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         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 341
 342         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 343         INIT_LIST_HEAD(&peer->ibp_conns);
 344         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 345
 346         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 347
 348         /* always called with a ref on ni, which prevents ni being shutdown */
 349         LASSERT(net->ibn_shutdown == 0);
 350
 351         /* npeers only grows with the global lock held */
 352         atomic_inc(&net->ibn_npeers);
 353
 354         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 355
 356         *peerp = peer;
 357         return 0;
 358 }
 359
 360 void
 361 kiblnd_destroy_peer (kib_peer_t *peer)
 362 {
 363         kib_net_t *net = peer->ibp_ni->ni_data;
 364
 365         LASSERT(net != NULL);
 366         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 367         LASSERT(!kiblnd_peer_active(peer));
 368         LASSERT(peer->ibp_connecting == 0);
 369         LASSERT(peer->ibp_accepting == 0);
 370         LASSERT(list_empty(&peer->ibp_conns));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 387         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head        *tmp;
 389         kib_peer_t              *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394
 395                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 396                         peer->ibp_accepting > 0 ||
 397                         !list_empty(&peer->ibp_conns)); /* active conn */
 398
 399                 if (peer->ibp_nid != nid)
 400                         continue;
 401
 402                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 403                        peer, libcfs_nid2str(nid),
 404                        atomic_read(&peer->ibp_refcount),
 405                        peer->ibp_version);
 406                 return peer;
 407         }
 408         return NULL;
 409 }
 410
 411 void
 412 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 413 {
 414         LASSERT(list_empty(&peer->ibp_conns));
 415
 416         LASSERT (kiblnd_peer_active(peer));
 417         list_del_init(&peer->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer);
 420 }
 421
 422 static int
 423 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 424                      lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_t              *peer;
 427         struct list_head        *ptmp;
 428         int                      i;
 429         unsigned long            flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 438                         LASSERT(peer->ibp_connecting > 0 ||
 439                                 peer->ibp_accepting > 0 ||
 440                                 !list_empty(&peer->ibp_conns));
 441
 442                         if (peer->ibp_ni != ni)
 443                                 continue;
 444
 445                         if (index-- > 0)
 446                                 continue;
 447
 448                         *nidp = peer->ibp_nid;
 449                         *count = atomic_read(&peer->ibp_refcount);
 450
 451                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 452                                                flags);
 453                         return 0;
 454                 }
 455         }
 456
 457         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 458         return -ENOENT;
 459 }
 460
 461 static void
 462 kiblnd_del_peer_locked (kib_peer_t *peer)
 463 {
 464         struct list_head        *ctmp;
 465         struct list_head        *cnxt;
 466         kib_conn_t              *conn;
 467
 468         if (list_empty(&peer->ibp_conns)) {
 469                 kiblnd_unlink_peer_locked(peer);
 470         } else {
 471                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 472                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 473
 474                         kiblnd_close_conn_locked(conn, 0);
 475                 }
 476                 /* NB closing peer's last conn unlinked it. */
 477         }
 478         /* NB peer now unlinked; might even be freed if the peer table had the
 479          * last ref on it. */
 480 }
 481
 482 static int
 483 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 484 {
 485         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 486         struct list_head        *ptmp;
 487         struct list_head        *pnxt;
 488         kib_peer_t              *peer;
 489         int                     lo;
 490         int                     hi;
 491         int                     i;
 492         unsigned long           flags;
 493         int                     rc = -ENOENT;
 494
 495         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 496
 497         if (nid != LNET_NID_ANY) {
 498                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 499         } else {
 500                 lo = 0;
 501                 hi = kiblnd_data.kib_peer_hash_size - 1;
 502         }
 503
 504         for (i = lo; i <= hi; i++) {
 505                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 506                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 507                         LASSERT(peer->ibp_connecting > 0 ||
 508                                 peer->ibp_accepting > 0 ||
 509                                 !list_empty(&peer->ibp_conns));
 510
 511                         if (peer->ibp_ni != ni)
 512                                 continue;
 513
 514                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 515                                 continue;
 516
 517                         if (!list_empty(&peer->ibp_tx_queue)) {
 518                                 LASSERT(list_empty(&peer->ibp_conns));
 519
 520                                 list_splice_init(&peer->ibp_tx_queue,
 521                                                  &zombies);
 522                         }
 523
 524                         kiblnd_del_peer_locked(peer);
 525                         rc = 0;         /* matched something */
 526                 }
 527         }
 528
 529         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 530
 531         kiblnd_txlist_done(ni, &zombies, -EIO);
 532
 533         return rc;
 534 }
 535
 536 static kib_conn_t *
 537 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 538 {
 539         kib_peer_t              *peer;
 540         struct list_head        *ptmp;
 541         kib_conn_t              *conn;
 542         struct list_head        *ctmp;
 543         int                     i;
 544         unsigned long           flags;
 545
 546         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 547
 548         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 549                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 550
 551                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 552                         LASSERT(peer->ibp_connecting > 0 ||
 553                                 peer->ibp_accepting > 0 ||
 554                                 !list_empty(&peer->ibp_conns));
 555
 556                         if (peer->ibp_ni != ni)
 557                                 continue;
 558
 559                         list_for_each(ctmp, &peer->ibp_conns) {
 560                                 if (index-- > 0)
 561                                         continue;
 562
 563                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 564                                 kiblnd_conn_addref(conn);
 565                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 566                                                        flags);
 567                                 return conn;
 568                         }
 569                 }
 570         }
 571
 572         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 573         return NULL;
 574 }
 575
 576 static void
 577 kiblnd_debug_rx (kib_rx_t *rx)
 578 {
 579         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 580                rx, rx->rx_status, rx->rx_msg->ibm_type,
 581                rx->rx_msg->ibm_credits);
 582 }
 583
 584 static void
 585 kiblnd_debug_tx (kib_tx_t *tx)
 586 {
 587         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 588                "cookie "LPX64" msg %s%s type %x cred %d\n",
 589                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 590                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 591                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 592                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 593                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 594 }
 595
 596 void
 597 kiblnd_debug_conn (kib_conn_t *conn)
 598 {
 599         struct list_head        *tmp;
 600         int                     i;
 601
 602         spin_lock(&conn->ibc_lock);
 603
 604         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 605                atomic_read(&conn->ibc_refcount), conn,
 606                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 607         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 608                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 609                conn->ibc_nsends_posted, conn->ibc_credits,
 610                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 611         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 612
 613         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 614         list_for_each(tmp, &conn->ibc_early_rxs)
 615                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 616
 617         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 618         list_for_each(tmp, &conn->ibc_tx_noops)
 619                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 620
 621         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 622         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 623                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 624
 625         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 626         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 627                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 628
 629         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 630         list_for_each(tmp, &conn->ibc_tx_queue)
 631                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 632
 633         CDEBUG(D_CONSOLE, "   active_txs:\n");
 634         list_for_each(tmp, &conn->ibc_active_txs)
 635                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 636
 637         CDEBUG(D_CONSOLE, "   rxs:\n");
 638         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 639                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 640
 641         spin_unlock(&conn->ibc_lock);
 642 }
 643
 644 int
 645 kiblnd_translate_mtu(int value)
 646 {
 647         switch (value) {
 648         default:
 649                 return -1;
 650         case 0:
 651                 return 0;
 652         case 256:
 653                 return IB_MTU_256;
 654         case 512:
 655                 return IB_MTU_512;
 656         case 1024:
 657                 return IB_MTU_1024;
 658         case 2048:
 659                 return IB_MTU_2048;
 660         case 4096:
 661                 return IB_MTU_4096;
 662         }
 663 }
 664
 665 static void
 666 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 667 {
 668         int           mtu;
 669
 670         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 671         if (cmid->route.path_rec == NULL)
 672                 return;
 673
 674         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 675         LASSERT (mtu >= 0);
 676         if (mtu != 0)
 677                 cmid->route.path_rec->mtu = mtu;
 678 }
 679
 680 static int
 681 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 682 {
 683         cpumask_t       *mask;
 684         int             vectors;
 685         int             off;
 686         int             i;
 687
 688         vectors = conn->ibc_cmid->device->num_comp_vectors;
 689         if (vectors <= 1)
 690                 return 0;
 691
 692         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 693
 694         /* hash NID to CPU id in this partition... */
 695         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 696         for_each_cpu(i, mask) {
 697                 if (off-- == 0)
 698                         return i % vectors;
 699         }
 700
 701         LBUG();
 702         return 1;
 703 }
 704
 705 kib_conn_t *
 706 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 707                    int state, int version, kib_connparams_t *cp)
 708 {
 709         /* CAVEAT EMPTOR:
 710          * If the new conn is created successfully it takes over the caller's
 711          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 712          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 713          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 714          * to destroy 'cmid' here since I'm called from the CM which still has
 715          * its ref on 'cmid'). */
 716         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 717         kib_net_t              *net = peer->ibp_ni->ni_data;
