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, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/page.h>
  42 #include "o2iblnd.h"
  43
  44 static lnd_t the_o2iblnd;
  45
  46 kib_data_t              kiblnd_data;
  47
  48 static __u32
  49 kiblnd_cksum (void *ptr, int nob)
  50 {
  51         char  *c  = ptr;
  52         __u32  sum = 0;
  53
  54         while (nob-- > 0)
  55                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  56
  57         /* ensure I don't return 0 (== no checksum) */
  58         return (sum == 0) ? 1 : sum;
  59 }
  60
  61 static char *
  62 kiblnd_msgtype2str(int type)
  63 {
  64         switch (type) {
  65         case IBLND_MSG_CONNREQ:
  66                 return "CONNREQ";
  67
  68         case IBLND_MSG_CONNACK:
  69                 return "CONNACK";
  70
  71         case IBLND_MSG_NOOP:
  72                 return "NOOP";
  73
  74         case IBLND_MSG_IMMEDIATE:
  75                 return "IMMEDIATE";
  76
  77         case IBLND_MSG_PUT_REQ:
  78                 return "PUT_REQ";
  79
  80         case IBLND_MSG_PUT_NAK:
  81                 return "PUT_NAK";
  82
  83         case IBLND_MSG_PUT_ACK:
  84                 return "PUT_ACK";
  85
  86         case IBLND_MSG_PUT_DONE:
  87                 return "PUT_DONE";
  88
  89         case IBLND_MSG_GET_REQ:
  90                 return "GET_REQ";
  91
  92         case IBLND_MSG_GET_DONE:
  93                 return "GET_DONE";
  94
  95         default:
  96                 return "???";
  97         }
  98 }
  99
 100 static int
 101 kiblnd_msgtype2size(int type)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104
 105         switch (type) {
 106         case IBLND_MSG_CONNREQ:
 107         case IBLND_MSG_CONNACK:
 108                 return hdr_size + sizeof(kib_connparams_t);
 109
 110         case IBLND_MSG_NOOP:
 111                 return hdr_size;
 112
 113         case IBLND_MSG_IMMEDIATE:
 114                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 115
 116         case IBLND_MSG_PUT_REQ:
 117                 return hdr_size + sizeof(kib_putreq_msg_t);
 118
 119         case IBLND_MSG_PUT_ACK:
 120                 return hdr_size + sizeof(kib_putack_msg_t);
 121
 122         case IBLND_MSG_GET_REQ:
 123                 return hdr_size + sizeof(kib_get_msg_t);
 124
 125         case IBLND_MSG_PUT_NAK:
 126         case IBLND_MSG_PUT_DONE:
 127         case IBLND_MSG_GET_DONE:
 128                 return hdr_size + sizeof(kib_completion_msg_t);
 129         default:
 130                 return -1;
 131         }
 132 }
 133
 134 static int
 135 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 136 {
 137         kib_rdma_desc_t   *rd;
 138         int                nob;
 139         int                n;
 140         int                i;
 141
 142         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 143                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 144
 145         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 146                               &msg->ibm_u.get.ibgm_rd :
 147                               &msg->ibm_u.putack.ibpam_rd;
 148
 149         if (flip) {
 150                 __swab32s(&rd->rd_key);
 151                 __swab32s(&rd->rd_nfrags);
 152         }
 153
 154         n = rd->rd_nfrags;
 155
 156         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 157                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 158                        n, IBLND_MAX_RDMA_FRAGS);
 159                 return 1;
 160         }
 161
 162         nob = offsetof (kib_msg_t, ibm_u) +
 163               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 164
 165         if (msg->ibm_nob < nob) {
 166                 CERROR("Short %s: %d(%d)\n",
 167                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 168                 return 1;
 169         }
 170
 171         if (!flip)
 172                 return 0;
 173
 174         for (i = 0; i < n; i++) {
 175                 __swab32s(&rd->rd_frags[i].rf_nob);
 176                 __swab64s(&rd->rd_frags[i].rf_addr);
 177         }
 178
 179         return 0;
 180 }
 181
 182 void
 183 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 184                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 185 {
 186         kib_net_t *net = ni->ni_data;
 187
 188         /* CAVEAT EMPTOR! all message fields not set here should have been
 189          * initialised previously. */
 190         msg->ibm_magic    = IBLND_MSG_MAGIC;
 191         msg->ibm_version  = version;
 192         /*   ibm_type */
 193         msg->ibm_credits  = credits;
 194         /*   ibm_nob */
 195         msg->ibm_cksum    = 0;
 196         msg->ibm_srcnid   = ni->ni_nid;
 197         msg->ibm_srcstamp = net->ibn_incarnation;
 198         msg->ibm_dstnid   = dstnid;
 199         msg->ibm_dststamp = dststamp;
 200
 201         if (*kiblnd_tunables.kib_cksum) {
 202                 /* NB ibm_cksum zero while computing cksum */
 203                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 204         }
 205 }
 206
 207 int
 208 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 209 {
 210         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 211         __u32     msg_cksum;
 212         __u16     version;
 213         int       msg_nob;
 214         int       flip;
 215
 216         /* 6 bytes are enough to have received magic + version */
 217         if (nob < 6) {
 218                 CERROR("Short message: %d\n", nob);
 219                 return -EPROTO;
 220         }
 221
 222         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 223                 flip = 0;
 224         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 225                 flip = 1;
 226         } else {
 227                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 228                 return -EPROTO;
 229         }
 230
 231         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 232         if (version != IBLND_MSG_VERSION &&
 233             version != IBLND_MSG_VERSION_1) {
 234                 CERROR("Bad version: %x\n", version);
 235                 return -EPROTO;
 236         }
 237
 238         if (nob < hdr_size) {
 239                 CERROR("Short message: %d\n", nob);
 240                 return -EPROTO;
 241         }
 242
 243         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 244         if (msg_nob > nob) {
 245                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 246                 return -EPROTO;
 247         }
 248
 249         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 250          * gets flipped */
 251         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 252         msg->ibm_cksum = 0;
 253         if (msg_cksum != 0 &&
 254             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 255                 CERROR("Bad checksum\n");
 256                 return -EPROTO;
 257         }
 258
 259         msg->ibm_cksum = msg_cksum;
 260
 261         if (flip) {
 262                 /* leave magic unflipped as a clue to peer endianness */
 263                 msg->ibm_version = version;
 264                 CLASSERT (sizeof(msg->ibm_type) == 1);
 265                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 266                 msg->ibm_nob     = msg_nob;
 267                 __swab64s(&msg->ibm_srcnid);
 268                 __swab64s(&msg->ibm_srcstamp);
 269                 __swab64s(&msg->ibm_dstnid);
 270                 __swab64s(&msg->ibm_dststamp);
 271         }
 272
 273         if (msg->ibm_srcnid == LNET_NID_ANY) {
 274                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 275                 return -EPROTO;
 276         }
 277
 278         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 279                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 280                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 281                 return -EPROTO;
 282         }
 283
 284         switch (msg->ibm_type) {
 285         default:
 286                 CERROR("Unknown message type %x\n", msg->ibm_type);
 287                 return -EPROTO;
 288
 289         case IBLND_MSG_NOOP:
 290         case IBLND_MSG_IMMEDIATE:
 291         case IBLND_MSG_PUT_REQ:
 292                 break;
 293
 294         case IBLND_MSG_PUT_ACK:
 295         case IBLND_MSG_GET_REQ:
 296                 if (kiblnd_unpack_rd(msg, flip))
 297                         return -EPROTO;
 298                 break;
 299
 300         case IBLND_MSG_PUT_NAK:
 301         case IBLND_MSG_PUT_DONE:
 302         case IBLND_MSG_GET_DONE:
 303                 if (flip)
 304                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 305                 break;
 306
