lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 kiblnd_cksum (void *ptr, int nob)
  50 {
  51         char  *c  = ptr;
  52         __u32  sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return (sum == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 340         peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
 341         peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
 342         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 343
 344         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 345         INIT_LIST_HEAD(&peer->ibp_conns);
 346         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 347
 348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 349
 350         /* always called with a ref on ni, which prevents ni being shutdown */
 351         LASSERT(net->ibn_shutdown == 0);
 352
 353         /* npeers only grows with the global lock held */
 354         atomic_inc(&net->ibn_npeers);
 355
 356         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 357
 358         *peerp = peer;
 359         return 0;
 360 }
 361
 362 void
 363 kiblnd_destroy_peer (kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(peer->ibp_connecting == 0);
 371         LASSERT(peer->ibp_accepting == 0);
 372         LASSERT(list_empty(&peer->ibp_conns));
 373         LASSERT(list_empty(&peer->ibp_tx_queue));
 374
 375         LIBCFS_FREE(peer, sizeof(*peer));
 376
 377         /* NB a peer's connections keep a reference on their peer until
 378          * they are destroyed, so we can be assured that _all_ state to do
 379          * with this peer has been cleaned up when its refcount drops to
 380          * zero. */
 381         atomic_dec(&net->ibn_npeers);
 382 }
 383
 384 kib_peer_t *
 385 kiblnd_find_peer_locked (lnet_nid_t nid)
 386 {
 387         /* the caller is responsible for accounting the additional reference
 388          * that this creates */
 389         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 390         struct list_head        *tmp;
 391         kib_peer_t              *peer;
 392
 393         list_for_each(tmp, peer_list) {
 394
 395                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 396
 397                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 398                         peer->ibp_accepting > 0 ||
 399                         !list_empty(&peer->ibp_conns)); /* active conn */
 400
 401                 if (peer->ibp_nid != nid)
 402                         continue;
 403
 404                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 405                        peer, libcfs_nid2str(nid),
 406                        atomic_read(&peer->ibp_refcount),
 407                        peer->ibp_version);
 408                 return peer;
 409         }
 410         return NULL;
 411 }
 412
 413 void
 414 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 415 {
 416         LASSERT(list_empty(&peer->ibp_conns));
 417
 418         LASSERT (kiblnd_peer_active(peer));
 419         list_del_init(&peer->ibp_list);
 420         /* lose peerlist's ref */
 421         kiblnd_peer_decref(peer);
 422 }
 423
 424 static int
 425 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 426                      lnet_nid_t *nidp, int *count)
 427 {
 428         kib_peer_t              *peer;
 429         struct list_head        *ptmp;
 430         int                      i;
 431         unsigned long            flags;
 432
 433         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 434
 435         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 436
 437                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 438
 439                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 440                         LASSERT(peer->ibp_connecting > 0 ||
 441                                 peer->ibp_accepting > 0 ||
 442                                 !list_empty(&peer->ibp_conns));
 443
 444                         if (peer->ibp_ni != ni)
 445                                 continue;
 446
 447                         if (index-- > 0)
 448                                 continue;
 449
 450                         *nidp = peer->ibp_nid;
 451                         *count = atomic_read(&peer->ibp_refcount);
 452
 453                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 454                                                flags);
 455                         return 0;
 456                 }
 457         }
 458
 459         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 460         return -ENOENT;
 461 }
 462
 463 static void
 464 kiblnd_del_peer_locked (kib_peer_t *peer)
 465 {
 466         struct list_head        *ctmp;
 467         struct list_head        *cnxt;
 468         kib_conn_t              *conn;
 469
 470         if (list_empty(&peer->ibp_conns)) {
 471                 kiblnd_unlink_peer_locked(peer);
 472         } else {
 473                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 474                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 475
 476                         kiblnd_close_conn_locked(conn, 0);
 477                 }
 478                 /* NB closing peer's last conn unlinked it. */
 479         }
 480         /* NB peer now unlinked; might even be freed if the peer table had the
 481          * last ref on it. */
 482 }
 483
 484 static int
 485 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 486 {
 487         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 488         struct list_head        *ptmp;
 489         struct list_head        *pnxt;
 490         kib_peer_t              *peer;
 491         int                     lo;
 492         int                     hi;
 493         int                     i;
 494         unsigned long           flags;
 495         int                     rc = -ENOENT;
 496
 497         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 498
 499         if (nid != LNET_NID_ANY) {
 500                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 501         } else {
 502                 lo = 0;
 503                 hi = kiblnd_data.kib_peer_hash_size - 1;
 504         }
 505
 506         for (i = lo; i <= hi; i++) {
 507                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 508                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 509                         LASSERT(peer->ibp_connecting > 0 ||
 510                                 peer->ibp_accepting > 0 ||
 511                                 !list_empty(&peer->ibp_conns));
 512
 513                         if (peer->ibp_ni != ni)
 514                                 continue;
 515
 516                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 517                                 continue;
 518
 519                         if (!list_empty(&peer->ibp_tx_queue)) {
 520                                 LASSERT(list_empty(&peer->ibp_conns));
 521
 522                                 list_splice_init(&peer->ibp_tx_queue,
 523                                                  &zombies);
 524                         }
 525
 526                         kiblnd_del_peer_locked(peer);
 527                         rc = 0;         /* matched something */
 528                 }
 529         }
 530
 531         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 532
 533         kiblnd_txlist_done(ni, &zombies, -EIO);
 534
 535         return rc;
 536 }
 537
 538 static kib_conn_t *
 539 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 540 {
 541         kib_peer_t              *peer;
 542         struct list_head        *ptmp;
 543         kib_conn_t              *conn;
 544         struct list_head        *ctmp;
 545         int                     i;
 546         unsigned long           flags;
 547
 548         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 549
 550         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 551                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 552
 553                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 554                         LASSERT(peer->ibp_connecting > 0 ||
 555                                 peer->ibp_accepting > 0 ||
 556                                 !list_empty(&peer->ibp_conns));
 557
 558                         if (peer->ibp_ni != ni)
 559                                 continue;
 560
 561                         list_for_each(ctmp, &peer->ibp_conns) {
 562                                 if (index-- > 0)
 563                                         continue;
 564
 565                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 566                                 kiblnd_conn_addref(conn);
 567                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 568                                                        flags);
 569                                 return conn;
 570                         }
 571                 }
 572         }
 573
 574         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 575         return NULL;
 576 }
 577
 578 static void
 579 kiblnd_debug_rx (kib_rx_t *rx)
 580 {
 581         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 582                rx, rx->rx_status, rx->rx_msg->ibm_type,
 583                rx->rx_msg->ibm_credits);
 584 }
 585
 586 static void
 587 kiblnd_debug_tx (kib_tx_t *tx)
 588 {
 589         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 590                "cookie "LPX64" msg %s%s type %x cred %d\n",
 591                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 592                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 593                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 594                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 595                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 596 }
 597
 598 void
 599 kiblnd_debug_conn (kib_conn_t *conn)
 600 {
 601         struct list_head        *tmp;
 602         int                     i;
 603
 604         spin_lock(&conn->ibc_lock);
 605
 606         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 607                atomic_read(&conn->ibc_refcount), conn,
 608                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 609         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 610                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 611                conn->ibc_nsends_posted, conn->ibc_credits,
 612                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 613         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 614
 615         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 616         list_for_each(tmp, &conn->ibc_early_rxs)
 617                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 620         list_for_each(tmp, &conn->ibc_tx_noops)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 628         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 632         list_for_each(tmp, &conn->ibc_tx_queue)
 633                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 634
 635         CDEBUG(D_CONSOLE, "   active_txs:\n");
 636         list_for_each(tmp, &conn->ibc_active_txs)
 637                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 638
 639         CDEBUG(D_CONSOLE, "   rxs:\n");
 640         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 641                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 642
 643         spin_unlock(&conn->ibc_lock);
 644 }
 645
 646 int
 647 kiblnd_translate_mtu(int value)
 648 {
 649         switch (value) {
 650         default:
 651                 return -1;
 652         case 0:
 653                 return 0;
 654         case 256:
 655                 return IB_MTU_256;
 656         case 512:
 657                 return IB_MTU_512;
 658         case 1024:
 659                 return IB_MTU_1024;
 660         case 2048:
 661                 return IB_MTU_2048;
 662         case 4096:
 663                 return IB_MTU_4096;
 664         }
 665 }
 666
 667 static void
 668 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 669 {
 670         int           mtu;
 671
 672         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 673         if (cmid->route.path_rec == NULL)
 674                 return;
 675
 676         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 677         LASSERT (mtu >= 0);
 678         if (mtu != 0)
 679                 cmid->route.path_rec->mtu = mtu;
 680 }
 681
 682 static int
 683 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 684 {
 685         cpumask_t       *mask;
 686         int             vectors;
 687         int             off;
 688         int             i;
 689
 690         vectors = conn->ibc_cmid->device->num_comp_vectors;
 691         if (vectors <= 1)
 692                 return 0;
 693
 694         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 695
 696         /* hash NID to CPU id in this partition... */
 697         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 698         for_each_cpu(i, mask) {
 699                 if (off-- == 0)
 700                         return i % vectors;
 701         }
 702
 703         LBUG();
 704         return 1;
 705 }
 706
 707 kib_conn_t *
 708 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 709                    int state, int version)
 710 {
 711         /* CAVEAT EMPTOR:
 712          * If the new conn is created successfully it takes over the caller's
