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