 307         case IBLND_MSG_CONNREQ:
 308         case IBLND_MSG_CONNACK:
 309                 if (flip) {
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 311                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 312                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 313                 }
 314                 break;
 315         }
 316         return 0;
 317 }
 318
 319 int
 320 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 321 {
 322         kib_peer_t      *peer;
 323         kib_net_t       *net = ni->ni_data;
 324         int             cpt = lnet_cpt_of_nid(nid);
 325         unsigned long   flags;
 326
 327         LASSERT(net != NULL);
 328         LASSERT(nid != LNET_NID_ANY);
 329
 330         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 331         if (peer == NULL) {
 332                 CERROR("Cannot allocate peer\n");
 333                 return -ENOMEM;
 334         }
 335
 336         peer->ibp_ni = ni;
 337         peer->ibp_nid = nid;
 338         peer->ibp_error = 0;
 339         peer->ibp_last_alive = 0;
 340         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 341
 342         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 343         INIT_LIST_HEAD(&peer->ibp_conns);
 344         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 345
 346         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 347
 348         /* always called with a ref on ni, which prevents ni being shutdown */
 349         LASSERT(net->ibn_shutdown == 0);
 350
 351         /* npeers only grows with the global lock held */
 352         atomic_inc(&net->ibn_npeers);
 353
 354         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 355
 356         *peerp = peer;
 357         return 0;
 358 }
 359
 360 void
 361 kiblnd_destroy_peer (kib_peer_t *peer)
 362 {
 363         kib_net_t *net = peer->ibp_ni->ni_data;
 364
 365         LASSERT(net != NULL);
 366         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 367         LASSERT(!kiblnd_peer_active(peer));
 368         LASSERT(peer->ibp_connecting == 0);
 369         LASSERT(peer->ibp_accepting == 0);
 370         LASSERT(list_empty(&peer->ibp_conns));
 371         LASSERT(list_empty(&peer->ibp_tx_queue));
 372
 373         LIBCFS_FREE(peer, sizeof(*peer));
 374
 375         /* NB a peer's connections keep a reference on their peer until
 376          * they are destroyed, so we can be assured that _all_ state to do
 377          * with this peer has been cleaned up when its refcount drops to
 378          * zero. */
 379         atomic_dec(&net->ibn_npeers);
 380 }
 381
 382 kib_peer_t *
 383 kiblnd_find_peer_locked (lnet_nid_t nid)
 384 {
 385         /* the caller is responsible for accounting the additional reference
 386          * that this creates */
 387         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 388         struct list_head        *tmp;
 389         kib_peer_t              *peer;
 390
 391         list_for_each(tmp, peer_list) {
 392
 393                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 394
 395                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 396                         peer->ibp_accepting > 0 ||
 397                         !list_empty(&peer->ibp_conns)); /* active conn */
 398
 399                 if (peer->ibp_nid != nid)
 400                         continue;
 401
 402                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 403                        peer, libcfs_nid2str(nid),
 404                        atomic_read(&peer->ibp_refcount),
 405                        peer->ibp_version);
 406                 return peer;
 407         }
 408         return NULL;
 409 }
 410
 411 void
 412 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 413 {
 414         LASSERT(list_empty(&peer->ibp_conns));
 415
 416         LASSERT (kiblnd_peer_active(peer));
 417         list_del_init(&peer->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer);
 420 }
 421
 422 static int
 423 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 424                      lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_t              *peer;
 427         struct list_head        *ptmp;
 428         int                      i;
 429         unsigned long            flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 438                         LASSERT(peer->ibp_connecting > 0 ||
 439                                 peer->ibp_accepting > 0 ||
 440                                 !list_empty(&peer->ibp_conns));
 441
 442                         if (peer->ibp_ni != ni)
 443                                 continue;
 444
 445                         if (index-- > 0)
 446                                 continue;
 447
 448                         *nidp = peer->ibp_nid;
 449                         *count = atomic_read(&peer->ibp_refcount);
 450
 451                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 452                                                flags);
 453                         return 0;
 454                 }
 455         }
 456
 457         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 458         return -ENOENT;
 459 }
 460
 461 static void
 462 kiblnd_del_peer_locked (kib_peer_t *peer)
 463 {
 464         struct list_head        *ctmp;
 465         struct list_head        *cnxt;
 466         kib_conn_t              *conn;
 467
 468         if (list_empty(&peer->ibp_conns)) {
 469                 kiblnd_unlink_peer_locked(peer);
 470         } else {
 471                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 472                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 473
 474                         kiblnd_close_conn_locked(conn, 0);
 475                 }
 476                 /* NB closing peer's last conn unlinked it. */
 477         }
 478         /* NB peer now unlinked; might even be freed if the peer table had the
 479          * last ref on it. */
 480 }
 481
 482 static int
 483 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 484 {
 485         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 486         struct list_head        *ptmp;
 487         struct list_head        *pnxt;
 488         kib_peer_t              *peer;
 489         int                     lo;
 490         int                     hi;
 491         int                     i;
 492         unsigned long           flags;
 493         int                     rc = -ENOENT;
 494
 495         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 496
 497         if (nid != LNET_NID_ANY) {
 498                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 499         } else {
 500                 lo = 0;
 501                 hi = kiblnd_data.kib_peer_hash_size - 1;
 502         }
 503
 504         for (i = lo; i <= hi; i++) {
 505                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 506                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 507                         LASSERT(peer->ibp_connecting > 0 ||
 508                                 peer->ibp_accepting > 0 ||
 509                                 !list_empty(&peer->ibp_conns));
 510
 511                         if (peer->ibp_ni != ni)
 512                                 continue;
 513
 514                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 515                                 continue;
 516
 517                         if (!list_empty(&peer->ibp_tx_queue)) {
 518                                 LASSERT(list_empty(&peer->ibp_conns));
 519
 520                                 list_splice_init(&peer->ibp_tx_queue,
 521                                                  &zombies);
 522                         }
 523
 524                         kiblnd_del_peer_locked(peer);
 525                         rc = 0;         /* matched something */
 526                 }
 527         }
 528
 529         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 530
 531         kiblnd_txlist_done(ni, &zombies, -EIO);
 532
 533         return rc;
 534 }
 535
 536 static kib_conn_t *
 537 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 538 {
 539         kib_peer_t              *peer;
 540         struct list_head        *ptmp;
 541         kib_conn_t              *conn;
 542         struct list_head        *ctmp;
 543         int                     i;
 544         unsigned long           flags;
 545
 546         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 547
 548         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 549                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 550
 551                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 552                         LASSERT(peer->ibp_connecting > 0 ||
 553                                 peer->ibp_accepting > 0 ||