 713          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 714          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 715          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 716          * to destroy 'cmid' here since I'm called from the CM which still has
 717          * its ref on 'cmid'). */
 718         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 719         kib_net_t              *net = peer->ibp_ni->ni_data;
 720         kib_dev_t              *dev;
 721         struct ib_qp_init_attr *init_qp_attr;
 722         struct kib_sched_info   *sched;
 723 #ifdef HAVE_IB_CQ_INIT_ATTR
 724         struct ib_cq_init_attr  cq_attr = {};
 725 #endif
 726         kib_conn_t              *conn;
 727         struct ib_cq            *cq;
 728         unsigned long           flags;
 729         int                     cpt;
 730         int                     rc;
 731         int                     i;
 732
 733         LASSERT(net != NULL);
 734         LASSERT(!in_interrupt());
 735
 736         dev = net->ibn_dev;
 737
 738         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 739         sched = kiblnd_data.kib_scheds[cpt];
 740
 741         LASSERT(sched->ibs_nthreads > 0);
 742
 743         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 744                          sizeof(*init_qp_attr));
 745         if (init_qp_attr == NULL) {
 746                 CERROR("Can't allocate qp_attr for %s\n",
 747                        libcfs_nid2str(peer->ibp_nid));
 748                 goto failed_0;
 749         }
 750
 751         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 752         if (conn == NULL) {
 753                 CERROR("Can't allocate connection for %s\n",
 754                        libcfs_nid2str(peer->ibp_nid));
 755                 goto failed_1;
 756         }
 757
 758         conn->ibc_state = IBLND_CONN_INIT;
 759         conn->ibc_version = version;
 760         conn->ibc_peer = peer;                  /* I take the caller's ref */
 761         cmid->context = conn;                   /* for future CM callbacks */
 762         conn->ibc_cmid = cmid;
 763         conn->ibc_max_frags = peer->ibp_max_frags;
 764         conn->ibc_queue_depth = peer->ibp_queue_depth;
 765
 766         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 767         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 768         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 769         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 770         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 771         INIT_LIST_HEAD(&conn->ibc_active_txs);
 772         spin_lock_init(&conn->ibc_lock);
 773
 774         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 775                          sizeof(*conn->ibc_connvars));
 776         if (conn->ibc_connvars == NULL) {
 777                 CERROR("Can't allocate in-progress connection state\n");
 778                 goto failed_2;
 779         }
 780
 781         write_lock_irqsave(glock, flags);
 782         if (dev->ibd_failover) {
 783                 write_unlock_irqrestore(glock, flags);
 784                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 785                 goto failed_2;
 786         }
 787
 788         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 789                 /* wakeup failover thread and teardown connection */
 790                 if (kiblnd_dev_can_failover(dev)) {
 791                         list_add_tail(&dev->ibd_fail_list,
 792                                       &kiblnd_data.kib_failed_devs);
 793                         wake_up(&kiblnd_data.kib_failover_waitq);
 794                 }
 795
 796                 write_unlock_irqrestore(glock, flags);
 797                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 798                        cmid->device->name, dev->ibd_ifname);
 799                 goto failed_2;
 800         }
 801
 802         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 803         conn->ibc_hdev = dev->ibd_hdev;
 804
 805         kiblnd_setup_mtu_locked(cmid);
 806
 807         write_unlock_irqrestore(glock, flags);
 808
 809         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 810                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 811         if (conn->ibc_rxs == NULL) {
 812                 CERROR("Cannot allocate RX buffers\n");
 813                 goto failed_2;
 814         }
 815
 816         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 817                                 IBLND_RX_MSG_PAGES(conn));
 818         if (rc != 0)
 819                 goto failed_2;
 820
 821         kiblnd_map_rx_descs(conn);
 822
 823 #ifdef HAVE_IB_CQ_INIT_ATTR
 824         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 825         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 826         cq = ib_create_cq(cmid->device,
 827                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 828                           &cq_attr);
 829 #else
 830         cq = ib_create_cq(cmid->device,
 831                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 832                           IBLND_CQ_ENTRIES(conn),
 833                           kiblnd_get_completion_vector(conn, cpt));
 834 #endif
 835         if (IS_ERR(cq)) {
 836                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 837                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 838                 goto failed_2;
 839         }
 840
 841         conn->ibc_cq = cq;
 842
 843         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 844         if (rc != 0) {
 845                 CERROR("Can't request completion notification: %d\n", rc);
 846                 goto failed_2;
 847         }
 848
 849         init_qp_attr->event_handler = kiblnd_qp_event;
 850         init_qp_attr->qp_context = conn;
 851         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 852         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 853         init_qp_attr->cap.max_send_sge = 1;
 854         init_qp_attr->cap.max_recv_sge = 1;
 855         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 856         init_qp_attr->qp_type = IB_QPT_RC;
 857         init_qp_attr->send_cq = cq;
 858         init_qp_attr->recv_cq = cq;
 859
 860         conn->ibc_sched = sched;
 861
 862         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 863         if (rc != 0) {
 864                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 865                        rc, init_qp_attr->cap.max_send_wr,
 866                        init_qp_attr->cap.max_recv_wr);
 867                 goto failed_2;
 868         }
 869
 870         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 871
 872         /* 1 ref for caller and each rxmsg */
 873         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 874         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 875
 876         /* post receives */
 877         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 878                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 879                                     IBLND_POSTRX_NO_CREDIT);
 880                 if (rc != 0) {
 881                         CERROR("Can't post rxmsg: %d\n", rc);
 882
 883                         /* Make posted receives complete */
 884                         kiblnd_abort_receives(conn);
 885
 886                         /* correct # of posted buffers
 887                          * NB locking needed now I'm racing with completion */
 888                         spin_lock_irqsave(&sched->ibs_lock, flags);
 889                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 890                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 891
 892                         /* cmid will be destroyed by CM(ofed) after cm_callback
 893                          * returned, so we can't refer it anymore
 894                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 895                         rdma_destroy_qp(conn->ibc_cmid);
 896                         conn->ibc_cmid = NULL;
 897
 898                         /* Drop my own and unused rxbuffer refcounts */
 899                         while (i++ <= IBLND_RX_MSGS(conn))
 900                                 kiblnd_conn_decref(conn);
 901
 902                         return NULL;
 903                 }
 904         }
 905
 906         /* Init successful! */
 907         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 908                  state == IBLND_CONN_PASSIVE_WAIT);
 909         conn->ibc_state = state;
 910
 911         /* 1 more conn */
 912         atomic_inc(&net->ibn_nconns);
 913         return conn;
 914
 915  failed_2:
 916         kiblnd_destroy_conn(conn);
 917  failed_1:
 918         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 919  failed_0:
 920         return NULL;
 921 }
 922
 923 void
 924 kiblnd_destroy_conn (kib_conn_t *conn)
 925 {
 926         struct rdma_cm_id *cmid = conn->ibc_cmid;
 927         kib_peer_t        *peer = conn->ibc_peer;
 928         int                rc;
 929
 930         LASSERT (!in_interrupt());
 931         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 932         LASSERT(list_empty(&conn->ibc_early_rxs));
 933         LASSERT(list_empty(&conn->ibc_tx_noops));
 934         LASSERT(list_empty(&conn->ibc_tx_queue));
 935         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 936         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 937         LASSERT(list_empty(&conn->ibc_active_txs));
 938         LASSERT (conn->ibc_noops_posted == 0);
 939         LASSERT (conn->ibc_nsends_posted == 0);
 940
 941         switch (conn->ibc_state) {
 942         default:
 943                 /* conn must be completely disengaged from the network */
 944                 LBUG();
 945
 946         case IBLND_CONN_DISCONNECTED:
 947                 /* connvars should have been freed already */
 948                 LASSERT (conn->ibc_connvars == NULL);
 949                 break;
 950
 951         case IBLND_CONN_INIT:
 952                 break;
 953         }
 954
 955         /* conn->ibc_cmid might be destroyed by CM already */
 956         if (cmid != NULL && cmid->qp != NULL)
 957                 rdma_destroy_qp(cmid);
 958
 959         if (conn->ibc_cq != NULL) {
 960                 rc = ib_destroy_cq(conn->ibc_cq);
 961                 if (rc != 0)
 962                         CWARN("Error destroying CQ: %d\n", rc);
 963         }
 964
 965         if (conn->ibc_rx_pages != NULL)
 966                 kiblnd_unmap_rx_descs(conn);
 967
 968         if (conn->ibc_rxs != NULL) {
 969                 LIBCFS_FREE(conn->ibc_rxs,
 970                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 971         }
 972
 973         if (conn->ibc_connvars != NULL)
 974                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 975
 976         if (conn->ibc_hdev != NULL)
 977                 kiblnd_hdev_decref(conn->ibc_hdev);
 978
 979         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 980         if (conn->ibc_state != IBLND_CONN_INIT) {
 981                 kib_net_t *net = peer->ibp_ni->ni_data;
 982
 983                 rdma_destroy_id(cmid);
 984                 atomic_dec(&net->ibn_nconns);
 985                 if (conn->ibc_conn_race) {
 986                         if (peer->ibp_accepting == 0 &&
 987                             !list_empty(&peer->ibp_tx_queue)) {
 988                                 kiblnd_connect_peer(peer);
 989                         } else  {
 990                                 rwlock_t *glock = &kiblnd_data.kib_global_lock;
 991                                 unsigned long flags;
 992
 993                                 write_lock_irqsave(glock, flags);
 994                                 peer->ibp_connecting--;
 995                                 write_unlock_irqrestore(glock, flags);
 996                         }
 997                 }
 998                 kiblnd_peer_decref(peer);
 999         }
1000
1001         LIBCFS_FREE(conn, sizeof(*conn));
1002 }
1003
1004 int
1005 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
1006 {
1007         kib_conn_t              *conn;
1008         struct list_head        *ctmp;
1009         struct list_head        *cnxt;
1010         int                     count = 0;
1011
1012         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1013                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1014
1015                 CDEBUG(D_NET, "Closing conn -> %s, "
1016                               "version: %x, reason: %d\n",
1017                        libcfs_nid2str(peer->ibp_nid),
1018                        conn->ibc_version, why);
1019
1020                 kiblnd_close_conn_locked(conn, why);
1021                 count++;
1022         }
1023
1024         return count;
1025 }
1026
1027 int
1028 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1029                                 int version, __u64 incarnation)
1030 {
1031         kib_conn_t              *conn;
1032         struct list_head        *ctmp;
1033         struct list_head        *cnxt;
1034         int                     count = 0;
1035
1036         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1037                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1038
1039                 if (conn->ibc_version     == version &&