 718         kib_dev_t              *dev;
 719         struct ib_qp_init_attr *init_qp_attr;
 720         struct kib_sched_info   *sched;
 721 #ifdef HAVE_IB_CQ_INIT_ATTR
 722         struct ib_cq_init_attr  cq_attr = {};
 723 #endif
 724         kib_conn_t              *conn;
 725         struct ib_cq            *cq;
 726         unsigned long           flags;
 727         int                     cpt;
 728         int                     rc;
 729         int                     i;
 730
 731         LASSERT(net != NULL);
 732         LASSERT(!in_interrupt());
 733
 734         dev = net->ibn_dev;
 735
 736         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 737         sched = kiblnd_data.kib_scheds[cpt];
 738
 739         LASSERT(sched->ibs_nthreads > 0);
 740
 741         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 742                          sizeof(*init_qp_attr));
 743         if (init_qp_attr == NULL) {
 744                 CERROR("Can't allocate qp_attr for %s\n",
 745                        libcfs_nid2str(peer->ibp_nid));
 746                 goto failed_0;
 747         }
 748
 749         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 750         if (conn == NULL) {
 751                 CERROR("Can't allocate connection for %s\n",
 752                        libcfs_nid2str(peer->ibp_nid));
 753                 goto failed_1;
 754         }
 755
 756         conn->ibc_state = IBLND_CONN_INIT;
 757         conn->ibc_version = version;
 758         conn->ibc_peer = peer;                  /* I take the caller's ref */
 759         cmid->context = conn;                   /* for future CM callbacks */
 760         conn->ibc_cmid = cmid;
 761
 762         if (cp == NULL) {
 763                 conn->ibc_max_frags = IBLND_CFG_RDMA_FRAGS;
 764                 conn->ibc_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 765         } else {
 766                 conn->ibc_max_frags = cp->ibcp_max_frags;
 767                 conn->ibc_queue_depth = cp->ibcp_queue_depth;
 768         }
 769
 770         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 771         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 772         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 773         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 774         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 775         INIT_LIST_HEAD(&conn->ibc_active_txs);
 776         spin_lock_init(&conn->ibc_lock);
 777
 778         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 779                          sizeof(*conn->ibc_connvars));
 780         if (conn->ibc_connvars == NULL) {
 781                 CERROR("Can't allocate in-progress connection state\n");
 782                 goto failed_2;
 783         }
 784
 785         write_lock_irqsave(glock, flags);
 786         if (dev->ibd_failover) {
 787                 write_unlock_irqrestore(glock, flags);
 788                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 789                 goto failed_2;
 790         }
 791
 792         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 793                 /* wakeup failover thread and teardown connection */
 794                 if (kiblnd_dev_can_failover(dev)) {
 795                         list_add_tail(&dev->ibd_fail_list,
 796                                       &kiblnd_data.kib_failed_devs);
 797                         wake_up(&kiblnd_data.kib_failover_waitq);
 798                 }
 799
 800                 write_unlock_irqrestore(glock, flags);
 801                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 802                        cmid->device->name, dev->ibd_ifname);
 803                 goto failed_2;
 804         }
 805
 806         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 807         conn->ibc_hdev = dev->ibd_hdev;
 808
 809         kiblnd_setup_mtu_locked(cmid);
 810
 811         write_unlock_irqrestore(glock, flags);
 812
 813         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 814                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 815         if (conn->ibc_rxs == NULL) {
 816                 CERROR("Cannot allocate RX buffers\n");
 817                 goto failed_2;
 818         }
 819
 820         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 821                                 IBLND_RX_MSG_PAGES(conn));
 822         if (rc != 0)
 823                 goto failed_2;
 824
 825         kiblnd_map_rx_descs(conn);
 826
 827 #ifdef HAVE_IB_CQ_INIT_ATTR
 828         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 829         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 830         cq = ib_create_cq(cmid->device,
 831                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 832                           &cq_attr);
 833 #else
 834         cq = ib_create_cq(cmid->device,
 835                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 836                           IBLND_CQ_ENTRIES(conn),
 837                           kiblnd_get_completion_vector(conn, cpt));
 838 #endif
 839         if (IS_ERR(cq)) {
 840                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 841                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 842                 goto failed_2;
 843         }
 844
 845         conn->ibc_cq = cq;
 846
 847         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 848         if (rc != 0) {
 849                 CERROR("Can't request completion notification: %d\n", rc);
 850                 goto failed_2;
 851         }
 852
 853         init_qp_attr->event_handler = kiblnd_qp_event;
 854         init_qp_attr->qp_context = conn;
 855         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 856         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 857         init_qp_attr->cap.max_send_sge = 1;
 858         init_qp_attr->cap.max_recv_sge = 1;
 859         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 860         init_qp_attr->qp_type = IB_QPT_RC;
 861         init_qp_attr->send_cq = cq;
 862         init_qp_attr->recv_cq = cq;
 863
 864         conn->ibc_sched = sched;
 865
 866         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 867         if (rc != 0) {
 868                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 869                        rc, init_qp_attr->cap.max_send_wr,
 870                        init_qp_attr->cap.max_recv_wr);
 871                 goto failed_2;
 872         }
 873
 874         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 875
 876         /* 1 ref for caller and each rxmsg */
 877         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 878         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 879
 880         /* post receives */
 881         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 882                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 883                                     IBLND_POSTRX_NO_CREDIT);
 884                 if (rc != 0) {
 885                         CERROR("Can't post rxmsg: %d\n", rc);
 886
 887                         /* Make posted receives complete */
 888                         kiblnd_abort_receives(conn);
 889
 890                         /* correct # of posted buffers
 891                          * NB locking needed now I'm racing with completion */
 892                         spin_lock_irqsave(&sched->ibs_lock, flags);
 893                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 894                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 895
 896                         /* cmid will be destroyed by CM(ofed) after cm_callback
 897                          * returned, so we can't refer it anymore
 898                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 899                         rdma_destroy_qp(conn->ibc_cmid);
 900                         conn->ibc_cmid = NULL;
 901
 902                         /* Drop my own and unused rxbuffer refcounts */
 903                         while (i++ <= IBLND_RX_MSGS(conn))
 904                                 kiblnd_conn_decref(conn);
 905
 906                         return NULL;
 907                 }
 908         }
 909
 910         /* Init successful! */
 911         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 912                  state == IBLND_CONN_PASSIVE_WAIT);
 913         conn->ibc_state = state;
 914
 915         /* 1 more conn */
 916         atomic_inc(&net->ibn_nconns);
 917         return conn;
 918
 919  failed_2:
 920         kiblnd_destroy_conn(conn);
 921  failed_1:
 922         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 923  failed_0:
 924         return NULL;
 925 }
 926
 927 void
 928 kiblnd_destroy_conn (kib_conn_t *conn)
 929 {
 930         struct rdma_cm_id *cmid = conn->ibc_cmid;
 931         kib_peer_t        *peer = conn->ibc_peer;
 932         int                rc;
 933
 934         LASSERT (!in_interrupt());
 935         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 936         LASSERT(list_empty(&conn->ibc_early_rxs));
 937         LASSERT(list_empty(&conn->ibc_tx_noops));
 938         LASSERT(list_empty(&conn->ibc_tx_queue));
 939         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 940         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 941         LASSERT(list_empty(&conn->ibc_active_txs));
 942         LASSERT (conn->ibc_noops_posted == 0);
 943         LASSERT (conn->ibc_nsends_posted == 0);
 944
 945         switch (conn->ibc_state) {
 946         default:
 947                 /* conn must be completely disengaged from the network */
 948                 LBUG();
 949
 950         case IBLND_CONN_DISCONNECTED:
 951                 /* connvars should have been freed already */
 952                 LASSERT (conn->ibc_connvars == NULL);
 953                 break;
 954
 955         case IBLND_CONN_INIT:
 956                 break;
 957         }
 958
 959         /* conn->ibc_cmid might be destroyed by CM already */
 960         if (cmid != NULL && cmid->qp != NULL)
 961                 rdma_destroy_qp(cmid);
 962
 963         if (conn->ibc_cq != NULL) {
 964                 rc = ib_destroy_cq(conn->ibc_cq);
 965                 if (rc != 0)
 966                         CWARN("Error destroying CQ: %d\n", rc);
 967         }
 968
 969         if (conn->ibc_rx_pages != NULL)
 970                 kiblnd_unmap_rx_descs(conn);
 971
 972         if (conn->ibc_rxs != NULL) {
 973                 LIBCFS_FREE(conn->ibc_rxs,
 974                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 975         }
 976
 977         if (conn->ibc_connvars != NULL)
 978                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 979
 980         if (conn->ibc_hdev != NULL)
 981                 kiblnd_hdev_decref(conn->ibc_hdev);
 982
 983         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 984         if (conn->ibc_state != IBLND_CONN_INIT) {
 985                 kib_net_t *net = peer->ibp_ni->ni_data;
 986
 987                 rdma_destroy_id(cmid);
 988                 atomic_dec(&net->ibn_nconns);
 989                 if (conn->ibc_conn_race) {
 990                         if (peer->ibp_accepting == 0 &&
 991                             !list_empty(&peer->ibp_tx_queue)) {
 992                                 kiblnd_connect_peer(peer);
 993                         } else  {
 994                                 rwlock_t *glock = &kiblnd_data.kib_global_lock;
 995                                 unsigned long flags;
 996
 997                                 write_lock_irqsave(glock, flags);
 998                                 peer->ibp_connecting--;
 999                                 write_unlock_irqrestore(glock, flags);
1000                         }
1001                 }
1002                 kiblnd_peer_decref(peer);
1003         }
1004
1005         LIBCFS_FREE(conn, sizeof(*conn));
1006 }
1007
1008 int
1009 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
1010 {
1011         kib_conn_t              *conn;
1012         struct list_head        *ctmp;
1013         struct list_head        *cnxt;
1014         int                     count = 0;
1015
1016         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1017                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1018
1019                 CDEBUG(D_NET, "Closing conn -> %s, "
1020                               "version: %x, reason: %d\n",
1021                        libcfs_nid2str(peer->ibp_nid),
1022                        conn->ibc_version, why);
1023
1024                 kiblnd_close_conn_locked(conn, why);
1025                 count++;
1026         }
1027
1028         return count;
1029 }
1030
1031 int
1032 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1033                                 int version, __u64 incarnation)
1034 {
1035         kib_conn_t              *conn;
1036         struct list_head        *ctmp;
1037         struct list_head        *cnxt;
1038         int                     count = 0;
1039
1040         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1041                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1042
1043                 if (conn->ibc_version     == version &&
1044                     conn->ibc_incarnation == incarnation)
1045                         continue;