 554                                 !list_empty(&peer->ibp_conns));
 555
 556                         if (peer->ibp_ni != ni)
 557                                 continue;
 558
 559                         list_for_each(ctmp, &peer->ibp_conns) {
 560                                 if (index-- > 0)
 561                                         continue;
 562
 563                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 564                                 kiblnd_conn_addref(conn);
 565                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 566                                                        flags);
 567                                 return conn;
 568                         }
 569                 }
 570         }
 571
 572         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 573         return NULL;
 574 }
 575
 576 static void
 577 kiblnd_debug_rx (kib_rx_t *rx)
 578 {
 579         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 580                rx, rx->rx_status, rx->rx_msg->ibm_type,
 581                rx->rx_msg->ibm_credits);
 582 }
 583
 584 static void
 585 kiblnd_debug_tx (kib_tx_t *tx)
 586 {
 587         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 588                "cookie "LPX64" msg %s%s type %x cred %d\n",
 589                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 590                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 591                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 592                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 593                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 594 }
 595
 596 void
 597 kiblnd_debug_conn (kib_conn_t *conn)
 598 {
 599         struct list_head        *tmp;
 600         int                     i;
 601
 602         spin_lock(&conn->ibc_lock);
 603
 604         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 605                atomic_read(&conn->ibc_refcount), conn,
 606                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 607         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 608                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 609                conn->ibc_nsends_posted, conn->ibc_credits,
 610                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 611         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 612
 613         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 614         list_for_each(tmp, &conn->ibc_early_rxs)
 615                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 616
 617         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 618         list_for_each(tmp, &conn->ibc_tx_noops)
 619                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 620
 621         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 622         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 623                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 624
 625         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 626         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 627                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 628
 629         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 630         list_for_each(tmp, &conn->ibc_tx_queue)
 631                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 632
 633         CDEBUG(D_CONSOLE, "   active_txs:\n");
 634         list_for_each(tmp, &conn->ibc_active_txs)
 635                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 636
 637         CDEBUG(D_CONSOLE, "   rxs:\n");
 638         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 639                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 640
 641         spin_unlock(&conn->ibc_lock);
 642 }
 643
 644 int
 645 kiblnd_translate_mtu(int value)
 646 {
 647         switch (value) {
 648         default:
 649                 return -1;
 650         case 0:
 651                 return 0;
 652         case 256:
 653                 return IB_MTU_256;
 654         case 512:
 655                 return IB_MTU_512;
 656         case 1024:
 657                 return IB_MTU_1024;
 658         case 2048:
 659                 return IB_MTU_2048;
 660         case 4096:
 661                 return IB_MTU_4096;
 662         }
 663 }
 664
 665 static void
 666 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 667 {
 668         int           mtu;
 669
 670         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 671         if (cmid->route.path_rec == NULL)
 672                 return;
 673
 674         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 675         LASSERT (mtu >= 0);
 676         if (mtu != 0)
 677                 cmid->route.path_rec->mtu = mtu;
 678 }
 679
 680 static int
 681 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 682 {
 683         cpumask_t       *mask;
 684         int             vectors;
 685         int             off;
 686         int             i;
 687
 688         vectors = conn->ibc_cmid->device->num_comp_vectors;
 689         if (vectors <= 1)
 690                 return 0;
 691
 692         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 693
 694         /* hash NID to CPU id in this partition... */
 695         off = conn->ibc_peer->ibp_nid % cpumask_weight(mask);
 696         for_each_cpu(i, mask) {
 697                 if (off-- == 0)
 698                         return i % vectors;
 699         }
 700
 701         LBUG();
 702         return 1;
 703 }
 704
 705 kib_conn_t *
 706 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 707                    int state, int version)
 708 {
 709         /* CAVEAT EMPTOR:
 710          * If the new conn is created successfully it takes over the caller's
 711          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 712          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 713          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 714          * to destroy 'cmid' here since I'm called from the CM which still has
 715          * its ref on 'cmid'). */
 716         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 717         kib_net_t              *net = peer->ibp_ni->ni_data;
 718         kib_dev_t              *dev;
 719         struct ib_qp_init_attr *init_qp_attr;
 720         struct kib_sched_info   *sched;
 721 #ifdef HAVE_IB_CQ_INIT_ATTR
 722         struct ib_cq_init_attr  cq_attr = {};
 723 #endif
 724         kib_conn_t              *conn;
 725         struct ib_cq            *cq;
 726         unsigned long           flags;
 727         int                     cpt;
 728         int                     rc;
 729         int                     i;
 730
 731         LASSERT(net != NULL);
 732         LASSERT(!in_interrupt());
 733
 734         dev = net->ibn_dev;
 735
 736         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 737         sched = kiblnd_data.kib_scheds[cpt];
 738
 739         LASSERT(sched->ibs_nthreads > 0);
 740
 741         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 742                          sizeof(*init_qp_attr));
 743         if (init_qp_attr == NULL) {
 744                 CERROR("Can't allocate qp_attr for %s\n",
 745                        libcfs_nid2str(peer->ibp_nid));
 746                 goto failed_0;
 747         }
 748
 749         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 750         if (conn == NULL) {
 751                 CERROR("Can't allocate connection for %s\n",
 752                        libcfs_nid2str(peer->ibp_nid));
 753                 goto failed_1;
 754         }
 755
 756         conn->ibc_state = IBLND_CONN_INIT;
 757         conn->ibc_version = version;
 758         conn->ibc_peer = peer;                  /* I take the caller's ref */
 759         cmid->context = conn;                   /* for future CM callbacks */
 760         conn->ibc_cmid = cmid;
 761
 762         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 763         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 764         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 765         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 766         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 767         INIT_LIST_HEAD(&conn->ibc_active_txs);
 768         spin_lock_init(&conn->ibc_lock);
 769
 770         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 771                          sizeof(*conn->ibc_connvars));
 772         if (conn->ibc_connvars == NULL) {
 773                 CERROR("Can't allocate in-progress connection state\n");
 774                 goto failed_2;
 775         }
 776
 777         write_lock_irqsave(glock, flags);
 778         if (dev->ibd_failover) {
 779                 write_unlock_irqrestore(glock, flags);