1040                     conn->ibc_incarnation == incarnation)
1041                         continue;
1042
1043                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1044                               "incarnation:"LPX64"(%x, "LPX64")\n",
1045                        libcfs_nid2str(peer->ibp_nid),
1046                        conn->ibc_version, conn->ibc_incarnation,
1047                        version, incarnation);
1048
1049                 kiblnd_close_conn_locked(conn, -ESTALE);
1050                 count++;
1051         }
1052
1053         return count;
1054 }
1055
1056 static int
1057 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1058 {
1059         kib_peer_t              *peer;
1060         struct list_head        *ptmp;
1061         struct list_head        *pnxt;
1062         int                     lo;
1063         int                     hi;
1064         int                     i;
1065         unsigned long           flags;
1066         int                     count = 0;
1067
1068         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1069
1070         if (nid != LNET_NID_ANY)
1071                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1072         else {
1073                 lo = 0;
1074                 hi = kiblnd_data.kib_peer_hash_size - 1;
1075         }
1076
1077         for (i = lo; i <= hi; i++) {
1078                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1079
1080                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1081                         LASSERT(peer->ibp_connecting > 0 ||
1082                                 peer->ibp_accepting > 0 ||
1083                                 !list_empty(&peer->ibp_conns));
1084
1085                         if (peer->ibp_ni != ni)
1086                                 continue;
1087
1088                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1089                                 continue;
1090
1091                         count += kiblnd_close_peer_conns_locked(peer, 0);
1092                 }
1093         }
1094
1095         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1096
1097         /* wildcards always succeed */
1098         if (nid == LNET_NID_ANY)
1099                 return 0;
1100
1101         return (count == 0) ? -ENOENT : 0;
1102 }
1103
1104 static int
1105 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1106 {
1107         struct libcfs_ioctl_data *data = arg;
1108         int                       rc = -EINVAL;
1109
1110         switch(cmd) {
1111         case IOC_LIBCFS_GET_PEER: {
1112                 lnet_nid_t   nid = 0;
1113                 int          count = 0;
1114
1115                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1116                                           &nid, &count);
1117                 data->ioc_nid    = nid;
1118                 data->ioc_count  = count;
1119                 break;
1120         }
1121
1122         case IOC_LIBCFS_DEL_PEER: {
1123                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1124                 break;
1125         }
1126         case IOC_LIBCFS_GET_CONN: {
1127                 kib_conn_t *conn;
1128
1129                 rc = 0;
1130                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1131                 if (conn == NULL) {
1132                         rc = -ENOENT;
1133                         break;
1134                 }
1135
1136                 LASSERT (conn->ibc_cmid != NULL);
1137                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1138                 if (conn->ibc_cmid->route.path_rec == NULL)
1139                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1140                 else
1141                         data->ioc_u32[0] =
1142                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1143                 kiblnd_conn_decref(conn);
1144                 break;
1145         }
1146         case IOC_LIBCFS_CLOSE_CONNECTION: {
1147                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1148                 break;
1149         }
1150
1151         default:
1152                 break;
1153         }
1154
1155         return rc;
1156 }
1157
1158 static void
1159 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1160 {
1161         cfs_time_t      last_alive = 0;
1162         cfs_time_t      now = cfs_time_current();
1163         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1164         kib_peer_t      *peer;
1165         unsigned long   flags;
1166
1167         read_lock_irqsave(glock, flags);
1168
1169         peer = kiblnd_find_peer_locked(nid);
1170         if (peer != NULL) {
1171                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1172                         peer->ibp_accepting > 0 ||
1173                         !list_empty(&peer->ibp_conns));  /* active conn */
1174                 last_alive = peer->ibp_last_alive;
1175         }
1176
1177         read_unlock_irqrestore(glock, flags);
1178
1179         if (last_alive != 0)
1180                 *when = last_alive;
1181
1182         /* peer is not persistent in hash, trigger peer creation
1183          * and connection establishment with a NULL tx */
1184         if (peer == NULL)
1185                 kiblnd_launch_tx(ni, NULL, nid);
1186
1187         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1188                libcfs_nid2str(nid), peer,
1189                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1190         return;
1191 }
1192
1193 static void
1194 kiblnd_free_pages(kib_pages_t *p)
1195 {
1196         int     npages = p->ibp_npages;
1197         int     i;
1198
1199         for (i = 0; i < npages; i++) {
1200                 if (p->ibp_pages[i] != NULL)
1201                         __free_page(p->ibp_pages[i]);
1202         }
1203
1204         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1205 }
1206
1207 int
1208 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1209 {
1210         kib_pages_t     *p;
1211         int             i;
1212
1213         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1214                          offsetof(kib_pages_t, ibp_pages[npages]));
1215         if (p == NULL) {
1216                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1217                 return -ENOMEM;
1218         }
1219
1220         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1221         p->ibp_npages = npages;
1222
1223         for (i = 0; i < npages; i++) {
1224                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1225                                                      GFP_NOFS);
1226                 if (p->ibp_pages[i] == NULL) {
1227                         CERROR("Can't allocate page %d of %d\n", i, npages);
1228                         kiblnd_free_pages(p);
1229                         return -ENOMEM;
1230                 }
1231         }
1232
1233         *pp = p;
1234         return 0;
1235 }
1236
1237 void
1238 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1239 {
1240         kib_rx_t *rx;
1241         int       i;
1242
1243         LASSERT (conn->ibc_rxs != NULL);
1244         LASSERT (conn->ibc_hdev != NULL);
1245
1246         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1247                 rx = &conn->ibc_rxs[i];
1248
1249                 LASSERT(rx->rx_nob >= 0); /* not posted */
1250
1251                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1252                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1253                                                           rx->rx_msgaddr),
1254                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1255         }
1256
1257         kiblnd_free_pages(conn->ibc_rx_pages);
1258
1259         conn->ibc_rx_pages = NULL;
1260 }
1261
1262 void
1263 kiblnd_map_rx_descs(kib_conn_t *conn)
1264 {
1265         kib_rx_t       *rx;
1266         struct page    *pg;
1267         int             pg_off;
1268         int             ipg;
1269         int             i;
1270
1271         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1272                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1273                 rx = &conn->ibc_rxs[i];
1274
1275                 rx->rx_conn = conn;
1276                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1277
1278                 rx->rx_msgaddr =
1279                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1280                                               rx->rx_msg, IBLND_MSG_SIZE,
1281                                               DMA_FROM_DEVICE);
1282                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1283                                                   rx->rx_msgaddr));
1284                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1285
1286                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1287                        i, rx->rx_msg, rx->rx_msgaddr,
1288                        (__u64)(page_to_phys(pg) + pg_off));
1289
1290                 pg_off += IBLND_MSG_SIZE;
1291                 LASSERT(pg_off <= PAGE_SIZE);
1292
1293                 if (pg_off == PAGE_SIZE) {
1294                         pg_off = 0;
1295                         ipg++;
1296                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1297                 }
1298         }
1299 }
1300
1301 static void
1302 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1303 {
1304         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1305         kib_tx_t       *tx;
1306         int             i;
1307
1308         LASSERT (tpo->tpo_pool.po_allocated == 0);
1309
1310         if (hdev == NULL)
1311                 return;
1312
1313         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1314                 tx = &tpo->tpo_tx_descs[i];
1315                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1316                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1317                                                           tx->tx_msgaddr),
1318                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1319         }
1320
1321         kiblnd_hdev_decref(hdev);
1322         tpo->tpo_hdev = NULL;
1323 }
1324
1325 static kib_hca_dev_t *
1326 kiblnd_current_hdev(kib_dev_t *dev)
1327 {
1328         kib_hca_dev_t *hdev;
1329         unsigned long  flags;
1330         int            i = 0;
1331
1332         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1333         while (dev->ibd_failover) {
1334                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1335                 if (i++ % 50 == 0)
1336                         CDEBUG(D_NET, "%s: Wait for failover\n",
1337                                dev->ibd_ifname);
1338                 schedule_timeout(cfs_time_seconds(1) / 100);
1339
1340                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1341         }
1342
1343         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1344         hdev = dev->ibd_hdev;
1345
1346         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1347
1348         return hdev;
1349 }
1350
1351 static void
1352 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1353 {
1354         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1355         kib_pool_t     *pool  = &tpo->tpo_pool;
1356         kib_net_t      *net   = pool->po_owner->ps_net;
1357         kib_dev_t      *dev;
1358         struct page    *page;
1359         kib_tx_t       *tx;
1360         int             page_offset;
1361         int             ipage;
1362         int             i;
1363
1364         LASSERT (net != NULL);
1365
1366         dev = net->ibn_dev;
1367
1368         /* pre-mapped messages are not bigger than 1 page */
1369         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1370
1371         /* No fancy arithmetic when we do the buffer calculations */
1372         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1373
1374         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1375
1376         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1377                 page = txpgs->ibp_pages[ipage];
1378                 tx = &tpo->tpo_tx_descs[i];
1379
1380                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1381                                            page_offset);
1382
1383                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1384                                                        tx->tx_msg,
1385                                                        IBLND_MSG_SIZE,
1386                                                        DMA_TO_DEVICE);
1387                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1388                                                   tx->tx_msgaddr));
1389                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1390
1391                 list_add(&tx->tx_list, &pool->po_free_list);
1392
1393                 page_offset += IBLND_MSG_SIZE;
1394                 LASSERT(page_offset <= PAGE_SIZE);
1395
1396                 if (page_offset == PAGE_SIZE) {
1397                         page_offset = 0;
1398                         ipage++;
1399                         LASSERT(ipage <= txpgs->ibp_npages);
1400                 }
1401         }
1402 }
1403
1404 struct ib_mr *
1405 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd,
1406                       int negotiated_nfrags)
1407 {
1408         __u16   nfrags = (negotiated_nfrags != -1) ?