1046
1047                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1048                               "incarnation:"LPX64"(%x, "LPX64")\n",
1049                        libcfs_nid2str(peer->ibp_nid),
1050                        conn->ibc_version, conn->ibc_incarnation,
1051                        version, incarnation);
1052
1053                 kiblnd_close_conn_locked(conn, -ESTALE);
1054                 count++;
1055         }
1056
1057         return count;
1058 }
1059
1060 static int
1061 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1062 {
1063         kib_peer_t              *peer;
1064         struct list_head        *ptmp;
1065         struct list_head        *pnxt;
1066         int                     lo;
1067         int                     hi;
1068         int                     i;
1069         unsigned long           flags;
1070         int                     count = 0;
1071
1072         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1073
1074         if (nid != LNET_NID_ANY)
1075                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1076         else {
1077                 lo = 0;
1078                 hi = kiblnd_data.kib_peer_hash_size - 1;
1079         }
1080
1081         for (i = lo; i <= hi; i++) {
1082                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1083
1084                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1085                         LASSERT(peer->ibp_connecting > 0 ||
1086                                 peer->ibp_accepting > 0 ||
1087                                 !list_empty(&peer->ibp_conns));
1088
1089                         if (peer->ibp_ni != ni)
1090                                 continue;
1091
1092                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1093                                 continue;
1094
1095                         count += kiblnd_close_peer_conns_locked(peer, 0);
1096                 }
1097         }
1098
1099         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1100
1101         /* wildcards always succeed */
1102         if (nid == LNET_NID_ANY)
1103                 return 0;
1104
1105         return (count == 0) ? -ENOENT : 0;
1106 }
1107
1108 static int
1109 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1110 {
1111         struct libcfs_ioctl_data *data = arg;
1112         int                       rc = -EINVAL;
1113
1114         switch(cmd) {
1115         case IOC_LIBCFS_GET_PEER: {
1116                 lnet_nid_t   nid = 0;
1117                 int          count = 0;
1118
1119                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1120                                           &nid, &count);
1121                 data->ioc_nid    = nid;
1122                 data->ioc_count  = count;
1123                 break;
1124         }
1125
1126         case IOC_LIBCFS_DEL_PEER: {
1127                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1128                 break;
1129         }
1130         case IOC_LIBCFS_GET_CONN: {
1131                 kib_conn_t *conn;
1132
1133                 rc = 0;
1134                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1135                 if (conn == NULL) {
1136                         rc = -ENOENT;
1137                         break;
1138                 }
1139
1140                 LASSERT (conn->ibc_cmid != NULL);
1141                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1142                 if (conn->ibc_cmid->route.path_rec == NULL)
1143                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1144                 else
1145                         data->ioc_u32[0] =
1146                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1147                 kiblnd_conn_decref(conn);
1148                 break;
1149         }
1150         case IOC_LIBCFS_CLOSE_CONNECTION: {
1151                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1152                 break;
1153         }
1154
1155         default:
1156                 break;
1157         }
1158
1159         return rc;
1160 }
1161
1162 static void
1163 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1164 {
1165         cfs_time_t      last_alive = 0;
1166         cfs_time_t      now = cfs_time_current();
1167         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1168         kib_peer_t      *peer;
1169         unsigned long   flags;
1170
1171         read_lock_irqsave(glock, flags);
1172
1173         peer = kiblnd_find_peer_locked(nid);
1174         if (peer != NULL) {
1175                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1176                         peer->ibp_accepting > 0 ||
1177                         !list_empty(&peer->ibp_conns));  /* active conn */
1178                 last_alive = peer->ibp_last_alive;
1179         }
1180
1181         read_unlock_irqrestore(glock, flags);
1182
1183         if (last_alive != 0)
1184                 *when = last_alive;
1185
1186         /* peer is not persistent in hash, trigger peer creation
1187          * and connection establishment with a NULL tx */
1188         if (peer == NULL)
1189                 kiblnd_launch_tx(ni, NULL, nid);
1190
1191         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1192                libcfs_nid2str(nid), peer,
1193                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1194         return;
1195 }
1196
1197 static void
1198 kiblnd_free_pages(kib_pages_t *p)
1199 {
1200         int     npages = p->ibp_npages;
1201         int     i;
1202
1203         for (i = 0; i < npages; i++) {
1204                 if (p->ibp_pages[i] != NULL)
1205                         __free_page(p->ibp_pages[i]);
1206         }
1207
1208         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1209 }
1210
1211 int
1212 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1213 {
1214         kib_pages_t     *p;
1215         int             i;
1216
1217         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1218                          offsetof(kib_pages_t, ibp_pages[npages]));
1219         if (p == NULL) {
1220                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1221                 return -ENOMEM;
1222         }
1223
1224         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1225         p->ibp_npages = npages;
1226
1227         for (i = 0; i < npages; i++) {
1228                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1229                                                      GFP_NOFS);
1230                 if (p->ibp_pages[i] == NULL) {
1231                         CERROR("Can't allocate page %d of %d\n", i, npages);
1232                         kiblnd_free_pages(p);
1233                         return -ENOMEM;
1234                 }
1235         }
1236
1237         *pp = p;
1238         return 0;
1239 }
1240
1241 void
1242 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1243 {
1244         kib_rx_t *rx;
1245         int       i;
1246
1247         LASSERT (conn->ibc_rxs != NULL);
1248         LASSERT (conn->ibc_hdev != NULL);
1249
1250         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1251                 rx = &conn->ibc_rxs[i];
1252
1253                 LASSERT(rx->rx_nob >= 0); /* not posted */
1254
1255                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1256                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1257                                                           rx->rx_msgaddr),
1258                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1259         }
1260
1261         kiblnd_free_pages(conn->ibc_rx_pages);
1262
1263         conn->ibc_rx_pages = NULL;
1264 }
1265
1266 void
1267 kiblnd_map_rx_descs(kib_conn_t *conn)
1268 {
1269         kib_rx_t       *rx;
1270         struct page    *pg;
1271         int             pg_off;
1272         int             ipg;
1273         int             i;
1274
1275         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1276                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1277                 rx = &conn->ibc_rxs[i];
1278
1279                 rx->rx_conn = conn;
1280                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1281
1282                 rx->rx_msgaddr =
1283                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1284                                               rx->rx_msg, IBLND_MSG_SIZE,
1285                                               DMA_FROM_DEVICE);
1286                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1287                                                   rx->rx_msgaddr));
1288                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1289
1290                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1291                        i, rx->rx_msg, rx->rx_msgaddr,
1292                        (__u64)(page_to_phys(pg) + pg_off));
1293
1294                 pg_off += IBLND_MSG_SIZE;
1295                 LASSERT(pg_off <= PAGE_SIZE);
1296
1297                 if (pg_off == PAGE_SIZE) {
1298                         pg_off = 0;
1299                         ipg++;
1300                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1301                 }
1302         }
1303 }
1304
1305 static void
1306 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1307 {
1308         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1309         kib_tx_t       *tx;
1310         int             i;
1311
1312         LASSERT (tpo->tpo_pool.po_allocated == 0);
1313
1314         if (hdev == NULL)
1315                 return;
1316
1317         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1318                 tx = &tpo->tpo_tx_descs[i];
1319                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1320                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1321                                                           tx->tx_msgaddr),
1322                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1323         }
1324
1325         kiblnd_hdev_decref(hdev);
1326         tpo->tpo_hdev = NULL;
1327 }
1328
1329 static kib_hca_dev_t *
1330 kiblnd_current_hdev(kib_dev_t *dev)
1331 {
1332         kib_hca_dev_t *hdev;
1333         unsigned long  flags;
1334         int            i = 0;
1335
1336         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1337         while (dev->ibd_failover) {
1338                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1339                 if (i++ % 50 == 0)
1340                         CDEBUG(D_NET, "%s: Wait for failover\n",
1341                                dev->ibd_ifname);
1342                 schedule_timeout(cfs_time_seconds(1) / 100);
1343
1344                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1345         }
1346
1347         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1348         hdev = dev->ibd_hdev;
1349
1350         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1351
1352         return hdev;
1353 }
1354
1355 static void
1356 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1357 {
1358         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1359         kib_pool_t     *pool  = &tpo->tpo_pool;
1360         kib_net_t      *net   = pool->po_owner->ps_net;
1361         kib_dev_t      *dev;
1362         struct page    *page;
1363         kib_tx_t       *tx;
1364         int             page_offset;
1365         int             ipage;
1366         int             i;
1367
1368         LASSERT (net != NULL);
1369
1370         dev = net->ibn_dev;
1371
1372         /* pre-mapped messages are not bigger than 1 page */
1373         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1374
1375         /* No fancy arithmetic when we do the buffer calculations */
1376         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1377
1378         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1379
1380         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1381                 page = txpgs->ibp_pages[ipage];
1382                 tx = &tpo->tpo_tx_descs[i];
1383
1384                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1385                                            page_offset);
1386
1387                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1388                                                        tx->tx_msg,
1389                                                        IBLND_MSG_SIZE,
1390                                                        DMA_TO_DEVICE);
1391                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1392                                                   tx->tx_msgaddr));
1393                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1394
1395                 list_add(&tx->tx_list, &pool->po_free_list);
1396
1397                 page_offset += IBLND_MSG_SIZE;
1398                 LASSERT(page_offset <= PAGE_SIZE);
1399
1400                 if (page_offset == PAGE_SIZE) {
1401                         page_offset = 0;
1402                         ipage++;
1403                         LASSERT(ipage <= txpgs->ibp_npages);
1404                 }
1405         }
1406 }
1407
1408 struct ib_mr *
1409 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1410                       int negotiated_nfrags)
1411 {
1412         __u16   nfrags = (negotiated_nfrags != -1) ?