 780                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 781                 goto failed_2;
 782         }
 783
 784         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 785                 /* wakeup failover thread and teardown connection */
 786                 if (kiblnd_dev_can_failover(dev)) {
 787                         list_add_tail(&dev->ibd_fail_list,
 788                                       &kiblnd_data.kib_failed_devs);
 789                         wake_up(&kiblnd_data.kib_failover_waitq);
 790                 }
 791
 792                 write_unlock_irqrestore(glock, flags);
 793                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 794                        cmid->device->name, dev->ibd_ifname);
 795                 goto failed_2;
 796         }
 797
 798         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 799         conn->ibc_hdev = dev->ibd_hdev;
 800
 801         kiblnd_setup_mtu_locked(cmid);
 802
 803         write_unlock_irqrestore(glock, flags);
 804
 805         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 806                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 807         if (conn->ibc_rxs == NULL) {
 808                 CERROR("Cannot allocate RX buffers\n");
 809                 goto failed_2;
 810         }
 811
 812         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 813                                 IBLND_RX_MSG_PAGES(version));
 814         if (rc != 0)
 815                 goto failed_2;
 816
 817         kiblnd_map_rx_descs(conn);
 818
 819 #ifdef HAVE_IB_CQ_INIT_ATTR
 820         cq_attr.cqe = IBLND_CQ_ENTRIES(version);
 821         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 822         cq = ib_create_cq(cmid->device,
 823                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 824                           &cq_attr);
 825 #else
 826         cq = ib_create_cq(cmid->device,
 827                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 828                           IBLND_CQ_ENTRIES(version),
 829                           kiblnd_get_completion_vector(conn, cpt));
 830 #endif
 831         if (IS_ERR(cq)) {
 832                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 833                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 834                 goto failed_2;
 835         }
 836
 837         conn->ibc_cq = cq;
 838
 839         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 840         if (rc != 0) {
 841                 CERROR("Can't request completion notification: %d\n", rc);
 842                 goto failed_2;
 843         }
 844
 845         init_qp_attr->event_handler = kiblnd_qp_event;
 846         init_qp_attr->qp_context = conn;
 847         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 848         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 849         init_qp_attr->cap.max_send_sge = 1;
 850         init_qp_attr->cap.max_recv_sge = 1;
 851         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 852         init_qp_attr->qp_type = IB_QPT_RC;
 853         init_qp_attr->send_cq = cq;
 854         init_qp_attr->recv_cq = cq;
 855
 856         conn->ibc_sched = sched;
 857
 858         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 859         if (rc != 0) {
 860                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 861                        rc, init_qp_attr->cap.max_send_wr,
 862                        init_qp_attr->cap.max_recv_wr);
 863                 goto failed_2;
 864         }
 865
 866         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 867
 868         /* 1 ref for caller and each rxmsg */
 869         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 870         conn->ibc_nrx = IBLND_RX_MSGS(version);
 871
 872         /* post receives */
 873         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 874                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 875                                     IBLND_POSTRX_NO_CREDIT);
 876                 if (rc != 0) {
 877                         CERROR("Can't post rxmsg: %d\n", rc);
 878
 879                         /* Make posted receives complete */
 880                         kiblnd_abort_receives(conn);
 881
 882                         /* correct # of posted buffers
 883                          * NB locking needed now I'm racing with completion */
 884                         spin_lock_irqsave(&sched->ibs_lock, flags);
 885                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 886                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 887
 888                         /* cmid will be destroyed by CM(ofed) after cm_callback
 889                          * returned, so we can't refer it anymore
 890                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 891                         rdma_destroy_qp(conn->ibc_cmid);
 892                         conn->ibc_cmid = NULL;
 893
 894                         /* Drop my own and unused rxbuffer refcounts */
 895                         while (i++ <= IBLND_RX_MSGS(version))
 896                                 kiblnd_conn_decref(conn);
 897
 898                         return NULL;
 899                 }
 900         }
 901
 902         /* Init successful! */
 903         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 904                  state == IBLND_CONN_PASSIVE_WAIT);
 905         conn->ibc_state = state;
 906
 907         /* 1 more conn */
 908         atomic_inc(&net->ibn_nconns);
 909         return conn;
 910
 911  failed_2:
 912         kiblnd_destroy_conn(conn);
 913  failed_1:
 914         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 915  failed_0:
 916         return NULL;
 917 }
 918
 919 void
 920 kiblnd_destroy_conn (kib_conn_t *conn)
 921 {
 922         struct rdma_cm_id *cmid = conn->ibc_cmid;
 923         kib_peer_t        *peer = conn->ibc_peer;
 924         int                rc;
 925
 926         LASSERT (!in_interrupt());
 927         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 928         LASSERT(list_empty(&conn->ibc_early_rxs));
 929         LASSERT(list_empty(&conn->ibc_tx_noops));
 930         LASSERT(list_empty(&conn->ibc_tx_queue));
 931         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 932         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 933         LASSERT(list_empty(&conn->ibc_active_txs));
 934         LASSERT (conn->ibc_noops_posted == 0);
 935         LASSERT (conn->ibc_nsends_posted == 0);
 936
 937         switch (conn->ibc_state) {
 938         default:
 939                 /* conn must be completely disengaged from the network */
 940                 LBUG();
 941
 942         case IBLND_CONN_DISCONNECTED:
 943                 /* connvars should have been freed already */
 944                 LASSERT (conn->ibc_connvars == NULL);
 945                 break;
 946
 947         case IBLND_CONN_INIT:
 948                 break;
 949         }
 950
 951         /* conn->ibc_cmid might be destroyed by CM already */
 952         if (cmid != NULL && cmid->qp != NULL)
 953                 rdma_destroy_qp(cmid);
 954
 955         if (conn->ibc_cq != NULL) {
 956                 rc = ib_destroy_cq(conn->ibc_cq);
 957                 if (rc != 0)
 958                         CWARN("Error destroying CQ: %d\n", rc);
 959         }
 960
 961         if (conn->ibc_rx_pages != NULL)
 962                 kiblnd_unmap_rx_descs(conn);
 963
 964         if (conn->ibc_rxs != NULL) {
 965                 LIBCFS_FREE(conn->ibc_rxs,
 966                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 967         }
 968
 969         if (conn->ibc_connvars != NULL)
 970                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 971
 972         if (conn->ibc_hdev != NULL)
 973                 kiblnd_hdev_decref(conn->ibc_hdev);
 974
 975         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 976         if (conn->ibc_state != IBLND_CONN_INIT) {
 977                 kib_net_t *net = peer->ibp_ni->ni_data;
 978
 979                 kiblnd_peer_decref(peer);
 980                 rdma_destroy_id(cmid);
 981                 atomic_dec(&net->ibn_nconns);
 982         }
 983
 984         LIBCFS_FREE(conn, sizeof(*conn));
 985 }
 986
 987 int
 988 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 989 {
 990         kib_conn_t              *conn;
 991         struct list_head        *ctmp;
 992         struct list_head        *cnxt;
 993         int                     count = 0;
 994
 995         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 996                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 997
 998                 CDEBUG(D_NET, "Closing conn -> %s, "
 999                               "version: %x, reason: %d\n",