1409           negotiated_nfrags : *kiblnd_tunables.kib_map_on_demand;
1410
1411         LASSERT(hdev->ibh_mrs != NULL);
1412
1413         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1414             nfrags <= rd->rd_nfrags)
1415                 return NULL;
1416
1417         return hdev->ibh_mrs;
1418 }
1419
1420 static void
1421 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1422 {
1423         LASSERT (pool->fpo_map_count == 0);
1424
1425         if (pool->fpo_fmr_pool != NULL)
1426                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1427
1428         if (pool->fpo_hdev != NULL)
1429                 kiblnd_hdev_decref(pool->fpo_hdev);
1430
1431         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1432 }
1433
1434 static void
1435 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1436 {
1437         kib_fmr_pool_t *pool;
1438
1439         while (!list_empty(head)) {
1440                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1441                 list_del(&pool->fpo_list);
1442                 kiblnd_destroy_fmr_pool(pool);
1443         }
1444 }
1445
1446 static int kiblnd_fmr_pool_size(int ncpts)
1447 {
1448         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1449
1450         return max(IBLND_FMR_POOL, size);
1451 }
1452
1453 static int kiblnd_fmr_flush_trigger(int ncpts)
1454 {
1455         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1456
1457         return max(IBLND_FMR_POOL_FLUSH, size);
1458 }
1459
1460 static int
1461 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1462 {
1463         /* FMR pool for RDMA */
1464         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1465         kib_fmr_pool_t          *fpo;
1466         struct ib_fmr_pool_param param = {
1467                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1468                 .page_shift        = PAGE_SHIFT,
1469                 .access            = (IB_ACCESS_LOCAL_WRITE |
1470                                       IB_ACCESS_REMOTE_WRITE),
1471                 .pool_size         = fps->fps_pool_size,
1472                 .dirty_watermark   = fps->fps_flush_trigger,
1473                 .flush_function    = NULL,
1474                 .flush_arg         = NULL,
1475                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1476         int rc;
1477
1478         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1479         if (fpo == NULL)
1480                 return -ENOMEM;
1481
1482         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1483
1484         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1485         if (IS_ERR(fpo->fpo_fmr_pool)) {
1486                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1487                 CERROR("Failed to create FMR pool: %d\n", rc);
1488
1489                 kiblnd_hdev_decref(fpo->fpo_hdev);
1490                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1491                 return rc;
1492         }
1493
1494         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1495         fpo->fpo_owner    = fps;
1496         *pp_fpo = fpo;
1497
1498         return 0;
1499 }
1500
1501 static void
1502 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1503 {
1504         if (fps->fps_net == NULL) /* intialized? */
1505                 return;
1506
1507         spin_lock(&fps->fps_lock);
1508
1509         while (!list_empty(&fps->fps_pool_list)) {
1510                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1511                                                  kib_fmr_pool_t, fpo_list);
1512                 fpo->fpo_failed = 1;
1513                 list_del(&fpo->fpo_list);
1514                 if (fpo->fpo_map_count == 0)
1515                         list_add(&fpo->fpo_list, zombies);
1516                 else
1517                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1518         }
1519
1520         spin_unlock(&fps->fps_lock);
1521 }
1522
1523 static void
1524 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1525 {
1526         if (fps->fps_net != NULL) { /* initialized? */
1527                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1528                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1529         }
1530 }
1531
1532 static int
1533 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1534                         int pool_size, int flush_trigger)
1535 {
1536         kib_fmr_pool_t *fpo;
1537         int             rc;
1538
1539         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1540
1541         fps->fps_net = net;
1542         fps->fps_cpt = cpt;
1543         fps->fps_pool_size = pool_size;
1544         fps->fps_flush_trigger = flush_trigger;
1545         spin_lock_init(&fps->fps_lock);
1546         INIT_LIST_HEAD(&fps->fps_pool_list);
1547         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1548
1549         rc = kiblnd_create_fmr_pool(fps, &fpo);
1550         if (rc == 0)
1551                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1552
1553         return rc;
1554 }
1555
1556 static int
1557 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1558 {
1559         if (fpo->fpo_map_count != 0) /* still in use */
1560                 return 0;
1561         if (fpo->fpo_failed)
1562                 return 1;
1563         return cfs_time_aftereq(now, fpo->fpo_deadline);
1564 }
1565
1566 void
1567 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1568 {
1569         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1570         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1571         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1572         cfs_time_t         now = cfs_time_current();
1573         kib_fmr_pool_t    *tmp;
1574         int                rc;
1575
1576         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1577         LASSERT(rc == 0);
1578
1579         if (status != 0) {
1580                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1581                 LASSERT(rc == 0);
1582         }
1583
1584         fmr->fmr_pool = NULL;
1585         fmr->fmr_pfmr = NULL;
1586
1587         spin_lock(&fps->fps_lock);
1588         fpo->fpo_map_count--;   /* decref the pool */
1589
1590         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1591                 /* the first pool is persistent */
1592                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1593                         continue;
1594
1595                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1596                         list_move(&fpo->fpo_list, &zombies);
1597                         fps->fps_version++;
1598                 }
1599         }
1600         spin_unlock(&fps->fps_lock);
1601
1602         if (!list_empty(&zombies))
1603                 kiblnd_destroy_fmr_pool_list(&zombies);
1604 }
1605
1606 int
1607 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1608                     __u64 iov, kib_fmr_t *fmr)
1609 {
1610         struct ib_pool_fmr *pfmr;
1611         kib_fmr_pool_t     *fpo;
1612         __u64               version;
1613         int                 rc;
1614
1615 again:
1616         spin_lock(&fps->fps_lock);
1617         version = fps->fps_version;
1618         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1619                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1620                 fpo->fpo_map_count++;
1621                 spin_unlock(&fps->fps_lock);
1622
1623                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1624                                             pages, npages, iov);
1625                 if (likely(!IS_ERR(pfmr))) {
1626                         fmr->fmr_pool = fpo;
1627                         fmr->fmr_pfmr = pfmr;
1628                         return 0;
1629                 }
1630
1631                 spin_lock(&fps->fps_lock);
1632                 fpo->fpo_map_count--;
1633                 if (PTR_ERR(pfmr) != -EAGAIN) {
1634                         spin_unlock(&fps->fps_lock);
1635                         return PTR_ERR(pfmr);
1636                 }
1637
1638                 /* EAGAIN and ... */
1639                 if (version != fps->fps_version) {
1640                         spin_unlock(&fps->fps_lock);
1641                         goto again;
1642                 }
1643         }
1644
1645         if (fps->fps_increasing) {
1646                 spin_unlock(&fps->fps_lock);
1647                 CDEBUG(D_NET, "Another thread is allocating new "
1648                        "FMR pool, waiting for her to complete\n");
1649                 schedule();
1650                 goto again;
1651
1652         }
1653
1654         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1655                 /* someone failed recently */
1656                 spin_unlock(&fps->fps_lock);
1657                 return -EAGAIN;
1658         }
1659
1660         fps->fps_increasing = 1;
1661         spin_unlock(&fps->fps_lock);
1662
1663         CDEBUG(D_NET, "Allocate new FMR pool\n");
1664         rc = kiblnd_create_fmr_pool(fps, &fpo);
1665         spin_lock(&fps->fps_lock);
1666         fps->fps_increasing = 0;
1667         if (rc == 0) {
1668                 fps->fps_version++;
1669                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1670         } else {
1671                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1672         }
1673         spin_unlock(&fps->fps_lock);
1674
1675         goto again;
1676 }
1677
1678 static void
1679 kiblnd_fini_pool(kib_pool_t *pool)
1680 {
1681         LASSERT(list_empty(&pool->po_free_list));
1682         LASSERT(pool->po_allocated == 0);
1683
1684         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1685 }
1686
1687 static void
1688 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1689 {
1690         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1691
1692         memset(pool, 0, sizeof(kib_pool_t));
1693         INIT_LIST_HEAD(&pool->po_free_list);
1694         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1695         pool->po_owner    = ps;
1696         pool->po_size     = size;
1697 }
1698
1699 static void
1700 kiblnd_destroy_pool_list(struct list_head *head)
1701 {
1702         kib_pool_t *pool;
1703
1704         while (!list_empty(head)) {
1705                 pool = list_entry(head->next, kib_pool_t, po_list);
1706                 list_del(&pool->po_list);
1707
1708                 LASSERT(pool->po_owner != NULL);
1709                 pool->po_owner->ps_pool_destroy(pool);
1710         }
1711 }
1712
1713 static void
1714 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1715 {
1716         if (ps->ps_net == NULL) /* intialized? */
1717                 return;
1718
1719         spin_lock(&ps->ps_lock);
1720         while (!list_empty(&ps->ps_pool_list)) {
1721                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1722                                             kib_pool_t, po_list);
1723                 po->po_failed = 1;
1724                 list_del(&po->po_list);