1413           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1414
1415         LASSERT(hdev->ibh_mrs != NULL);
1416
1417         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1418             nfrags <= rd->rd_nfrags)
1419                 return NULL;
1420
1421         return hdev->ibh_mrs;
1422 }
1423
1424 static void
1425 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1426 {
1427         LASSERT (pool->fpo_map_count == 0);
1428
1429         if (pool->fpo_fmr_pool != NULL)
1430                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1431
1432         if (pool->fpo_hdev != NULL)
1433                 kiblnd_hdev_decref(pool->fpo_hdev);
1434
1435         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1436 }
1437
1438 static void
1439 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1440 {
1441         kib_fmr_pool_t *pool;
1442
1443         while (!list_empty(head)) {
1444                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1445                 list_del(&pool->fpo_list);
1446                 kiblnd_destroy_fmr_pool(pool);
1447         }
1448 }
1449
1450 static int kiblnd_fmr_pool_size(int ncpts)
1451 {
1452         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1453
1454         return max(IBLND_FMR_POOL, size);
1455 }
1456
1457 static int kiblnd_fmr_flush_trigger(int ncpts)
1458 {
1459         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1460
1461         return max(IBLND_FMR_POOL_FLUSH, size);
1462 }
1463
1464 static int
1465 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1466 {
1467         /* FMR pool for RDMA */
1468         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1469         kib_fmr_pool_t          *fpo;
1470         struct ib_fmr_pool_param param = {
1471                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1472                 .page_shift        = PAGE_SHIFT,
1473                 .access            = (IB_ACCESS_LOCAL_WRITE |
1474                                       IB_ACCESS_REMOTE_WRITE),
1475                 .pool_size         = fps->fps_pool_size,
1476                 .dirty_watermark   = fps->fps_flush_trigger,
1477                 .flush_function    = NULL,
1478                 .flush_arg         = NULL,
1479                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1480         int rc;
1481
1482         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1483         if (fpo == NULL)
1484                 return -ENOMEM;
1485
1486         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1487
1488         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1489         if (IS_ERR(fpo->fpo_fmr_pool)) {
1490                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1491                 CERROR("Failed to create FMR pool: %d\n", rc);
1492
1493                 kiblnd_hdev_decref(fpo->fpo_hdev);
1494                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1495                 return rc;
1496         }
1497
1498         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1499         fpo->fpo_owner    = fps;
1500         *pp_fpo = fpo;
1501
1502         return 0;
1503 }
1504
1505 static void
1506 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1507 {
1508         if (fps->fps_net == NULL) /* intialized? */
1509                 return;
1510
1511         spin_lock(&fps->fps_lock);
1512
1513         while (!list_empty(&fps->fps_pool_list)) {
1514                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1515                                                  kib_fmr_pool_t, fpo_list);
1516                 fpo->fpo_failed = 1;
1517                 list_del(&fpo->fpo_list);
1518                 if (fpo->fpo_map_count == 0)
1519                         list_add(&fpo->fpo_list, zombies);
1520                 else
1521                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1522         }
1523
1524         spin_unlock(&fps->fps_lock);
1525 }
1526
1527 static void
1528 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1529 {
1530         if (fps->fps_net != NULL) { /* initialized? */
1531                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1532                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1533         }
1534 }
1535
1536 static int
1537 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1538                         int pool_size, int flush_trigger)
1539 {
1540         kib_fmr_pool_t *fpo;
1541         int             rc;
1542
1543         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1544
1545         fps->fps_net = net;
1546         fps->fps_cpt = cpt;
1547         fps->fps_pool_size = pool_size;
1548         fps->fps_flush_trigger = flush_trigger;
1549         spin_lock_init(&fps->fps_lock);
1550         INIT_LIST_HEAD(&fps->fps_pool_list);
1551         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1552
1553         rc = kiblnd_create_fmr_pool(fps, &fpo);
1554         if (rc == 0)
1555                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1556
1557         return rc;
1558 }
1559
1560 static int
1561 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1562 {
1563         if (fpo->fpo_map_count != 0) /* still in use */
1564                 return 0;
1565         if (fpo->fpo_failed)
1566                 return 1;
1567         return cfs_time_aftereq(now, fpo->fpo_deadline);
1568 }
1569
1570 void
1571 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1572 {
1573         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1574         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1575         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1576         cfs_time_t         now = cfs_time_current();
1577         kib_fmr_pool_t    *tmp;
1578         int                rc;
1579
1580         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1581         LASSERT(rc == 0);
1582
1583         if (status != 0) {
1584                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1585                 LASSERT(rc == 0);
1586         }
1587
1588         fmr->fmr_pool = NULL;
1589         fmr->fmr_pfmr = NULL;
1590
1591         spin_lock(&fps->fps_lock);
1592         fpo->fpo_map_count--;   /* decref the pool */
1593
1594         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1595                 /* the first pool is persistent */
1596                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1597                         continue;
1598
1599                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1600                         list_move(&fpo->fpo_list, &zombies);
1601                         fps->fps_version++;
1602                 }
1603         }
1604         spin_unlock(&fps->fps_lock);
1605
1606         if (!list_empty(&zombies))
1607                 kiblnd_destroy_fmr_pool_list(&zombies);
1608 }
1609
1610 int
1611 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1612                     __u64 iov, kib_fmr_t *fmr)
1613 {
1614         struct ib_pool_fmr *pfmr;
1615         kib_fmr_pool_t     *fpo;
1616         __u64               version;
1617         int                 rc;
1618
1619 again:
1620         spin_lock(&fps->fps_lock);
1621         version = fps->fps_version;
1622         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1623                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1624                 fpo->fpo_map_count++;
1625                 spin_unlock(&fps->fps_lock);
1626
1627                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1628                                             pages, npages, iov);
1629                 if (likely(!IS_ERR(pfmr))) {
1630                         fmr->fmr_pool = fpo;
1631                         fmr->fmr_pfmr = pfmr;
1632                         return 0;
1633                 }
1634
1635                 spin_lock(&fps->fps_lock);
1636                 fpo->fpo_map_count--;
1637                 if (PTR_ERR(pfmr) != -EAGAIN) {
1638                         spin_unlock(&fps->fps_lock);
1639                         return PTR_ERR(pfmr);
1640                 }
1641
1642                 /* EAGAIN and ... */
1643                 if (version != fps->fps_version) {
1644                         spin_unlock(&fps->fps_lock);
1645                         goto again;
1646                 }
1647         }
1648
1649         if (fps->fps_increasing) {
1650                 spin_unlock(&fps->fps_lock);
1651                 CDEBUG(D_NET, "Another thread is allocating new "
1652                        "FMR pool, waiting for her to complete\n");
1653                 schedule();
1654                 goto again;
1655
1656         }
1657
1658         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1659                 /* someone failed recently */
1660                 spin_unlock(&fps->fps_lock);
1661                 return -EAGAIN;
1662         }
1663
1664         fps->fps_increasing = 1;
1665         spin_unlock(&fps->fps_lock);
1666
1667         CDEBUG(D_NET, "Allocate new FMR pool\n");
1668         rc = kiblnd_create_fmr_pool(fps, &fpo);
1669         spin_lock(&fps->fps_lock);
1670         fps->fps_increasing = 0;
1671         if (rc == 0) {
1672                 fps->fps_version++;
1673                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1674         } else {
1675                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1676         }
1677         spin_unlock(&fps->fps_lock);
1678
1679         goto again;
1680 }
1681
1682 static void
1683 kiblnd_fini_pool(kib_pool_t *pool)
1684 {
1685         LASSERT(list_empty(&pool->po_free_list));
1686         LASSERT(pool->po_allocated == 0);
1687
1688         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1689 }
1690
1691 static void
1692 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1693 {
1694         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1695
1696         memset(pool, 0, sizeof(kib_pool_t));
1697         INIT_LIST_HEAD(&pool->po_free_list);
1698         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1699         pool->po_owner    = ps;
1700         pool->po_size     = size;
1701 }
1702
1703 static void
1704 kiblnd_destroy_pool_list(struct list_head *head)
1705 {
1706         kib_pool_t *pool;
1707
1708         while (!list_empty(head)) {
1709                 pool = list_entry(head->next, kib_pool_t, po_list);
1710                 list_del(&pool->po_list);
1711
1712                 LASSERT(pool->po_owner != NULL);
1713                 pool->po_owner->ps_pool_destroy(pool);
1714         }
1715 }
1716
1717 static void
1718 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1719 {
1720         if (ps->ps_net == NULL) /* intialized? */
1721                 return;
1722
1723         spin_lock(&ps->ps_lock);
1724         while (!list_empty(&ps->ps_pool_list)) {
1725                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1726                                             kib_pool_t, po_list);
1727                 po->po_failed = 1;
1728                 list_del(&po->po_list);