1000                        libcfs_nid2str(peer->ibp_nid),
1001                        conn->ibc_version, why);
1002
1003                 kiblnd_close_conn_locked(conn, why);
1004                 count++;
1005         }
1006
1007         return count;
1008 }
1009
1010 int
1011 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1012                                 int version, __u64 incarnation)
1013 {
1014         kib_conn_t              *conn;
1015         struct list_head        *ctmp;
1016         struct list_head        *cnxt;
1017         int                     count = 0;
1018
1019         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1020                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1021
1022                 if (conn->ibc_version     == version &&
1023                     conn->ibc_incarnation == incarnation)
1024                         continue;
1025
1026                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1027                               "incarnation:"LPX64"(%x, "LPX64")\n",
1028                        libcfs_nid2str(peer->ibp_nid),
1029                        conn->ibc_version, conn->ibc_incarnation,
1030                        version, incarnation);
1031
1032                 kiblnd_close_conn_locked(conn, -ESTALE);
1033                 count++;
1034         }
1035
1036         return count;
1037 }
1038
1039 static int
1040 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1041 {
1042         kib_peer_t              *peer;
1043         struct list_head        *ptmp;
1044         struct list_head        *pnxt;
1045         int                     lo;
1046         int                     hi;
1047         int                     i;
1048         unsigned long           flags;
1049         int                     count = 0;
1050
1051         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1052
1053         if (nid != LNET_NID_ANY)
1054                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1055         else {
1056                 lo = 0;
1057                 hi = kiblnd_data.kib_peer_hash_size - 1;
1058         }
1059
1060         for (i = lo; i <= hi; i++) {
1061                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1062
1063                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1064                         LASSERT(peer->ibp_connecting > 0 ||
1065                                 peer->ibp_accepting > 0 ||
1066                                 !list_empty(&peer->ibp_conns));
1067
1068                         if (peer->ibp_ni != ni)
1069                                 continue;
1070
1071                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1072                                 continue;
1073
1074                         count += kiblnd_close_peer_conns_locked(peer, 0);
1075                 }
1076         }
1077
1078         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1079
1080         /* wildcards always succeed */
1081         if (nid == LNET_NID_ANY)
1082                 return 0;
1083
1084         return (count == 0) ? -ENOENT : 0;
1085 }
1086
1087 static int
1088 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1089 {
1090         struct libcfs_ioctl_data *data = arg;
1091         int                       rc = -EINVAL;
1092
1093         switch(cmd) {
1094         case IOC_LIBCFS_GET_PEER: {
1095                 lnet_nid_t   nid = 0;
1096                 int          count = 0;
1097
1098                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1099                                           &nid, &count);
1100                 data->ioc_nid    = nid;
1101                 data->ioc_count  = count;
1102                 break;
1103         }
1104
1105         case IOC_LIBCFS_DEL_PEER: {
1106                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1107                 break;
1108         }
1109         case IOC_LIBCFS_GET_CONN: {
1110                 kib_conn_t *conn;
1111
1112                 rc = 0;
1113                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1114                 if (conn == NULL) {
1115                         rc = -ENOENT;
1116                         break;
1117                 }
1118
1119                 LASSERT (conn->ibc_cmid != NULL);
1120                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1121                 if (conn->ibc_cmid->route.path_rec == NULL)
1122                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1123                 else
1124                         data->ioc_u32[0] =
1125                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1126                 kiblnd_conn_decref(conn);
1127                 break;
1128         }
1129         case IOC_LIBCFS_CLOSE_CONNECTION: {
1130                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1131                 break;
1132         }
1133
1134         default:
1135                 break;
1136         }
1137
1138         return rc;
1139 }
1140
1141 static void
1142 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1143 {
1144         cfs_time_t      last_alive = 0;
1145         cfs_time_t      now = cfs_time_current();
1146         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1147         kib_peer_t      *peer;
1148         unsigned long   flags;
1149
1150         read_lock_irqsave(glock, flags);
1151
1152         peer = kiblnd_find_peer_locked(nid);
1153         if (peer != NULL) {
1154                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1155                         peer->ibp_accepting > 0 ||
1156                         !list_empty(&peer->ibp_conns));  /* active conn */
1157                 last_alive = peer->ibp_last_alive;
1158         }
1159
1160         read_unlock_irqrestore(glock, flags);
1161
1162         if (last_alive != 0)
1163                 *when = last_alive;
1164
1165         /* peer is not persistent in hash, trigger peer creation
1166          * and connection establishment with a NULL tx */
1167         if (peer == NULL)
1168                 kiblnd_launch_tx(ni, NULL, nid);
1169
1170         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1171                libcfs_nid2str(nid), peer,
1172                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1173         return;
1174 }
1175
1176 static void
1177 kiblnd_free_pages(kib_pages_t *p)
1178 {
1179         int     npages = p->ibp_npages;
1180         int     i;
1181
1182         for (i = 0; i < npages; i++) {
1183                 if (p->ibp_pages[i] != NULL)
1184                         __free_page(p->ibp_pages[i]);
1185         }
1186
1187         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1188 }
1189
1190 int
1191 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1192 {
1193         kib_pages_t     *p;
1194         int             i;
1195
1196         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1197                          offsetof(kib_pages_t, ibp_pages[npages]));
1198         if (p == NULL) {
1199                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1200                 return -ENOMEM;
1201         }
1202
1203         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1204         p->ibp_npages = npages;
1205
1206         for (i = 0; i < npages; i++) {
1207                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1208                                                      GFP_NOFS);
1209                 if (p->ibp_pages[i] == NULL) {
1210                         CERROR("Can't allocate page %d of %d\n", i, npages);
1211                         kiblnd_free_pages(p);
1212                         return -ENOMEM;
1213                 }
1214         }
1215
1216         *pp = p;
1217         return 0;
1218 }
1219
1220 void
1221 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1222 {
1223         kib_rx_t *rx;
1224         int       i;
1225
1226         LASSERT (conn->ibc_rxs != NULL);
1227         LASSERT (conn->ibc_hdev != NULL);
1228
1229         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1230                 rx = &conn->ibc_rxs[i];
1231
1232                 LASSERT (rx->rx_nob >= 0); /* not posted */
1233
1234                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1235                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1236                                                           rx->rx_msgaddr),
1237                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1238         }
1239
1240         kiblnd_free_pages(conn->ibc_rx_pages);
1241
1242         conn->ibc_rx_pages = NULL;
1243 }
1244
1245 void
1246 kiblnd_map_rx_descs(kib_conn_t *conn)
1247 {
1248         kib_rx_t       *rx;
1249         struct page    *pg;
1250         int             pg_off;
1251         int             ipg;
1252         int             i;
1253
1254         for (pg_off = ipg = i = 0;