1725                 if (po->po_allocated == 0)
1726                         list_add(&po->po_list, zombies);
1727                 else
1728                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1729         }
1730         spin_unlock(&ps->ps_lock);
1731 }
1732
1733 static void
1734 kiblnd_fini_poolset(kib_poolset_t *ps)
1735 {
1736         if (ps->ps_net != NULL) { /* initialized? */
1737                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1738                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1739         }
1740 }
1741
1742 static int
1743 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1744                     kib_net_t *net, char *name, int size,
1745                     kib_ps_pool_create_t po_create,
1746                     kib_ps_pool_destroy_t po_destroy,
1747                     kib_ps_node_init_t nd_init,
1748                     kib_ps_node_fini_t nd_fini)
1749 {
1750         kib_pool_t      *pool;
1751         int             rc;
1752
1753         memset(ps, 0, sizeof(kib_poolset_t));
1754
1755         ps->ps_cpt          = cpt;
1756         ps->ps_net          = net;
1757         ps->ps_pool_create  = po_create;
1758         ps->ps_pool_destroy = po_destroy;
1759         ps->ps_node_init    = nd_init;
1760         ps->ps_node_fini    = nd_fini;
1761         ps->ps_pool_size    = size;
1762         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1763             >= sizeof(ps->ps_name))
1764                 return -E2BIG;
1765         spin_lock_init(&ps->ps_lock);
1766         INIT_LIST_HEAD(&ps->ps_pool_list);
1767         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1768
1769         rc = ps->ps_pool_create(ps, size, &pool);
1770         if (rc == 0)
1771                 list_add(&pool->po_list, &ps->ps_pool_list);
1772         else
1773                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1774
1775         return rc;
1776 }
1777
1778 static int
1779 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1780 {
1781         if (pool->po_allocated != 0) /* still in use */
1782                 return 0;
1783         if (pool->po_failed)
1784                 return 1;
1785         return cfs_time_aftereq(now, pool->po_deadline);
1786 }
1787
1788 void
1789 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1790 {
1791         struct list_head zombies = LIST_HEAD_INIT(zombies);
1792         kib_poolset_t   *ps = pool->po_owner;
1793         kib_pool_t      *tmp;
1794         cfs_time_t       now = cfs_time_current();
1795
1796         spin_lock(&ps->ps_lock);
1797
1798         if (ps->ps_node_fini != NULL)
1799                 ps->ps_node_fini(pool, node);
1800
1801         LASSERT(pool->po_allocated > 0);
1802         list_add(node, &pool->po_free_list);
1803         pool->po_allocated--;
1804
1805         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1806                 /* the first pool is persistent */
1807                 if (ps->ps_pool_list.next == &pool->po_list)
1808                         continue;
1809
1810                 if (kiblnd_pool_is_idle(pool, now))
1811                         list_move(&pool->po_list, &zombies);
1812         }
1813         spin_unlock(&ps->ps_lock);
1814
1815         if (!list_empty(&zombies))
1816                 kiblnd_destroy_pool_list(&zombies);
1817 }
1818
1819 struct list_head *
1820 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1821 {
1822         struct list_head        *node;
1823         kib_pool_t              *pool;
1824         int                     rc;
1825
1826 again:
1827         spin_lock(&ps->ps_lock);
1828         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1829                 if (list_empty(&pool->po_free_list))
1830                         continue;
1831
1832                 pool->po_allocated++;
1833                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1834                 node = pool->po_free_list.next;
1835                 list_del(node);
1836
1837                 if (ps->ps_node_init != NULL) {
1838                         /* still hold the lock */
1839                         ps->ps_node_init(pool, node);
1840                 }
1841                 spin_unlock(&ps->ps_lock);
1842                 return node;
1843         }
1844
1845         /* no available tx pool and ... */
1846         if (ps->ps_increasing) {
1847                 /* another thread is allocating a new pool */
1848                 spin_unlock(&ps->ps_lock);
1849                 CDEBUG(D_NET, "Another thread is allocating new "
1850                        "%s pool, waiting for her to complete\n",
1851                        ps->ps_name);
1852                 schedule();
1853                 goto again;
1854         }
1855
1856         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1857                 /* someone failed recently */
1858                 spin_unlock(&ps->ps_lock);
1859                 return NULL;
1860         }
1861
1862         ps->ps_increasing = 1;
1863         spin_unlock(&ps->ps_lock);
1864
1865         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1866
1867         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1868
1869         spin_lock(&ps->ps_lock);
1870         ps->ps_increasing = 0;
1871         if (rc == 0) {
1872                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1873         } else {
1874                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1875                 CERROR("Can't allocate new %s pool because out of memory\n",
1876                        ps->ps_name);
1877         }
1878         spin_unlock(&ps->ps_lock);
1879
1880         goto again;
1881 }
1882
1883 static void
1884 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1885 {
1886         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1887         int             i;
1888
1889         LASSERT (pool->po_allocated == 0);
1890
1891         if (tpo->tpo_tx_pages != NULL) {
1892                 kiblnd_unmap_tx_pool(tpo);
1893                 kiblnd_free_pages(tpo->tpo_tx_pages);
1894         }
1895
1896         if (tpo->tpo_tx_descs == NULL)
1897                 goto out;
1898
1899         for (i = 0; i < pool->po_size; i++) {
1900                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1901
1902                 list_del(&tx->tx_list);
1903                 if (tx->tx_pages != NULL)
1904                         LIBCFS_FREE(tx->tx_pages,
1905                                     LNET_MAX_IOV *
1906                                     sizeof(*tx->tx_pages));
1907                 if (tx->tx_frags != NULL)
1908                         LIBCFS_FREE(tx->tx_frags,
1909                                     IBLND_MAX_RDMA_FRAGS *
1910                                             sizeof(*tx->tx_frags));
1911                 if (tx->tx_wrq != NULL)
1912                         LIBCFS_FREE(tx->tx_wrq,
1913                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1914                                     sizeof(*tx->tx_wrq));
1915                 if (tx->tx_sge != NULL)
1916                         LIBCFS_FREE(tx->tx_sge,
1917                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1918                                     sizeof(*tx->tx_sge));
1919                 if (tx->tx_rd != NULL)
1920                         LIBCFS_FREE(tx->tx_rd,
1921                                     offsetof(kib_rdma_desc_t,
1922                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1923         }
1924
1925         LIBCFS_FREE(tpo->tpo_tx_descs,
1926                     pool->po_size * sizeof(kib_tx_t));
1927 out:
1928         kiblnd_fini_pool(pool);
1929         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1930 }
1931
1932 static int kiblnd_tx_pool_size(int ncpts)
1933 {
1934         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1935
1936         return max(IBLND_TX_POOL, ntx);
1937 }
1938
1939 static int
1940 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1941 {
1942         int            i;
1943         int            npg;
1944         kib_pool_t    *pool;
1945         kib_tx_pool_t *tpo;
1946
1947         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1948         if (tpo == NULL) {
1949                 CERROR("Failed to allocate TX pool\n");
1950                 return -ENOMEM;
1951         }
1952
1953         pool = &tpo->tpo_pool;
1954         kiblnd_init_pool(ps, pool, size);
1955         tpo->tpo_tx_descs = NULL;
1956         tpo->tpo_tx_pages = NULL;
1957
1958         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1959         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1960                 CERROR("Can't allocate tx pages: %d\n", npg);
1961                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1962                 return -ENOMEM;
1963         }
1964
1965         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1966                          size * sizeof(kib_tx_t));
1967         if (tpo->tpo_tx_descs == NULL) {
1968                 CERROR("Can't allocate %d tx descriptors\n", size);
1969                 ps->ps_pool_destroy(pool);
1970                 return -ENOMEM;
1971         }
1972
1973         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1974
1975         for (i = 0; i < size; i++) {
1976                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1977
1978                 tx->tx_pool = tpo;
1979                 if (ps->ps_net->ibn_fmr_ps != NULL) {
1980                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1981                                          lnet_cpt_table(), ps->ps_cpt,
1982                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1983                         if (tx->tx_pages == NULL)
1984                                 break;
1985                 }
1986
1987                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1988                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1989                 if (tx->tx_frags == NULL)
1990                         break;
1991
1992                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1993
1994                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1995                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1996                                  sizeof(*tx->tx_wrq));
1997                 if (tx->tx_wrq == NULL)
1998                         break;
1999
2000                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2001                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2002                                  sizeof(*tx->tx_sge));
2003                 if (tx->tx_sge == NULL)
2004                         break;
2005
2006                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2007                                  offsetof(kib_rdma_desc_t,
2008                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2009                 if (tx->tx_rd == NULL)
2010                         break;
2011         }
2012
2013         if (i == size) {