1729                 if (po->po_allocated == 0)
1730                         list_add(&po->po_list, zombies);
1731                 else
1732                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1733         }
1734         spin_unlock(&ps->ps_lock);
1735 }
1736
1737 static void
1738 kiblnd_fini_poolset(kib_poolset_t *ps)
1739 {
1740         if (ps->ps_net != NULL) { /* initialized? */
1741                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1742                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1743         }
1744 }
1745
1746 static int
1747 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1748                     kib_net_t *net, char *name, int size,
1749                     kib_ps_pool_create_t po_create,
1750                     kib_ps_pool_destroy_t po_destroy,
1751                     kib_ps_node_init_t nd_init,
1752                     kib_ps_node_fini_t nd_fini)
1753 {
1754         kib_pool_t      *pool;
1755         int             rc;
1756
1757         memset(ps, 0, sizeof(kib_poolset_t));
1758
1759         ps->ps_cpt          = cpt;
1760         ps->ps_net          = net;
1761         ps->ps_pool_create  = po_create;
1762         ps->ps_pool_destroy = po_destroy;
1763         ps->ps_node_init    = nd_init;
1764         ps->ps_node_fini    = nd_fini;
1765         ps->ps_pool_size    = size;
1766         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1767             >= sizeof(ps->ps_name))
1768                 return -E2BIG;
1769         spin_lock_init(&ps->ps_lock);
1770         INIT_LIST_HEAD(&ps->ps_pool_list);
1771         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1772
1773         rc = ps->ps_pool_create(ps, size, &pool);
1774         if (rc == 0)
1775                 list_add(&pool->po_list, &ps->ps_pool_list);
1776         else
1777                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1778
1779         return rc;
1780 }
1781
1782 static int
1783 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1784 {
1785         if (pool->po_allocated != 0) /* still in use */
1786                 return 0;
1787         if (pool->po_failed)
1788                 return 1;
1789         return cfs_time_aftereq(now, pool->po_deadline);
1790 }
1791
1792 void
1793 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1794 {
1795         struct list_head zombies = LIST_HEAD_INIT(zombies);
1796         kib_poolset_t   *ps = pool->po_owner;
1797         kib_pool_t      *tmp;
1798         cfs_time_t       now = cfs_time_current();
1799
1800         spin_lock(&ps->ps_lock);
1801
1802         if (ps->ps_node_fini != NULL)
1803                 ps->ps_node_fini(pool, node);
1804
1805         LASSERT(pool->po_allocated > 0);
1806         list_add(node, &pool->po_free_list);
1807         pool->po_allocated--;
1808
1809         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1810                 /* the first pool is persistent */
1811                 if (ps->ps_pool_list.next == &pool->po_list)
1812                         continue;
1813
1814                 if (kiblnd_pool_is_idle(pool, now))
1815                         list_move(&pool->po_list, &zombies);
1816         }
1817         spin_unlock(&ps->ps_lock);
1818
1819         if (!list_empty(&zombies))
1820                 kiblnd_destroy_pool_list(&zombies);
1821 }
1822
1823 struct list_head *
1824 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1825 {
1826         struct list_head        *node;
1827         kib_pool_t              *pool;
1828         int                     rc;
1829
1830 again:
1831         spin_lock(&ps->ps_lock);
1832         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1833                 if (list_empty(&pool->po_free_list))
1834                         continue;
1835
1836                 pool->po_allocated++;
1837                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1838                 node = pool->po_free_list.next;
1839                 list_del(node);
1840
1841                 if (ps->ps_node_init != NULL) {
1842                         /* still hold the lock */
1843                         ps->ps_node_init(pool, node);
1844                 }
1845                 spin_unlock(&ps->ps_lock);
1846                 return node;
1847         }
1848
1849         /* no available tx pool and ... */
1850         if (ps->ps_increasing) {
1851                 /* another thread is allocating a new pool */
1852                 spin_unlock(&ps->ps_lock);
1853                 CDEBUG(D_NET, "Another thread is allocating new "
1854                        "%s pool, waiting for her to complete\n",
1855                        ps->ps_name);
1856                 schedule();
1857                 goto again;
1858         }
1859
1860         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1861                 /* someone failed recently */
1862                 spin_unlock(&ps->ps_lock);
1863                 return NULL;
1864         }
1865
1866         ps->ps_increasing = 1;
1867         spin_unlock(&ps->ps_lock);
1868
1869         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1870
1871         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1872
1873         spin_lock(&ps->ps_lock);
1874         ps->ps_increasing = 0;
1875         if (rc == 0) {
1876                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1877         } else {
1878                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1879                 CERROR("Can't allocate new %s pool because out of memory\n",
1880                        ps->ps_name);
1881         }
1882         spin_unlock(&ps->ps_lock);
1883
1884         goto again;
1885 }
1886
1887 static void
1888 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1889 {
1890         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1891         int             i;
1892
1893         LASSERT (pool->po_allocated == 0);
1894
1895         if (tpo->tpo_tx_pages != NULL) {
1896                 kiblnd_unmap_tx_pool(tpo);
1897                 kiblnd_free_pages(tpo->tpo_tx_pages);
1898         }
1899
1900         if (tpo->tpo_tx_descs == NULL)
1901                 goto out;
1902
1903         for (i = 0; i < pool->po_size; i++) {
1904                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1905
1906                 list_del(&tx->tx_list);
1907                 if (tx->tx_pages != NULL)
1908                         LIBCFS_FREE(tx->tx_pages,
1909                                     LNET_MAX_IOV *
1910                                     sizeof(*tx->tx_pages));
1911                 if (tx->tx_frags != NULL)
1912                         LIBCFS_FREE(tx->tx_frags,
1913                                     IBLND_MAX_RDMA_FRAGS *
1914                                             sizeof(*tx->tx_frags));
1915                 if (tx->tx_wrq != NULL)
1916                         LIBCFS_FREE(tx->tx_wrq,
1917                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1918                                     sizeof(*tx->tx_wrq));
1919                 if (tx->tx_sge != NULL)
1920                         LIBCFS_FREE(tx->tx_sge,
1921                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1922                                     sizeof(*tx->tx_sge));
1923                 if (tx->tx_rd != NULL)
1924                         LIBCFS_FREE(tx->tx_rd,
1925                                     offsetof(kib_rdma_desc_t,
1926                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1927         }
1928
1929         LIBCFS_FREE(tpo->tpo_tx_descs,
1930                     pool->po_size * sizeof(kib_tx_t));
1931 out:
1932         kiblnd_fini_pool(pool);
1933         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1934 }
1935
1936 static int kiblnd_tx_pool_size(int ncpts)
1937 {
1938         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1939
1940         return max(IBLND_TX_POOL, ntx);
1941 }
1942
1943 static int
1944 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1945 {
1946         int            i;
1947         int            npg;
1948         kib_pool_t    *pool;
1949         kib_tx_pool_t *tpo;
1950
1951         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1952         if (tpo == NULL) {
1953                 CERROR("Failed to allocate TX pool\n");
1954                 return -ENOMEM;
1955         }
1956
1957         pool = &tpo->tpo_pool;
1958         kiblnd_init_pool(ps, pool, size);
1959         tpo->tpo_tx_descs = NULL;
1960         tpo->tpo_tx_pages = NULL;
1961
1962         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1963         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1964                 CERROR("Can't allocate tx pages: %d\n", npg);
1965                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1966                 return -ENOMEM;
1967         }
1968
1969         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1970                          size * sizeof(kib_tx_t));
1971         if (tpo->tpo_tx_descs == NULL) {
1972                 CERROR("Can't allocate %d tx descriptors\n", size);
1973                 ps->ps_pool_destroy(pool);
1974                 return -ENOMEM;
1975         }
1976
1977         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1978
1979         for (i = 0; i < size; i++) {
1980                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1981
1982                 tx->tx_pool = tpo;
1983                 if (ps->ps_net->ibn_fmr_ps != NULL) {
1984                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1985                                          lnet_cpt_table(), ps->ps_cpt,
1986                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1987                         if (tx->tx_pages == NULL)
1988                                 break;
1989                 }
1990
1991                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1992                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1993                 if (tx->tx_frags == NULL)
1994                         break;
1995
1996                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1997
1998                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1999                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2000                                  sizeof(*tx->tx_wrq));
2001                 if (tx->tx_wrq == NULL)
2002                         break;
2003
2004                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2005                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2006                                  sizeof(*tx->tx_sge));
2007                 if (tx->tx_sge == NULL)
2008                         break;
2009
2010                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2011                                  offsetof(kib_rdma_desc_t,
2012                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2013                 if (tx->tx_rd == NULL)
2014                         break;
2015         }
2016
2017         if (i == size) {