1255              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1256                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1257                 rx = &conn->ibc_rxs[i];
1258
1259                 rx->rx_conn = conn;
1260                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1261
1262                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1263                                                        rx->rx_msg, IBLND_MSG_SIZE,
1264                                                        DMA_FROM_DEVICE);
1265                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1266                                                    rx->rx_msgaddr));
1267                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1268
1269                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1270                        i, rx->rx_msg, rx->rx_msgaddr,
1271                        (__u64)(page_to_phys(pg) + pg_off));
1272
1273                 pg_off += IBLND_MSG_SIZE;
1274                 LASSERT (pg_off <= PAGE_SIZE);
1275
1276                 if (pg_off == PAGE_SIZE) {
1277                         pg_off = 0;
1278                         ipg++;
1279                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1280                 }
1281         }
1282 }
1283
1284 static void
1285 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1286 {
1287         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1288         kib_tx_t       *tx;
1289         int             i;
1290
1291         LASSERT (tpo->tpo_pool.po_allocated == 0);
1292
1293         if (hdev == NULL)
1294                 return;
1295
1296         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1297                 tx = &tpo->tpo_tx_descs[i];
1298                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1299                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1300                                                           tx->tx_msgaddr),
1301                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1302         }
1303
1304         kiblnd_hdev_decref(hdev);
1305         tpo->tpo_hdev = NULL;
1306 }
1307
1308 static kib_hca_dev_t *
1309 kiblnd_current_hdev(kib_dev_t *dev)
1310 {
1311         kib_hca_dev_t *hdev;
1312         unsigned long  flags;
1313         int            i = 0;
1314
1315         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1316         while (dev->ibd_failover) {
1317                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1318                 if (i++ % 50 == 0)
1319                         CDEBUG(D_NET, "%s: Wait for failover\n",
1320                                dev->ibd_ifname);
1321                 schedule_timeout(cfs_time_seconds(1) / 100);
1322
1323                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1324         }
1325
1326         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1327         hdev = dev->ibd_hdev;
1328
1329         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1330
1331         return hdev;
1332 }
1333
1334 static void
1335 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1336 {
1337         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1338         kib_pool_t     *pool  = &tpo->tpo_pool;
1339         kib_net_t      *net   = pool->po_owner->ps_net;
1340         kib_dev_t      *dev;
1341         struct page    *page;
1342         kib_tx_t       *tx;
1343         int             page_offset;
1344         int             ipage;
1345         int             i;
1346
1347         LASSERT (net != NULL);
1348
1349         dev = net->ibn_dev;
1350
1351         /* pre-mapped messages are not bigger than 1 page */
1352         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1353
1354         /* No fancy arithmetic when we do the buffer calculations */
1355         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1356
1357         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1358
1359         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1360                 page = txpgs->ibp_pages[ipage];
1361                 tx = &tpo->tpo_tx_descs[i];
1362
1363                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1364                                            page_offset);
1365
1366                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1367                                                        tx->tx_msg,
1368                                                        IBLND_MSG_SIZE,
1369                                                        DMA_TO_DEVICE);
1370                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1371                                                   tx->tx_msgaddr));
1372                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1373
1374                 list_add(&tx->tx_list, &pool->po_free_list);
1375
1376                 page_offset += IBLND_MSG_SIZE;
1377                 LASSERT(page_offset <= PAGE_SIZE);
1378
1379                 if (page_offset == PAGE_SIZE) {
1380                         page_offset = 0;
1381                         ipage++;
1382                         LASSERT(ipage <= txpgs->ibp_npages);
1383                 }
1384         }
1385 }
1386
1387 struct ib_mr *
1388 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1389 {
1390         __u64   index;
1391
1392         LASSERT (hdev->ibh_mrs[0] != NULL);
1393
1394         if (hdev->ibh_nmrs == 1)
1395                 return hdev->ibh_mrs[0];
1396
1397         index = addr >> hdev->ibh_mr_shift;
1398
1399         if (index <  hdev->ibh_nmrs &&
1400             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1401                 return hdev->ibh_mrs[index];
1402
1403         return NULL;
1404 }
1405
1406 struct ib_mr *
1407 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1408 {
1409         struct ib_mr *prev_mr;
1410         struct ib_mr *mr;
1411         int           i;
1412
1413         LASSERT (hdev->ibh_mrs[0] != NULL);
1414
1415         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1416             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1417                 return NULL;
1418
1419         if (hdev->ibh_nmrs == 1)
1420                 return hdev->ibh_mrs[0];
1421
1422         for (i = 0, mr = prev_mr = NULL;
1423              i < rd->rd_nfrags; i++) {
1424                 mr = kiblnd_find_dma_mr(hdev,
1425                                         rd->rd_frags[i].rf_addr,
1426                                         rd->rd_frags[i].rf_nob);
1427                 if (prev_mr == NULL)
1428                         prev_mr = mr;
1429
1430                 if (mr == NULL || prev_mr != mr) {
1431                         /* Can't covered by one single MR */
1432                         mr = NULL;
1433                         break;
1434                 }
1435         }
1436
1437         return mr;
1438 }
1439
1440 static void
1441 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1442 {
1443         LASSERT (pool->fpo_map_count == 0);
1444
1445         if (pool->fpo_fmr_pool != NULL)
1446                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1447
1448         if (pool->fpo_hdev != NULL)
1449                 kiblnd_hdev_decref(pool->fpo_hdev);
1450
1451         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1452 }
1453
1454 static void
1455 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1456 {
1457         kib_fmr_pool_t *pool;
1458
1459         while (!list_empty(head)) {
1460                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1461                 list_del(&pool->fpo_list);
1462                 kiblnd_destroy_fmr_pool(pool);
1463         }
1464 }
1465
1466 static int kiblnd_fmr_pool_size(int ncpts)
1467 {
1468         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1469
1470         return max(IBLND_FMR_POOL, size);
1471 }
1472
1473 static int kiblnd_fmr_flush_trigger(int ncpts)
1474 {
1475         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1476
1477         return max(IBLND_FMR_POOL_FLUSH, size);
1478 }
1479
1480 static int
1481 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1482 {
1483         /* FMR pool for RDMA */
1484         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1485         kib_fmr_pool_t          *fpo;
1486         struct ib_fmr_pool_param param = {
1487                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1488                 .page_shift        = PAGE_SHIFT,
1489                 .access            = (IB_ACCESS_LOCAL_WRITE |
1490                                       IB_ACCESS_REMOTE_WRITE),
1491                 .pool_size         = fps->fps_pool_size,
1492                 .dirty_watermark   = fps->fps_flush_trigger,
1493                 .flush_function    = NULL,
1494                 .flush_arg         = NULL,