2014                 kiblnd_map_tx_pool(tpo);
2015                 *pp_po = pool;
2016                 return 0;
2017         }
2018
2019         ps->ps_pool_destroy(pool);
2020         return -ENOMEM;
2021 }
2022
2023 static void
2024 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2025 {
2026         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2027                                              tps_poolset);
2028         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2029
2030         tx->tx_cookie = tps->tps_next_tx_cookie++;
2031 }
2032
2033 static void
2034 kiblnd_net_fini_pools(kib_net_t *net)
2035 {
2036         int     i;
2037
2038         cfs_cpt_for_each(i, lnet_cpt_table()) {
2039                 kib_tx_poolset_t        *tps;
2040                 kib_fmr_poolset_t       *fps;
2041
2042                 if (net->ibn_tx_ps != NULL) {
2043                         tps = net->ibn_tx_ps[i];
2044                         kiblnd_fini_poolset(&tps->tps_poolset);
2045                 }
2046
2047                 if (net->ibn_fmr_ps != NULL) {
2048                         fps = net->ibn_fmr_ps[i];
2049                         kiblnd_fini_fmr_poolset(fps);
2050                 }
2051         }
2052
2053         if (net->ibn_tx_ps != NULL) {
2054                 cfs_percpt_free(net->ibn_tx_ps);
2055                 net->ibn_tx_ps = NULL;
2056         }
2057
2058         if (net->ibn_fmr_ps != NULL) {
2059                 cfs_percpt_free(net->ibn_fmr_ps);
2060                 net->ibn_fmr_ps = NULL;
2061         }
2062 }
2063
2064 static int
2065 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2066 {
2067         unsigned long   flags;
2068         int             cpt;
2069         int             rc;
2070         int             i;
2071
2072         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2073         if (*kiblnd_tunables.kib_map_on_demand == 0) {
2074                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2075                                            flags);
2076                 goto create_tx_pool;
2077         }
2078
2079         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2080
2081         if (*kiblnd_tunables.kib_fmr_pool_size <
2082             *kiblnd_tunables.kib_ntx / 4) {
2083                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2084                        *kiblnd_tunables.kib_fmr_pool_size,
2085                        *kiblnd_tunables.kib_ntx / 4);
2086                 rc = -EINVAL;
2087                 goto failed;
2088         }
2089
2090         /* TX pool must be created later than FMR, see LU-2268
2091          * for details */
2092         LASSERT(net->ibn_tx_ps == NULL);
2093
2094         /* premapping can fail if ibd_nmr > 1, so we always create
2095          * FMR pool and map-on-demand if premapping failed */
2096
2097         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2098                                            sizeof(kib_fmr_poolset_t));
2099         if (net->ibn_fmr_ps == NULL) {
2100                 CERROR("Failed to allocate FMR pool array\n");
2101                 rc = -ENOMEM;
2102                 goto failed;
2103         }
2104
2105         for (i = 0; i < ncpts; i++) {
2106                 cpt = (cpts == NULL) ? i : cpts[i];
2107                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2108                                              kiblnd_fmr_pool_size(ncpts),
2109                                              kiblnd_fmr_flush_trigger(ncpts));
2110                 if (rc != 0) {
2111                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2112                                cpt, rc);
2113                         goto failed;
2114                 }
2115         }
2116
2117         if (i > 0)
2118                 LASSERT(i == ncpts);
2119
2120  create_tx_pool:
2121         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2122                                           sizeof(kib_tx_poolset_t));
2123         if (net->ibn_tx_ps == NULL) {
2124                 CERROR("Failed to allocate tx pool array\n");
2125                 rc = -ENOMEM;
2126                 goto failed;
2127         }
2128
2129         for (i = 0; i < ncpts; i++) {
2130                 cpt = (cpts == NULL) ? i : cpts[i];
2131                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2132                                          cpt, net, "TX",
2133                                          kiblnd_tx_pool_size(ncpts),
2134                                          kiblnd_create_tx_pool,
2135                                          kiblnd_destroy_tx_pool,
2136                                          kiblnd_tx_init, NULL);
2137                 if (rc != 0) {
2138                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2139                                cpt, rc);
2140                         goto failed;
2141                 }
2142         }
2143
2144         return 0;
2145  failed:
2146         kiblnd_net_fini_pools(net);
2147         LASSERT(rc != 0);
2148         return rc;
2149 }
2150
2151 static int
2152 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2153 {
2154         struct ib_device_attr *attr;
2155         int                    rc;
2156
2157         /* It's safe to assume a HCA can handle a page size
2158          * matching that of the native system */
2159         hdev->ibh_page_shift = PAGE_SHIFT;
2160         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2161         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2162
2163         LIBCFS_ALLOC(attr, sizeof(*attr));
2164         if (attr == NULL) {
2165                 CERROR("Out of memory\n");
2166                 return -ENOMEM;
2167         }
2168
2169         rc = ib_query_device(hdev->ibh_ibdev, attr);
2170         if (rc == 0)
2171                 hdev->ibh_mr_size = attr->max_mr_size;
2172
2173         LIBCFS_FREE(attr, sizeof(*attr));
2174
2175         if (rc != 0) {
2176                 CERROR("Failed to query IB device: %d\n", rc);
2177                 return rc;
2178         }
2179
2180         if (hdev->ibh_mr_size == ~0ULL) {
2181                 hdev->ibh_mr_shift = 64;
2182                 return 0;
2183         }
2184
2185         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2186         return -EINVAL;
2187 }
2188
2189 static void
2190 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2191 {
2192         if (hdev->ibh_mrs == NULL)
2193                 return;
2194
2195         ib_dereg_mr(hdev->ibh_mrs);
2196
2197         hdev->ibh_mrs = NULL;
2198 }
2199
2200 void
2201 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2202 {
2203         kiblnd_hdev_cleanup_mrs(hdev);
2204
2205         if (hdev->ibh_pd != NULL)
2206                 ib_dealloc_pd(hdev->ibh_pd);
2207
2208         if (hdev->ibh_cmid != NULL)
2209                 rdma_destroy_id(hdev->ibh_cmid);
2210
2211         LIBCFS_FREE(hdev, sizeof(*hdev));
2212 }
2213
2214 static int
2215 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2216 {
2217         struct ib_mr *mr;
2218         int           rc;
2219         int           acflags = IB_ACCESS_LOCAL_WRITE |
2220                                 IB_ACCESS_REMOTE_WRITE;
2221
2222         rc = kiblnd_hdev_get_attr(hdev);
2223         if (rc != 0)
2224                 return rc;
2225
2226         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2227         if (IS_ERR(mr)) {
2228                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2229                 kiblnd_hdev_cleanup_mrs(hdev);
2230                 return PTR_ERR(mr);
2231         }
2232
2233         hdev->ibh_mrs = mr;
2234
2235         return 0;
2236 }
2237
2238 static int
2239 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2240 {       /* DUMMY */
2241         return 0;
2242 }
2243
2244 static int
2245 kiblnd_dev_need_failover(kib_dev_t *dev)
2246 {
2247         struct rdma_cm_id  *cmid;
2248         struct sockaddr_in  srcaddr;
2249         struct sockaddr_in  dstaddr;
2250         int                 rc;
2251
2252         if (dev->ibd_hdev == NULL || /* initializing */
2253             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2254             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2255                 return 1;
2256
2257         /* XXX: it's UGLY, but I don't have better way to find
2258          * ib-bonding HCA failover because:
2259          *
2260          * a. no reliable CM event for HCA failover...
2261          * b. no OFED API to get ib_device for current net_device...
2262          *
2263          * We have only two choices at this point:
2264          *
2265          * a. rdma_bind_addr(), it will conflict with listener cmid
2266          * b. rdma_resolve_addr() to zero addr */
2267         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2268                                      IB_QPT_RC);
2269         if (IS_ERR(cmid)) {
2270                 rc = PTR_ERR(cmid);
2271                 CERROR("Failed to create cmid for failover: %d\n", rc);
2272                 return rc;
2273         }
2274
2275         memset(&srcaddr, 0, sizeof(srcaddr));
2276         srcaddr.sin_family      = AF_INET;
2277         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2278
2279         memset(&dstaddr, 0, sizeof(dstaddr));
2280         dstaddr.sin_family = AF_INET;
2281         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2282                                (struct sockaddr *)&dstaddr, 1);
2283         if (rc != 0 || cmid->device == NULL) {
2284                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2285                        dev->ibd_ifname, &dev->ibd_ifip,
2286                        cmid->device, rc);
2287                 rdma_destroy_id(cmid);
2288                 return rc;
2289         }
2290
2291         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2292         rdma_destroy_id(cmid);
2293         return rc;
2294 }
2295
2296 int
2297 kiblnd_dev_failover(kib_dev_t *dev)
2298 {
2299         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2300         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2301         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2302         struct rdma_cm_id  *cmid  = NULL;
2303         kib_hca_dev_t      *hdev  = NULL;
2304         kib_hca_dev_t      *old;
2305         struct ib_pd       *pd;
2306         kib_net_t          *net;
2307         struct sockaddr_in  addr;
2308         unsigned long       flags;
2309         int                 rc = 0;
2310         int                 i;
2311
2312         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2313                  dev->ibd_can_failover ||
2314                  dev->ibd_hdev == NULL);
2315
2316         rc = kiblnd_dev_need_failover(dev);
2317         if (rc <= 0)
2318                 goto out;
2319
2320         if (dev->ibd_hdev != NULL &&
2321             dev->ibd_hdev->ibh_cmid != NULL) {
2322                 /* XXX it's not good to close old listener at here,
2323                  * because we can fail to create new listener.
