822 {
2823         char                     *ifname;
2824         kib_dev_t                *ibdev = NULL;
2825         kib_net_t                *net;
2826         struct timeval            tv;
2827         unsigned long             flags;
2828         int                       rc;
2829         int                       newdev;
2830
2831         LASSERT (ni->ni_lnd == &the_o2iblnd);
2832
2833         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2834                 rc = kiblnd_base_startup();
2835                 if (rc != 0)
2836                         return rc;
2837         }
2838
2839         LIBCFS_ALLOC(net, sizeof(*net));
2840         ni->ni_data = net;
2841         if (net == NULL)
2842                 goto failed;
2843
2844         memset(net, 0, sizeof(*net));
2845
2846         do_gettimeofday(&tv);
2847         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2848
2849         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2850         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2851         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2852         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2853
2854         if (ni->ni_interfaces[0] != NULL) {
2855                 /* Use the IPoIB interface specified in 'networks=' */
2856
2857                 CLASSERT (LNET_MAX_INTERFACES > 1);
2858                 if (ni->ni_interfaces[1] != NULL) {
2859                         CERROR("Multiple interfaces not supported\n");
2860                         goto failed;
2861                 }
2862
2863                 ifname = ni->ni_interfaces[0];
2864         } else {
2865                 ifname = *kiblnd_tunables.kib_default_ipif;
2866         }
2867
2868         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2869                 CERROR("IPoIB interface name too long: %s\n", ifname);
2870                 goto failed;
2871         }
2872
2873         ibdev = kiblnd_dev_search(ifname);
2874
2875         newdev = ibdev == NULL;
2876         /* hmm...create kib_dev even for alias */
2877         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
2878                 ibdev = kiblnd_create_dev(ifname);
2879
2880         if (ibdev == NULL)
2881                 goto failed;
2882
2883         net->ibn_dev = ibdev;
2884         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2885
2886         rc = kiblnd_dev_start_threads(ibdev, newdev,
2887                                       ni->ni_cpts, ni->ni_ncpts);
2888         if (rc != 0)
2889                 goto failed;
2890
2891         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2892         if (rc != 0) {
2893                 CERROR("Failed to initialize NI pools: %d\n", rc);
2894                 goto failed;
2895         }
2896
2897         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2898         ibdev->ibd_nnets++;
2899         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2900         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2901
2902         net->ibn_init = IBLND_INIT_ALL;
2903
2904         return 0;
2905
2906 failed:
2907         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
2908                 kiblnd_destroy_dev(ibdev);
2909
2910         kiblnd_shutdown(ni);
2911
2912         CDEBUG(D_NET, "kiblnd_startup failed\n");
2913         return -ENETDOWN;
2914 }
2915
2916 static lnd_t the_o2iblnd = {
2917         .lnd_type       = O2IBLND,
2918         .lnd_startup    = kiblnd_startup,
2919         .lnd_shutdown   = kiblnd_shutdown,
2920         .lnd_ctl        = kiblnd_ctl,
2921         .lnd_query      = kiblnd_query,
2922         .lnd_send       = kiblnd_send,
2923         .lnd_recv       = kiblnd_recv,
2924 };
2925
2926 static void __exit ko2iblnd_exit(void)
2927 {
2928         lnet_unregister_lnd(&the_o2iblnd);
2929         kiblnd_tunables_fini();
2930 }
2931
2932 static int __init ko2iblnd_init(void)
2933 {
2934         int rc;
2935
2936         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2937         CLASSERT(offsetof(kib_msg_t,
2938                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
2939                  IBLND_MSG_SIZE);
2940         CLASSERT(offsetof(kib_msg_t,
2941                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2942                  <= IBLND_MSG_SIZE);
2943
2944         rc = kiblnd_tunables_init();
2945         if (rc != 0)
2946                 return rc;
2947
2948         lnet_register_lnd(&the_o2iblnd);
2949
2950         return 0;
2951 }
2952
2953 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2954 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
2955 MODULE_VERSION("2.8.0");
2956 MODULE_LICENSE("GPL");
2957
2958 module_init(ko2iblnd_init);
2959 module_exit(ko2iblnd_exit);