1495                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1496         int rc;
1497
1498         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1499         if (fpo == NULL)
1500                 return -ENOMEM;
1501
1502         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1503
1504         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1505         if (IS_ERR(fpo->fpo_fmr_pool)) {
1506                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1507                 CERROR("Failed to create FMR pool: %d\n", rc);
1508
1509                 kiblnd_hdev_decref(fpo->fpo_hdev);
1510                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1511                 return rc;
1512         }
1513
1514         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1515         fpo->fpo_owner    = fps;
1516         *pp_fpo = fpo;
1517
1518         return 0;
1519 }
1520
1521 static void
1522 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1523 {
1524         if (fps->fps_net == NULL) /* intialized? */
1525                 return;
1526
1527         spin_lock(&fps->fps_lock);
1528
1529         while (!list_empty(&fps->fps_pool_list)) {
1530                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1531                                                  kib_fmr_pool_t, fpo_list);
1532                 fpo->fpo_failed = 1;
1533                 list_del(&fpo->fpo_list);
1534                 if (fpo->fpo_map_count == 0)
1535                         list_add(&fpo->fpo_list, zombies);
1536                 else
1537                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1538         }
1539
1540         spin_unlock(&fps->fps_lock);
1541 }
1542
1543 static void
1544 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1545 {
1546         if (fps->fps_net != NULL) { /* initialized? */
1547                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1548                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1549         }
1550 }
1551
1552 static int
1553 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1554                         int pool_size, int flush_trigger)
1555 {
1556         kib_fmr_pool_t *fpo;
1557         int             rc;
1558
1559         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1560
1561         fps->fps_net = net;
1562         fps->fps_cpt = cpt;
1563         fps->fps_pool_size = pool_size;
1564         fps->fps_flush_trigger = flush_trigger;
1565         spin_lock_init(&fps->fps_lock);
1566         INIT_LIST_HEAD(&fps->fps_pool_list);
1567         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1568
1569         rc = kiblnd_create_fmr_pool(fps, &fpo);
1570         if (rc == 0)
1571                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1572
1573         return rc;
1574 }
1575
1576 static int
1577 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1578 {
1579         if (fpo->fpo_map_count != 0) /* still in use */
1580                 return 0;
1581         if (fpo->fpo_failed)
1582                 return 1;
1583         return cfs_time_aftereq(now, fpo->fpo_deadline);
1584 }
1585
1586 void
1587 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1588 {
1589         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1590         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1591         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1592         cfs_time_t         now = cfs_time_current();
1593         kib_fmr_pool_t    *tmp;
1594         int                rc;
1595
1596         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1597         LASSERT(rc == 0);
1598
1599         if (status != 0) {
1600                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1601                 LASSERT(rc == 0);
1602         }
1603
1604         fmr->fmr_pool = NULL;
1605         fmr->fmr_pfmr = NULL;
1606
1607         spin_lock(&fps->fps_lock);
1608         fpo->fpo_map_count--;   /* decref the pool */
1609
1610         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1611                 /* the first pool is persistent */
1612                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1613                         continue;
1614
1615                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1616                         list_move(&fpo->fpo_list, &zombies);
1617                         fps->fps_version++;
1618                 }
1619         }
1620         spin_unlock(&fps->fps_lock);
1621
1622         if (!list_empty(&zombies))
1623                 kiblnd_destroy_fmr_pool_list(&zombies);
1624 }
1625
1626 int
1627 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1628                     __u64 iov, kib_fmr_t *fmr)
1629 {
1630         struct ib_pool_fmr *pfmr;
1631         kib_fmr_pool_t     *fpo;
1632         __u64               version;
1633         int                 rc;
1634
1635 again:
1636         spin_lock(&fps->fps_lock);
1637         version = fps->fps_version;
1638         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1639                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1640                 fpo->fpo_map_count++;
1641                 spin_unlock(&fps->fps_lock);
1642
1643                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1644                                             pages, npages, iov);
1645                 if (likely(!IS_ERR(pfmr))) {
1646                         fmr->fmr_pool = fpo;
1647                         fmr->fmr_pfmr = pfmr;
1648                         return 0;
1649                 }
1650
1651                 spin_lock(&fps->fps_lock);
1652                 fpo->fpo_map_count--;
1653                 if (PTR_ERR(pfmr) != -EAGAIN) {
1654                         spin_unlock(&fps->fps_lock);
1655                         return PTR_ERR(pfmr);
1656                 }
1657
1658                 /* EAGAIN and ... */
1659                 if (version != fps->fps_version) {
1660                         spin_unlock(&fps->fps_lock);
1661                         goto again;
1662                 }
1663         }
1664
1665         if (fps->fps_increasing) {
1666                 spin_unlock(&fps->fps_lock);
1667                 CDEBUG(D_NET, "Another thread is allocating new "
1668                        "FMR pool, waiting for her to complete\n");
1669                 schedule();
1670                 goto again;
1671
1672         }
1673
1674         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1675                 /* someone failed recently */
1676                 spin_unlock(&fps->fps_lock);
1677                 return -EAGAIN;
1678         }
1679
1680         fps->fps_increasing = 1;
1681         spin_unlock(&fps->fps_lock);
1682
1683         CDEBUG(D_NET, "Allocate new FMR pool\n");
1684         rc = kiblnd_create_fmr_pool(fps, &fpo);
1685         spin_lock(&fps->fps_lock);
1686         fps->fps_increasing = 0;
1687         if (rc == 0) {
1688                 fps->fps_version++;
1689                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1690         } else {
1691                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1692         }
1693         spin_unlock(&fps->fps_lock);
1694
1695         goto again;
1696 }
1697
1698 static void
1699 kiblnd_fini_pool(kib_pool_t *pool)
1700 {
1701         LASSERT(list_empty(&pool->po_free_list));
1702         LASSERT(pool->po_allocated == 0);
1703
1704         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1705 }
1706
1707 static void
1708 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1709 {
1710         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1711
1712         memset(pool, 0, sizeof(kib_pool_t));
1713         INIT_LIST_HEAD(&pool->po_free_list);
1714         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1715         pool->po_owner    = ps;
1716         pool->po_size     = size;
1717 }
1718
1719 static void
1720 kiblnd_destroy_pool_list(struct list_head *head)
1721 {
1722         kib_pool_t *pool;
1723
1724         while (!list_empty(head)) {
1725                 pool = list_entry(head->next, kib_pool_t, po_list);
1726                 list_del(&pool->po_list);
1727
1728                 LASSERT(pool->po_owner != NULL);
1729                 pool->po_owner->ps_pool_destroy(pool);
1730         }
1731 }
1732
1733 static void
1734 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1735 {
1736         if (ps->ps_net == NULL) /* intialized? */
1737                 return;
1738
1739         spin_lock(&ps->ps_lock);
1740         while (!list_empty(&ps->ps_pool_list)) {
1741                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1742                                             kib_pool_t, po_list);
1743                 po->po_failed = 1;
1744                 list_del(&po->po_list);
1745                 if (po->po_allocated == 0)