2324                  * But we have to close it now, otherwise rdma_bind_addr
2325                  * will return EADDRINUSE... How crap! */
2326                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2327
2328                 cmid = dev->ibd_hdev->ibh_cmid;
2329                 /* make next schedule of kiblnd_dev_need_failover()
2330                  * return 1 for me */
2331                 dev->ibd_hdev->ibh_cmid  = NULL;
2332                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2333
2334                 rdma_destroy_id(cmid);
2335         }
2336
2337         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2338                                      IB_QPT_RC);
2339         if (IS_ERR(cmid)) {
2340                 rc = PTR_ERR(cmid);
2341                 CERROR("Failed to create cmid for failover: %d\n", rc);
2342                 goto out;
2343         }
2344
2345         memset(&addr, 0, sizeof(addr));
2346         addr.sin_family      = AF_INET;
2347         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2348         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2349
2350         /* Bind to failover device or port */
2351         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2352         if (rc != 0 || cmid->device == NULL) {
2353                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2354                        dev->ibd_ifname, &dev->ibd_ifip,
2355                        cmid->device, rc);
2356                 rdma_destroy_id(cmid);
2357                 goto out;
2358         }
2359
2360         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2361         if (hdev == NULL) {
2362                 CERROR("Failed to allocate kib_hca_dev\n");
2363                 rdma_destroy_id(cmid);
2364                 rc = -ENOMEM;
2365                 goto out;
2366         }
2367
2368         atomic_set(&hdev->ibh_ref, 1);
2369         hdev->ibh_dev   = dev;
2370         hdev->ibh_cmid  = cmid;
2371         hdev->ibh_ibdev = cmid->device;
2372
2373         pd = ib_alloc_pd(cmid->device);
2374         if (IS_ERR(pd)) {
2375                 rc = PTR_ERR(pd);
2376                 CERROR("Can't allocate PD: %d\n", rc);
2377                 goto out;
2378         }
2379
2380         hdev->ibh_pd = pd;
2381
2382         rc = rdma_listen(cmid, 0);
2383         if (rc != 0) {
2384                 CERROR("Can't start new listener: %d\n", rc);
2385                 goto out;
2386         }
2387
2388         rc = kiblnd_hdev_setup_mrs(hdev);
2389         if (rc != 0) {
2390                 CERROR("Can't setup device: %d\n", rc);
2391                 goto out;
2392         }
2393
2394         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2395
2396         old = dev->ibd_hdev;
2397         dev->ibd_hdev = hdev;   /* take over the refcount */
2398         hdev = old;
2399
2400         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2401                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2402                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2403                                             &zombie_tpo);
2404
2405                         if (net->ibn_fmr_ps != NULL)
2406                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2407                                                         &zombie_fpo);
2408                 }
2409         }
2410
2411         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2412  out:
2413         if (!list_empty(&zombie_tpo))
2414                 kiblnd_destroy_pool_list(&zombie_tpo);
2415         if (!list_empty(&zombie_ppo))
2416                 kiblnd_destroy_pool_list(&zombie_ppo);
2417         if (!list_empty(&zombie_fpo))
2418                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2419         if (hdev != NULL)
2420                 kiblnd_hdev_decref(hdev);
2421
2422         if (rc != 0)
2423                 dev->ibd_failed_failover++;
2424         else
2425                 dev->ibd_failed_failover = 0;
2426
2427         return rc;
2428 }
2429
2430 void
2431 kiblnd_destroy_dev (kib_dev_t *dev)
2432 {
2433         LASSERT (dev->ibd_nnets == 0);
2434         LASSERT(list_empty(&dev->ibd_nets));
2435
2436         list_del(&dev->ibd_fail_list);
2437         list_del(&dev->ibd_list);
2438
2439         if (dev->ibd_hdev != NULL)
2440                 kiblnd_hdev_decref(dev->ibd_hdev);
2441
2442         LIBCFS_FREE(dev, sizeof(*dev));
2443 }
2444
2445 static kib_dev_t *
2446 kiblnd_create_dev(char *ifname)
2447 {
2448         struct net_device *netdev;
2449         kib_dev_t         *dev;
2450         __u32              netmask;
2451         __u32              ip;
2452         int                up;
2453         int                rc;
2454
2455         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2456         if (rc != 0) {
2457                 CERROR("Can't query IPoIB interface %s: %d\n",
2458                        ifname, rc);
2459                 return NULL;
2460         }
2461
2462         if (!up) {
2463                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2464                 return NULL;
2465         }
2466
2467         LIBCFS_ALLOC(dev, sizeof(*dev));
2468         if (dev == NULL)
2469                 return NULL;
2470
2471         memset(dev, 0, sizeof(*dev));
2472         netdev = dev_get_by_name(&init_net, ifname);
2473         if (netdev == NULL) {
2474                 dev->ibd_can_failover = 0;
2475         } else {
2476                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2477                 dev_put(netdev);
2478         }
2479
2480         INIT_LIST_HEAD(&dev->ibd_nets);
2481         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2482         INIT_LIST_HEAD(&dev->ibd_fail_list);
2483         dev->ibd_ifip = ip;
2484         strcpy(&dev->ibd_ifname[0], ifname);
2485
2486         /* initialize the device */
2487         rc = kiblnd_dev_failover(dev);
2488         if (rc != 0) {
2489                 CERROR("Can't initialize device: %d\n", rc);
2490                 LIBCFS_FREE(dev, sizeof(*dev));
2491                 return NULL;
2492         }
2493
2494         list_add_tail(&dev->ibd_list,
2495                           &kiblnd_data.kib_devs);
2496         return dev;
2497 }
2498
2499 static void
2500 kiblnd_base_shutdown(void)
2501 {
2502         struct kib_sched_info   *sched;
2503         int                     i;
2504
2505         LASSERT(list_empty(&kiblnd_data.kib_devs));
2506
2507         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2508                atomic_read(&libcfs_kmemory));
2509
2510         switch (kiblnd_data.kib_init) {
2511         default:
2512                 LBUG();
2513
2514         case IBLND_INIT_ALL:
2515         case IBLND_INIT_DATA:
2516                 LASSERT (kiblnd_data.kib_peers != NULL);
2517                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2518                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2519                 }
2520                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2521                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2522
2523                 /* flag threads to terminate; wake and wait for them to die */
2524                 kiblnd_data.kib_shutdown = 1;
2525
2526                 /* NB: we really want to stop scheduler threads net by net
2527                  * instead of the whole module, this should be improved
2528                  * with dynamic configuration LNet */
2529                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2530                         wake_up_all(&sched->ibs_waitq);
2531
2532                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2533                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2534
2535                 i = 2;
2536                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2537                         i++;
2538                         /* power of 2? */
2539                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2540                                "Waiting for %d threads to terminate\n",
2541                                atomic_read(&kiblnd_data.kib_nthreads));
2542                         set_current_state(TASK_UNINTERRUPTIBLE);
2543                         schedule_timeout(cfs_time_seconds(1));
2544                 }
2545
2546                 /* fall through */
2547
2548         case IBLND_INIT_NOTHING:
2549                 break;
2550         }
2551
2552         if (kiblnd_data.kib_peers != NULL) {
2553                 LIBCFS_FREE(kiblnd_data.kib_peers,
2554                             sizeof(struct list_head) *
2555                             kiblnd_data.kib_peer_hash_size);
2556         }
2557
2558         if (kiblnd_data.kib_scheds != NULL)
2559                 cfs_percpt_free(kiblnd_data.kib_scheds);
2560
2561         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2562                atomic_read(&libcfs_kmemory));
2563
2564         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2565         module_put(THIS_MODULE);
2566 }
2567
2568 static void
2569 kiblnd_shutdown (lnet_ni_t *ni)
2570 {
2571         kib_net_t        *net = ni->ni_data;
2572         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2573         int               i;
2574         unsigned long     flags;
2575
2576         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2577
2578         if (net == NULL)
2579                 goto out;
2580
2581         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2582                atomic_read(&libcfs_kmemory));
2583
2584         write_lock_irqsave(g_lock, flags);
2585         net->ibn_shutdown = 1;
2586         write_unlock_irqrestore(g_lock, flags);
2587
2588         switch (net->ibn_init) {
2589         default:
2590                 LBUG();
2591
2592         case IBLND_INIT_ALL:
2593                 /* nuke all existing peers within this net */
2594                 kiblnd_del_peer(ni, LNET_NID_ANY);
2595
2596                 /* Wait for all peer state to clean up */
2597                 i = 2;
2598                 while (atomic_read(&net->ibn_npeers) != 0) {
2599                         i++;
2600                         /* power of 2? */
2601                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2602                                "%s: waiting for %d peers to disconnect\n",
2603                                libcfs_nid2str(ni->ni_nid),
2604                                atomic_read(&net->ibn_npeers));
2605                         set_current_state(TASK_UNINTERRUPTIBLE);
2606                         schedule_timeout(cfs_time_seconds(1));
2607                 }
2608
2609                 kiblnd_net_fini_pools(net);
2610
2611                 write_lock_irqsave(g_lock, flags);
2612                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2613                 net->ibn_dev->ibd_nnets--;
2614                 list_del(&net->ibn_list);
2615                 write_unlock_irqrestore(g_lock, flags);
2616
2617                 /* fall through */
2618
2619         case IBLND_INIT_NOTHING:
2620                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2621
2622                 if (net->ibn_dev != NULL &&
2623                     net->ibn_dev->ibd_nnets == 0)
2624                         kiblnd_destroy_dev(net->ibn_dev);
2625
2626                 break;
2627         }
2628
2629         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2630                atomic_read(&libcfs_kmemory));
2631
2632         net->ibn_init = IBLND_INIT_NOTHING;
2633         ni->ni_data = NULL;
2634
2635         LIBCFS_FREE(net, sizeof(*net));
2636
2637 out:
2638         if (list_empty(&kiblnd_data.kib_devs))
2639                 kiblnd_base_shutdown();
2640         return;
2641 }
2642
2643 static int
2644 kiblnd_base_startup(void)
2645 {
2646         struct kib_sched_info   *sched;