2018                 kiblnd_map_tx_pool(tpo);
2019                 *pp_po = pool;
2020                 return 0;
2021         }
2022
2023         ps->ps_pool_destroy(pool);
2024         return -ENOMEM;
2025 }
2026
2027 static void
2028 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2029 {
2030         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2031                                              tps_poolset);
2032         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2033
2034         tx->tx_cookie = tps->tps_next_tx_cookie++;
2035 }
2036
2037 static void
2038 kiblnd_net_fini_pools(kib_net_t *net)
2039 {
2040         int     i;
2041
2042         cfs_cpt_for_each(i, lnet_cpt_table()) {
2043                 kib_tx_poolset_t        *tps;
2044                 kib_fmr_poolset_t       *fps;
2045
2046                 if (net->ibn_tx_ps != NULL) {
2047                         tps = net->ibn_tx_ps[i];
2048                         kiblnd_fini_poolset(&tps->tps_poolset);
2049                 }
2050
2051                 if (net->ibn_fmr_ps != NULL) {
2052                         fps = net->ibn_fmr_ps[i];
2053                         kiblnd_fini_fmr_poolset(fps);
2054                 }
2055         }
2056
2057         if (net->ibn_tx_ps != NULL) {
2058                 cfs_percpt_free(net->ibn_tx_ps);
2059                 net->ibn_tx_ps = NULL;
2060         }
2061
2062         if (net->ibn_fmr_ps != NULL) {
2063                 cfs_percpt_free(net->ibn_fmr_ps);
2064                 net->ibn_fmr_ps = NULL;
2065         }
2066 }
2067
2068 static int
2069 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2070 {
2071         unsigned long   flags;
2072         int             cpt;
2073         int             rc;
2074         int             i;
2075
2076         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2077         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2078                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2079                                            flags);
2080                 goto create_tx_pool;
2081         }
2082
2083         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2084
2085         if (*kiblnd_tunables.kib_fmr_pool_size <
2086             *kiblnd_tunables.kib_ntx / 4) {
2087                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2088                        *kiblnd_tunables.kib_fmr_pool_size,
2089                        *kiblnd_tunables.kib_ntx / 4);
2090                 rc = -EINVAL;
2091                 goto failed;
2092         }
2093
2094         /* TX pool must be created later than FMR, see LU-2268
2095          * for details */
2096         LASSERT(net->ibn_tx_ps == NULL);
2097
2098         /* premapping can fail if ibd_nmr > 1, so we always create
2099          * FMR pool and map-on-demand if premapping failed */
2100
2101         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2102                                            sizeof(kib_fmr_poolset_t));
2103         if (net->ibn_fmr_ps == NULL) {
2104                 CERROR("Failed to allocate FMR pool array\n");
2105                 rc = -ENOMEM;
2106                 goto failed;
2107         }
2108
2109         for (i = 0; i < ncpts; i++) {
2110                 cpt = (cpts == NULL) ? i : cpts[i];
2111                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2112                                              kiblnd_fmr_pool_size(ncpts),
2113                                              kiblnd_fmr_flush_trigger(ncpts));
2114                 if (rc != 0) {
2115                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2116                                cpt, rc);
2117                         goto failed;
2118                 }
2119         }
2120
2121         if (i > 0)
2122                 LASSERT(i == ncpts);
2123
2124  create_tx_pool:
2125         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2126                                           sizeof(kib_tx_poolset_t));
2127         if (net->ibn_tx_ps == NULL) {
2128                 CERROR("Failed to allocate tx pool array\n");
2129                 rc = -ENOMEM;
2130                 goto failed;
2131         }
2132
2133         for (i = 0; i < ncpts; i++) {
2134                 cpt = (cpts == NULL) ? i : cpts[i];
2135                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2136                                          cpt, net, "TX",
2137                                          kiblnd_tx_pool_size(ncpts),
2138                                          kiblnd_create_tx_pool,
2139                                          kiblnd_destroy_tx_pool,
2140                                          kiblnd_tx_init, NULL);
2141                 if (rc != 0) {
2142                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2143                                cpt, rc);
2144                         goto failed;
2145                 }
2146         }
2147
2148         return 0;
2149  failed:
2150         kiblnd_net_fini_pools(net);
2151         LASSERT(rc != 0);
2152         return rc;
2153 }
2154
2155 static int
2156 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2157 {
2158         struct ib_device_attr *attr;
2159         int                    rc;
2160
2161         /* It's safe to assume a HCA can handle a page size
2162          * matching that of the native system */
2163         hdev->ibh_page_shift = PAGE_SHIFT;
2164         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2165         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2166
2167         LIBCFS_ALLOC(attr, sizeof(*attr));
2168         if (attr == NULL) {
2169                 CERROR("Out of memory\n");
2170                 return -ENOMEM;
2171         }
2172
2173         rc = ib_query_device(hdev->ibh_ibdev, attr);
2174         if (rc == 0)
2175                 hdev->ibh_mr_size = attr->max_mr_size;
2176
2177         LIBCFS_FREE(attr, sizeof(*attr));
2178
2179         if (rc != 0) {
2180                 CERROR("Failed to query IB device: %d\n", rc);
2181                 return rc;
2182         }
2183
2184         if (hdev->ibh_mr_size == ~0ULL) {
2185                 hdev->ibh_mr_shift = 64;
2186                 return 0;
2187         }
2188
2189         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2190         return -EINVAL;
2191 }
2192
2193 static void
2194 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2195 {
2196         if (hdev->ibh_mrs == NULL)
2197                 return;
2198
2199         ib_dereg_mr(hdev->ibh_mrs);
2200
2201         hdev->ibh_mrs = NULL;
2202 }
2203
2204 void
2205 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2206 {
2207         kiblnd_hdev_cleanup_mrs(hdev);
2208
2209         if (hdev->ibh_pd != NULL)
2210                 ib_dealloc_pd(hdev->ibh_pd);
2211
2212         if (hdev->ibh_cmid != NULL)
2213                 rdma_destroy_id(hdev->ibh_cmid);
2214
2215         LIBCFS_FREE(hdev, sizeof(*hdev));
2216 }
2217
2218 static int
2219 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2220 {
2221         struct ib_mr *mr;
2222         int           rc;
2223         int           acflags = IB_ACCESS_LOCAL_WRITE |
2224                                 IB_ACCESS_REMOTE_WRITE;
2225
2226         rc = kiblnd_hdev_get_attr(hdev);
2227         if (rc != 0)
2228                 return rc;
2229
2230         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2231         if (IS_ERR(mr)) {
2232                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2233                 kiblnd_hdev_cleanup_mrs(hdev);
2234                 return PTR_ERR(mr);
2235         }
2236
2237         hdev->ibh_mrs = mr;
2238
2239         return 0;
2240 }
2241
2242 static int
2243 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2244 {       /* DUMMY */
2245         return 0;
2246 }
2247
2248 static int
2249 kiblnd_dev_need_failover(kib_dev_t *dev)
2250 {
2251         struct rdma_cm_id  *cmid;
2252         struct sockaddr_in  srcaddr;
2253         struct sockaddr_in  dstaddr;
2254         int                 rc;
2255
2256         if (dev->ibd_hdev == NULL || /* initializing */
2257             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2258             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2259                 return 1;
2260
2261         /* XXX: it's UGLY, but I don't have better way to find
2262          * ib-bonding HCA failover because:
2263          *
2264          * a. no reliable CM event for HCA failover...
2265          * b. no OFED API to get ib_device for current net_device...
2266          *
2267          * We have only two choices at this point:
2268          *
2269          * a. rdma_bind_addr(), it will conflict with listener cmid
2270          * b. rdma_resolve_addr() to zero addr */
2271         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2272                                      IB_QPT_RC);
2273         if (IS_ERR(cmid)) {
2274                 rc = PTR_ERR(cmid);
2275                 CERROR("Failed to create cmid for failover: %d\n", rc);
2276                 return rc;
2277         }
2278
2279         memset(&srcaddr, 0, sizeof(srcaddr));
2280         srcaddr.sin_family      = AF_INET;
2281         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2282
2283         memset(&dstaddr, 0, sizeof(dstaddr));
2284         dstaddr.sin_family = AF_INET;
2285         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2286                                (struct sockaddr *)&dstaddr, 1);
2287         if (rc != 0 || cmid->device == NULL) {
2288                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2289                        dev->ibd_ifname, &dev->ibd_ifip,
2290                        cmid->device, rc);
2291                 rdma_destroy_id(cmid);
2292                 return rc;
2293         }
2294
2295         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2296         rdma_destroy_id(cmid);
2297         return rc;
2298 }
2299
2300 int
2301 kiblnd_dev_failover(kib_dev_t *dev)
2302 {
2303         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2304         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2305         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2306         struct rdma_cm_id  *cmid  = NULL;
2307         kib_hca_dev_t      *hdev  = NULL;
2308         kib_hca_dev_t      *old;
2309         struct ib_pd       *pd;
2310         kib_net_t          *net;
2311         struct sockaddr_in  addr;
2312         unsigned long       flags;
2313         int                 rc = 0;
2314         int                 i;
2315
2316         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2317                  dev->ibd_can_failover ||
2318                  dev->ibd_hdev == NULL);
2319
2320         rc = kiblnd_dev_need_failover(dev);
2321         if (rc <= 0)
2322                 goto out;
2323
2324         if (dev->ibd_hdev != NULL &&
2325             dev->ibd_hdev->ibh_cmid != NULL) {
2326                 /* XXX it's not good to close old listener at here,
2327                  * because we can fail to create new listener.