2647         int                     rc;
2648         int                     i;
2649
2650         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2651
2652         try_module_get(THIS_MODULE);
2653         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2654
2655         rwlock_init(&kiblnd_data.kib_global_lock);
2656
2657         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2658         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2659
2660         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2661         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2662                      sizeof(struct list_head) *
2663                      kiblnd_data.kib_peer_hash_size);
2664         if (kiblnd_data.kib_peers == NULL)
2665                 goto failed;
2666
2667         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2668                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2669
2670         spin_lock_init(&kiblnd_data.kib_connd_lock);
2671         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2672         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2673         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2674         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2675
2676         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2677                                                   sizeof(*sched));
2678         if (kiblnd_data.kib_scheds == NULL)
2679                 goto failed;
2680
2681         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2682                 int     nthrs;
2683
2684                 spin_lock_init(&sched->ibs_lock);
2685                 INIT_LIST_HEAD(&sched->ibs_conns);
2686                 init_waitqueue_head(&sched->ibs_waitq);
2687
2688                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2689                 if (*kiblnd_tunables.kib_nscheds > 0) {
2690                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2691                 } else {
2692                         /* max to half of CPUs, another half is reserved for
2693                          * upper layer modules */
2694                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2695                 }
2696
2697                 sched->ibs_nthreads_max = nthrs;
2698                 sched->ibs_cpt = i;
2699         }
2700
2701         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2702
2703         /* lists/ptrs/locks initialised */
2704         kiblnd_data.kib_init = IBLND_INIT_DATA;
2705         /*****************************************************/
2706
2707         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2708         if (rc != 0) {
2709                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2710                 goto failed;
2711         }
2712
2713         if (*kiblnd_tunables.kib_dev_failover != 0)
2714                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2715                                          "kiblnd_failover");
2716
2717         if (rc != 0) {
2718                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2719                 goto failed;
2720         }
2721
2722         /* flag everything initialised */
2723         kiblnd_data.kib_init = IBLND_INIT_ALL;
2724         /*****************************************************/
2725
2726         return 0;
2727
2728  failed:
2729         kiblnd_base_shutdown();
2730         return -ENETDOWN;
2731 }
2732
2733 static int
2734 kiblnd_start_schedulers(struct kib_sched_info *sched)
2735 {
2736         int     rc = 0;
2737         int     nthrs;
2738         int     i;
2739
2740         if (sched->ibs_nthreads == 0) {
2741                 if (*kiblnd_tunables.kib_nscheds > 0) {
2742                         nthrs = sched->ibs_nthreads_max;
2743                 } else {
2744                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2745                                                sched->ibs_cpt);
2746                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2747                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2748                 }
2749         } else {
2750                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2751                 /* increase one thread if there is new interface */
2752                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2753         }
2754
2755         for (i = 0; i < nthrs; i++) {
2756                 long    id;
2757                 char    name[20];
2758                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2759                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2760                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2761                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2762                 if (rc == 0)
2763                         continue;
2764
2765                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2766                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2767                 break;
2768         }
2769
2770         sched->ibs_nthreads += i;
2771         return rc;
2772 }
2773
2774 static int
2775 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
2776 {
2777         int     cpt;
2778         int     rc;
2779         int     i;
2780
2781         for (i = 0; i < ncpts; i++) {
2782                 struct kib_sched_info *sched;
2783
2784                 cpt = (cpts == NULL) ? i : cpts[i];
2785                 sched = kiblnd_data.kib_scheds[cpt];
2786
2787                 if (!newdev && sched->ibs_nthreads > 0)
2788                         continue;
2789
2790                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2791                 if (rc != 0) {
2792                         CERROR("Failed to start scheduler threads for %s\n",
2793                                dev->ibd_ifname);
2794                         return rc;
2795                 }
2796         }
2797         return 0;
2798 }
2799
2800 static kib_dev_t *
2801 kiblnd_dev_search(char *ifname)
2802 {
2803         kib_dev_t       *alias = NULL;
2804         kib_dev_t       *dev;
2805         char            *colon;
2806         char            *colon2;
2807
2808         colon = strchr(ifname, ':');
2809         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2810                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2811                         return dev;
2812
2813                 if (alias != NULL)
2814                         continue;
2815
2816                 colon2 = strchr(dev->ibd_ifname, ':');
2817                 if (colon != NULL)
2818                         *colon = 0;
2819                 if (colon2 != NULL)
2820                         *colon2 = 0;
2821
2822                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2823                         alias = dev;
2824
2825                 if (colon != NULL)
2826                         *colon = ':';
2827                 if (colon2 != NULL)
2828                         *colon2 = ':';
2829         }
2830         return alias;
2831 }
2832
2833 static int
2834 kiblnd_startup (lnet_ni_t *ni)
2835 {
2836         char                     *ifname;
2837         kib_dev_t                *ibdev = NULL;
2838         kib_net_t                *net;
2839         struct timeval            tv;
2840         unsigned long             flags;
2841         int                       rc;
2842         int                       newdev;
2843
2844         LASSERT (ni->ni_lnd == &the_o2iblnd);
2845
2846         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2847                 rc = kiblnd_base_startup();
2848                 if (rc != 0)
2849                         return rc;
2850         }
2851
2852         LIBCFS_ALLOC(net, sizeof(*net));
2853         ni->ni_data = net;
2854         if (net == NULL)
2855                 goto failed;
2856
2857         memset(net, 0, sizeof(*net));
2858
2859         do_gettimeofday(&tv);
2860         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2861
2862         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2863         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2864         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2865         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2866
2867         if (ni->ni_interfaces[0] != NULL) {
2868                 /* Use the IPoIB interface specified in 'networks=' */
2869
2870                 CLASSERT (LNET_MAX_INTERFACES > 1);
2871                 if (ni->ni_interfaces[1] != NULL) {
2872                         CERROR("Multiple interfaces not supported\n");
2873                         goto failed;
2874                 }
2875
2876                 ifname = ni->ni_interfaces[0];
2877         } else {
2878                 ifname = *kiblnd_tunables.kib_default_ipif;
2879         }
2880
2881         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2882                 CERROR("IPoIB interface name too long: %s\n", ifname);
2883                 goto failed;
2884         }
2885
2886         ibdev = kiblnd_dev_search(ifname);
2887
2888         newdev = ibdev == NULL;
2889         /* hmm...create kib_dev even for alias */
2890         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
2891                 ibdev = kiblnd_create_dev(ifname);
2892
2893         if (ibdev == NULL)
2894                 goto failed;
2895
2896         net->ibn_dev = ibdev;
2897         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2898
2899         rc = kiblnd_dev_start_threads(ibdev, newdev,
2900                                       ni->ni_cpts, ni->ni_ncpts);
2901         if (rc != 0)
2902                 goto failed;
2903
2904         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2905         if (rc != 0) {
2906                 CERROR("Failed to initialize NI pools: %d\n", rc);
2907                 goto failed;
2908         }
2909
2910         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2911         ibdev->ibd_nnets++;
2912         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2913         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2914
2915         net->ibn_init = IBLND_INIT_ALL;
2916
2917         return 0;
2918
2919 failed:
2920         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
2921                 kiblnd_destroy_dev(ibdev);
2922
2923         kiblnd_shutdown(ni);
2924
2925         CDEBUG(D_NET, "kiblnd_startup failed\n");
2926         return -ENETDOWN;
2927 }
2928
2929 static lnd_t the_o2iblnd = {
2930         .lnd_type       = O2IBLND,
2931         .lnd_startup    = kiblnd_startup,
2932         .lnd_shutdown   = kiblnd_shutdown,
2933         .lnd_ctl        = kiblnd_ctl,
2934         .lnd_query      = kiblnd_query,
2935         .lnd_send       = kiblnd_send,
2936         .lnd_recv       = kiblnd_recv,
2937 };
2938
2939 static void __exit ko2iblnd_exit(void)
2940 {
2941         lnet_unregister_lnd(&the_o2iblnd);
2942         kiblnd_tunables_fini();
2943 }
2944
2945 static int __init ko2iblnd_init(void)
2946 {
2947         int rc;
2948
2949         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2950         CLASSERT(offsetof(kib_msg_t,
2951                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
2952                  IBLND_MSG_SIZE);
2953         CLASSERT(offsetof(kib_msg_t,
2954                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2955                  <= IBLND_MSG_SIZE);
2956
2957         rc = kiblnd_tunables_init();
2958         if (rc != 0)
2959                 return rc;
2960
2961         lnet_register_lnd(&the_o2iblnd);
2962
2963         return 0;
2964 }
2965
2966 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2967 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
2968 MODULE_VERSION("2.8.0");
2969 MODULE_LICENSE("GPL");
2970
2971 module_init(ko2iblnd_init);
2972 module_exit(ko2iblnd_exit);