2328                  * But we have to close it now, otherwise rdma_bind_addr
2329                  * will return EADDRINUSE... How crap! */
2330                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2331
2332                 cmid = dev->ibd_hdev->ibh_cmid;
2333                 /* make next schedule of kiblnd_dev_need_failover()
2334                  * return 1 for me */
2335                 dev->ibd_hdev->ibh_cmid  = NULL;
2336                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2337
2338                 rdma_destroy_id(cmid);
2339         }
2340
2341         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2342                                      IB_QPT_RC);
2343         if (IS_ERR(cmid)) {
2344                 rc = PTR_ERR(cmid);
2345                 CERROR("Failed to create cmid for failover: %d\n", rc);
2346                 goto out;
2347         }
2348
2349         memset(&addr, 0, sizeof(addr));
2350         addr.sin_family      = AF_INET;
2351         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2352         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2353
2354         /* Bind to failover device or port */
2355         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2356         if (rc != 0 || cmid->device == NULL) {
2357                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2358                        dev->ibd_ifname, &dev->ibd_ifip,
2359                        cmid->device, rc);
2360                 rdma_destroy_id(cmid);
2361                 goto out;
2362         }
2363
2364         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2365         if (hdev == NULL) {
2366                 CERROR("Failed to allocate kib_hca_dev\n");
2367                 rdma_destroy_id(cmid);
2368                 rc = -ENOMEM;
2369                 goto out;
2370         }
2371
2372         atomic_set(&hdev->ibh_ref, 1);
2373         hdev->ibh_dev   = dev;
2374         hdev->ibh_cmid  = cmid;
2375         hdev->ibh_ibdev = cmid->device;
2376
2377         pd = ib_alloc_pd(cmid->device);
2378         if (IS_ERR(pd)) {
2379                 rc = PTR_ERR(pd);
2380                 CERROR("Can't allocate PD: %d\n", rc);
2381                 goto out;
2382         }
2383
2384         hdev->ibh_pd = pd;
2385
2386         rc = rdma_listen(cmid, 0);
2387         if (rc != 0) {
2388                 CERROR("Can't start new listener: %d\n", rc);
2389                 goto out;
2390         }
2391
2392         rc = kiblnd_hdev_setup_mrs(hdev);
2393         if (rc != 0) {
2394                 CERROR("Can't setup device: %d\n", rc);
2395                 goto out;
2396         }
2397
2398         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2399
2400         old = dev->ibd_hdev;
2401         dev->ibd_hdev = hdev;   /* take over the refcount */
2402         hdev = old;
2403
2404         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2405                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2406                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2407                                             &zombie_tpo);
2408
2409                         if (net->ibn_fmr_ps != NULL)
2410                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2411                                                         &zombie_fpo);
2412                 }
2413         }
2414
2415         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2416  out:
2417         if (!list_empty(&zombie_tpo))
2418                 kiblnd_destroy_pool_list(&zombie_tpo);
2419         if (!list_empty(&zombie_ppo))
2420                 kiblnd_destroy_pool_list(&zombie_ppo);
2421         if (!list_empty(&zombie_fpo))
2422                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2423         if (hdev != NULL)
2424                 kiblnd_hdev_decref(hdev);
2425
2426         if (rc != 0)
2427                 dev->ibd_failed_failover++;
2428         else
2429                 dev->ibd_failed_failover = 0;
2430
2431         return rc;
2432 }
2433
2434 void
2435 kiblnd_destroy_dev (kib_dev_t *dev)
2436 {
2437         LASSERT (dev->ibd_nnets == 0);
2438         LASSERT(list_empty(&dev->ibd_nets));
2439
2440         list_del(&dev->ibd_fail_list);
2441         list_del(&dev->ibd_list);
2442
2443         if (dev->ibd_hdev != NULL)
2444                 kiblnd_hdev_decref(dev->ibd_hdev);
2445
2446         LIBCFS_FREE(dev, sizeof(*dev));
2447 }
2448
2449 static kib_dev_t *
2450 kiblnd_create_dev(char *ifname)
2451 {
2452         struct net_device *netdev;
2453         kib_dev_t         *dev;
2454         __u32              netmask;
2455         __u32              ip;
2456         int                up;
2457         int                rc;
2458
2459         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2460         if (rc != 0) {
2461                 CERROR("Can't query IPoIB interface %s: %d\n",
2462                        ifname, rc);
2463                 return NULL;
2464         }
2465
2466         if (!up) {
2467                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2468                 return NULL;
2469         }
2470
2471         LIBCFS_ALLOC(dev, sizeof(*dev));
2472         if (dev == NULL)
2473                 return NULL;
2474
2475         memset(dev, 0, sizeof(*dev));
2476         netdev = dev_get_by_name(&init_net, ifname);
2477         if (netdev == NULL) {
2478                 dev->ibd_can_failover = 0;
2479         } else {
2480                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2481                 dev_put(netdev);
2482         }
2483
2484         INIT_LIST_HEAD(&dev->ibd_nets);
2485         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2486         INIT_LIST_HEAD(&dev->ibd_fail_list);
2487         dev->ibd_ifip = ip;
2488         strcpy(&dev->ibd_ifname[0], ifname);
2489
2490         /* initialize the device */
2491         rc = kiblnd_dev_failover(dev);
2492         if (rc != 0) {
2493                 CERROR("Can't initialize device: %d\n", rc);
2494                 LIBCFS_FREE(dev, sizeof(*dev));
2495                 return NULL;
2496         }
2497
2498         list_add_tail(&dev->ibd_list,
2499                           &kiblnd_data.kib_devs);
2500         return dev;
2501 }
2502
2503 static void
2504 kiblnd_base_shutdown(void)
2505 {
2506         struct kib_sched_info   *sched;
2507         int                     i;
2508
2509         LASSERT(list_empty(&kiblnd_data.kib_devs));
2510
2511         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2512                atomic_read(&libcfs_kmemory));
2513
2514         switch (kiblnd_data.kib_init) {
2515         default:
2516                 LBUG();
2517
2518         case IBLND_INIT_ALL:
2519         case IBLND_INIT_DATA:
2520                 LASSERT (kiblnd_data.kib_peers != NULL);
2521                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2522                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2523                 }
2524                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2525                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2526
2527                 /* flag threads to terminate; wake and wait for them to die */
2528                 kiblnd_data.kib_shutdown = 1;
2529
2530                 /* NB: we really want to stop scheduler threads net by net
2531                  * instead of the whole module, this should be improved
2532                  * with dynamic configuration LNet */
2533                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2534                         wake_up_all(&sched->ibs_waitq);
2535
2536                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2537                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2538
2539                 i = 2;
2540                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2541                         i++;
2542                         /* power of 2? */
2543                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2544                                "Waiting for %d threads to terminate\n",
2545                                atomic_read(&kiblnd_data.kib_nthreads));
2546                         set_current_state(TASK_UNINTERRUPTIBLE);
2547                         schedule_timeout(cfs_time_seconds(1));
2548                 }
2549
2550                 /* fall through */
2551
2552         case IBLND_INIT_NOTHING:
2553                 break;
2554         }
2555
2556         if (kiblnd_data.kib_peers != NULL) {
2557                 LIBCFS_FREE(kiblnd_data.kib_peers,
2558                             sizeof(struct list_head) *
2559                             kiblnd_data.kib_peer_hash_size);
2560         }
2561
2562         if (kiblnd_data.kib_scheds != NULL)
2563                 cfs_percpt_free(kiblnd_data.kib_scheds);
2564
2565         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2566                atomic_read(&libcfs_kmemory));
2567
2568         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2569         module_put(THIS_MODULE);
2570 }
2571
2572 static void
2573 kiblnd_shutdown (lnet_ni_t *ni)
2574 {
2575         kib_net_t        *net = ni->ni_data;
2576         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2577         int               i;
2578         unsigned long     flags;
2579
2580         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2581
2582         if (net == NULL)
2583                 goto out;
2584
2585         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2586                atomic_read(&libcfs_kmemory));
2587
2588         write_lock_irqsave(g_lock, flags);
2589         net->ibn_shutdown = 1;
2590         write_unlock_irqrestore(g_lock, flags);
2591
2592         switch (net->ibn_init) {
2593         default:
2594                 LBUG();
2595
2596         case IBLND_INIT_ALL:
2597                 /* nuke all existing peers within this net */
2598                 kiblnd_del_peer(ni, LNET_NID_ANY);
2599
2600                 /* Wait for all peer state to clean up */
2601                 i = 2;
2602                 while (atomic_read(&net->ibn_npeers) != 0) {
2603                         i++;
2604                         /* power of 2? */
2605                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2606                                "%s: waiting for %d peers to disconnect\n",
2607                                libcfs_nid2str(ni->ni_nid),
2608                                atomic_read(&net->ibn_npeers));
2609                         set_current_state(TASK_UNINTERRUPTIBLE);
2610                         schedule_timeout(cfs_time_seconds(1));
2611                 }
2612
2613                 kiblnd_net_fini_pools(net);
2614
2615                 write_lock_irqsave(g_lock, flags);
2616                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2617                 net->ibn_dev->ibd_nnets--;
2618                 list_del(&net->ibn_list);
2619                 write_unlock_irqrestore(g_lock, flags);
2620
2621                 /* fall through */
2622
2623         case IBLND_INIT_NOTHING:
2624                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2625
2626                 if (net->ibn_dev != NULL &&
2627                     net->ibn_dev->ibd_nnets == 0)
2628                         kiblnd_destroy_dev(net->ibn_dev);
2629
2630                 break;
2631         }
2632
2633         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2634                atomic_read(&libcfs_kmemory));
2635
2636         net->ibn_init = IBLND_INIT_NOTHING;
2637         ni->ni_data = NULL;
2638
2639         LIBCFS_FREE(net, sizeof(*net));
2640
2641 out:
2642         if (list_empty(&kiblnd_data.kib_devs))
2643                 kiblnd_base_shutdown();
2644         return;
2645 }
2646
2647 static int
2648 kiblnd_base_startup(void)
2649 {
2650         struct kib_sched_info   *sched;
2651         int                     rc;
2652         int                     i;
2653
2654         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2655
2656         try_module_get(THIS_MODULE);
2657         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2658
2659         rwlock_init(&kiblnd_data.kib_global_lock);
2660
2661         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2662         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2663
2664         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2665         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2666                      sizeof(struct list_head) *
2667                      kiblnd_data.kib_peer_hash_size);
2668         if (kiblnd_data.kib_peers == NULL)
2669                 goto failed;
2670
2671         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2672                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2673
2674         spin_lock_init(&kiblnd_data.kib_connd_lock);
2675         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2676         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2677         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2678         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2679
2680         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2681                                                   sizeof(*sched));
2682         if (kiblnd_data.kib_scheds == NULL)
2683                 goto failed;
2684
2685         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2686                 int     nthrs;
2687
2688                 spin_lock_init(&sched->ibs_lock);
2689                 INIT_LIST_HEAD(&sched->ibs_conns);
2690                 init_waitqueue_head(&sched->ibs_waitq);
2691
2692                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2693                 if (*kiblnd_tunables.kib_nscheds > 0) {
2694                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2695                 } else {
2696                         /* max to half of CPUs, another half is reserved for
2697                          * upper layer modules */
2698                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2699                 }
2700
2701                 sched->ibs_nthreads_max = nthrs;
2702                 sched->ibs_cpt = i;
2703         }
2704
2705         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2706
2707         /* lists/ptrs/locks initialised */
2708         kiblnd_data.kib_init = IBLND_INIT_DATA;
2709         /*****************************************************/
2710
2711         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2712         if (rc != 0) {
2713                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2714                 goto failed;
2715         }
2716
2717         if (*kiblnd_tunables.kib_dev_failover != 0)
2718                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2719                                          "kiblnd_failover");
2720
2721         if (rc != 0) {
2722                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2723                 goto failed;
2724         }
2725
2726         /* flag everything initialised */
2727         kiblnd_data.kib_init = IBLND_INIT_ALL;
2728         /*****************************************************/
2729
2730         return 0;
2731
2732  failed:
2733         kiblnd_base_shutdown();
2734         return -ENETDOWN;
2735 }
2736
2737 static int
2738 kiblnd_start_schedulers(struct kib_sched_info *sched)
2739 {
2740         int     rc = 0;
2741         int     nthrs;
2742         int     i;
2743
2744         if (sched->ibs_nthreads == 0) {
2745                 if (*kiblnd_tunables.kib_nscheds > 0) {
2746                         nthrs = sched->ibs_nthreads_max;
2747                 } else {
2748                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2749                                                sched->ibs_cpt);
2750                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2751                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2752                 }
2753         } else {
2754                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2755                 /* increase one thread if there is new interface */
2756                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2757         }
2758
2759         for (i = 0; i < nthrs; i++) {
2760                 long    id;
2761                 char    name[20];
2762                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2763                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2764                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2765                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2766                 if (rc == 0)
2767                         continue;
2768
2769                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2770                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2771                 break;
2772         }
2773
2774         sched->ibs_nthreads += i;
2775         return rc;
2776 }
2777
2778 static int
2779 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2780 {
2781         int     cpt;
2782         int     rc;
2783         int     i;
2784
2785         for (i = 0; i < ncpts; i++) {
2786                 struct kib_sched_info *sched;
2787
2788                 cpt = (cpts == NULL) ? i : cpts[i];
2789                 sched = kiblnd_data.kib_scheds[cpt];
2790
2791                 if (!newdev && sched->ibs_nthreads > 0)
2792                         continue;
2793
2794                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2795                 if (rc != 0) {
2796                         CERROR("Failed to start scheduler threads for %s\n",
2797                                dev->ibd_ifname);
2798                         return rc;
2799                 }
2800         }
2801         return 0;
2802 }
2803
2804 static kib_dev_t *
2805 kiblnd_dev_search(char *ifname)
2806 {
2807         kib_dev_t       *alias = NULL;
2808         kib_dev_t       *dev;
2809         char            *colon;
2810         char            *colon2;
2811
2812         colon = strchr(ifname, ':');
2813         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2814                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2815                         return dev;
2816
2817                 if (alias != NULL)
2818                         continue;
2819
2820                 colon2 = strchr(dev->ibd_ifname, ':');
2821                 if (colon != NULL)
2822                         *colon = 0;
2823                 if (colon2 != NULL)
2824                         *colon2 = 0;
2825
2826                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2827                         alias = dev;
2828
2829                 if (colon != NULL)
2830                         *colon = ':';
2831                 if (colon2 != NULL)
2832                         *colon2 = ':';
2833         }
2834         return alias;
2835 }
2836
2837 static int
2838 kiblnd_startup (lnet_ni_t *ni)
2839 {
2840         char                     *ifname;
2841         kib_dev_t                *ibdev = NULL;
2842         kib_net_t                *net;
2843         struct timeval            tv;
2844         unsigned long             flags;
2845         int                       rc;
2846         int                       newdev;
2847
2848         LASSERT (ni->ni_lnd == &the_o2iblnd);
2849
2850         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2851                 rc = kiblnd_base_startup();
2852                 if (rc != 0)
2853                         return rc;
2854         }
2855
2856         LIBCFS_ALLOC(net, sizeof(*net));
2857         ni->ni_data = net;
2858         if (net == NULL)
2859                 goto failed;
2860
2861         memset(net, 0, sizeof(*net));
2862
2863         do_gettimeofday(&tv);
2864         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2865
2866         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2867         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2868         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2869         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2870
2871         if (ni->ni_interfaces[0] != NULL) {
2872                 /* Use the IPoIB interface specified in 'networks=' */
2873
2874                 CLASSERT (LNET_MAX_INTERFACES > 1);
2875                 if (ni->ni_interfaces[1] != NULL) {
2876                         CERROR("Multiple interfaces not supported\n");
2877                         goto failed;
2878                 }
2879
2880                 ifname = ni->ni_interfaces[0];
2881         } else {
2882                 ifname = *kiblnd_tunables.kib_default_ipif;
2883         }
2884
2885         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2886                 CERROR("IPoIB interface name too long: %s\n", ifname);
2887                 goto failed;
2888         }
2889
2890         ibdev = kiblnd_dev_search(ifname);
2891
2892         newdev = ibdev == NULL;
2893         /* hmm...create kib_dev even for alias */
2894         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
2895                 ibdev = kiblnd_create_dev(ifname);
2896
2897         if (ibdev == NULL)
2898                 goto failed;
2899
2900         net->ibn_dev = ibdev;
2901         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2902
2903         rc = kiblnd_dev_start_threads(ibdev, newdev,
2904                                       ni->ni_cpts, ni->ni_ncpts);
2905         if (rc != 0)
2906                 goto failed;
2907
2908         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2909         if (rc != 0) {
2910                 CERROR("Failed to initialize NI pools: %d\n", rc);
2911                 goto failed;
2912         }
2913
2914         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2915         ibdev->ibd_nnets++;
2916         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2917         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2918
2919         net->ibn_init = IBLND_INIT_ALL;
2920
2921         return 0;
2922
2923 failed:
2924         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
2925                 kiblnd_destroy_dev(ibdev);
2926
2927         kiblnd_shutdown(ni);
2928
2929         CDEBUG(D_NET, "kiblnd_startup failed\n");
2930         return -ENETDOWN;
2931 }
2932
2933 static lnd_t the_o2iblnd = {
2934         .lnd_type       = O2IBLND,
2935         .lnd_startup    = kiblnd_startup,
2936         .lnd_shutdown   = kiblnd_shutdown,
2937         .lnd_ctl        = kiblnd_ctl,
2938         .lnd_query      = kiblnd_query,
2939         .lnd_send       = kiblnd_send,
2940         .lnd_recv       = kiblnd_recv,
2941 };
2942
2943 static void __exit
2944 kiblnd_module_fini (void)
2945 {
2946         lnet_unregister_lnd(&the_o2iblnd);
2947         kiblnd_tunables_fini();
2948 }
2949
2950 static int __init
2951 kiblnd_module_init (void)
2952 {
2953         int    rc;
2954
2955         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2956         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2957                   <= IBLND_MSG_SIZE);
2958         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2959                   <= IBLND_MSG_SIZE);
2960
2961         rc = kiblnd_tunables_init();
2962         if (rc != 0)
2963                 return rc;
2964
2965         lnet_register_lnd(&the_o2iblnd);
2966
2967         return 0;
2968 }
2969
2970 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2971 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2972 MODULE_LICENSE("GPL");
2973
2974 module_init(kiblnd_module_init);
2975 module_exit(kiblnd_module_fini);