1746                         list_add(&po->po_list, zombies);
1747                 else
1748                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1749         }
1750         spin_unlock(&ps->ps_lock);
1751 }
1752
1753 static void
1754 kiblnd_fini_poolset(kib_poolset_t *ps)
1755 {
1756         if (ps->ps_net != NULL) { /* initialized? */
1757                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1758                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1759         }
1760 }
1761
1762 static int
1763 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1764                     kib_net_t *net, char *name, int size,
1765                     kib_ps_pool_create_t po_create,
1766                     kib_ps_pool_destroy_t po_destroy,
1767                     kib_ps_node_init_t nd_init,
1768                     kib_ps_node_fini_t nd_fini)
1769 {
1770         kib_pool_t      *pool;
1771         int             rc;
1772
1773         memset(ps, 0, sizeof(kib_poolset_t));
1774
1775         ps->ps_cpt          = cpt;
1776         ps->ps_net          = net;
1777         ps->ps_pool_create  = po_create;
1778         ps->ps_pool_destroy = po_destroy;
1779         ps->ps_node_init    = nd_init;
1780         ps->ps_node_fini    = nd_fini;
1781         ps->ps_pool_size    = size;
1782         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1783             >= sizeof(ps->ps_name))
1784                 return -E2BIG;
1785         spin_lock_init(&ps->ps_lock);
1786         INIT_LIST_HEAD(&ps->ps_pool_list);
1787         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1788
1789         rc = ps->ps_pool_create(ps, size, &pool);
1790         if (rc == 0)
1791                 list_add(&pool->po_list, &ps->ps_pool_list);
1792         else
1793                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1794
1795         return rc;
1796 }
1797
1798 static int
1799 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1800 {
1801         if (pool->po_allocated != 0) /* still in use */
1802                 return 0;
1803         if (pool->po_failed)
1804                 return 1;
1805         return cfs_time_aftereq(now, pool->po_deadline);
1806 }
1807
1808 void
1809 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1810 {
1811         struct list_head zombies = LIST_HEAD_INIT(zombies);
1812         kib_poolset_t   *ps = pool->po_owner;
1813         kib_pool_t      *tmp;
1814         cfs_time_t       now = cfs_time_current();
1815
1816         spin_lock(&ps->ps_lock);
1817
1818         if (ps->ps_node_fini != NULL)
1819                 ps->ps_node_fini(pool, node);
1820
1821         LASSERT(pool->po_allocated > 0);
1822         list_add(node, &pool->po_free_list);
1823         pool->po_allocated--;
1824
1825         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1826                 /* the first pool is persistent */
1827                 if (ps->ps_pool_list.next == &pool->po_list)
1828                         continue;
1829
1830                 if (kiblnd_pool_is_idle(pool, now))
1831                         list_move(&pool->po_list, &zombies);
1832         }
1833         spin_unlock(&ps->ps_lock);
1834
1835         if (!list_empty(&zombies))
1836                 kiblnd_destroy_pool_list(&zombies);
1837 }
1838
1839 struct list_head *
1840 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1841 {
1842         struct list_head        *node;
1843         kib_pool_t              *pool;
1844         int                     rc;
1845
1846 again:
1847         spin_lock(&ps->ps_lock);
1848         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1849                 if (list_empty(&pool->po_free_list))
1850                         continue;
1851
1852                 pool->po_allocated++;
1853                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1854                 node = pool->po_free_list.next;
1855                 list_del(node);
1856
1857                 if (ps->ps_node_init != NULL) {
1858                         /* still hold the lock */
1859                         ps->ps_node_init(pool, node);
1860                 }
1861                 spin_unlock(&ps->ps_lock);
1862                 return node;
1863         }
1864
1865         /* no available tx pool and ... */
1866         if (ps->ps_increasing) {
1867                 /* another thread is allocating a new pool */
1868                 spin_unlock(&ps->ps_lock);
1869                 CDEBUG(D_NET, "Another thread is allocating new "
1870                        "%s pool, waiting for her to complete\n",
1871                        ps->ps_name);
1872                 schedule();
1873                 goto again;
1874         }
1875
1876         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1877                 /* someone failed recently */
1878                 spin_unlock(&ps->ps_lock);
1879                 return NULL;
1880         }
1881
1882         ps->ps_increasing = 1;
1883         spin_unlock(&ps->ps_lock);
1884
1885         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1886
1887         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1888
1889         spin_lock(&ps->ps_lock);
1890         ps->ps_increasing = 0;
1891         if (rc == 0) {
1892                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1893         } else {
1894                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1895                 CERROR("Can't allocate new %s pool because out of memory\n",
1896                        ps->ps_name);
1897         }
1898         spin_unlock(&ps->ps_lock);
1899
1900         goto again;
1901 }
1902
1903 void
1904 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1905 {
1906         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1907         struct ib_mr        *mr  = pmr->pmr_mr;
1908
1909         pmr->pmr_mr = NULL;
1910         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1911         if (mr != NULL)
1912                 ib_dereg_mr(mr);
1913 }
1914
1915 int
1916 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1917                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1918 {
1919         kib_phys_mr_t    *pmr;
1920         struct list_head *node;
1921         int               rc;
1922         int               i;
1923
1924         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1925         if (node == NULL) {
1926                 CERROR("Failed to allocate PMR descriptor\n");
1927                 return -ENOMEM;
1928         }
1929
1930         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1931         if (pmr->pmr_pool->ppo_hdev != hdev) {
1932                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1933                 return -EAGAIN;
1934         }
1935
1936         for (i = 0; i < rd->rd_nfrags; i ++) {
1937                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1938                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1939         }
1940
1941         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1942                                      pmr->pmr_ipb, rd->rd_nfrags,
1943                                      IB_ACCESS_LOCAL_WRITE |
1944                                      IB_ACCESS_REMOTE_WRITE,
1945                                      iova);
1946         if (!IS_ERR(pmr->pmr_mr)) {
1947                 pmr->pmr_iova = *iova;
1948                 *pp_pmr = pmr;
1949                 return 0;
1950         }
1951
1952         rc = PTR_ERR(pmr->pmr_mr);
1953         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1954
1955         pmr->pmr_mr = NULL;
1956         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1957
1958         return rc;
1959 }
1960
1961 static void
1962 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1963 {
1964         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1965         kib_phys_mr_t  *pmr;
1966
1967         LASSERT (pool->po_allocated == 0);
1968
1969         while (!list_empty(&pool->po_free_list)) {
1970                 pmr = list_entry(pool->po_free_list.next,
1971                                      kib_phys_mr_t, pmr_list);
1972
1973                 LASSERT (pmr->pmr_mr == NULL);
1974                 list_del(&pmr->pmr_list);
1975
1976                 if (pmr->pmr_ipb != NULL) {
1977                         LIBCFS_FREE(pmr->pmr_ipb,
1978                                     IBLND_MAX_RDMA_FRAGS *
1979                                     sizeof(struct ib_phys_buf));
1980                 }
1981
1982                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1983         }
1984
1985         kiblnd_fini_pool(pool);
1986         if (ppo->ppo_hdev != NULL)
1987                 kiblnd_hdev_decref(ppo->ppo_hdev);
1988
1989         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1990 }
1991
1992 static inline int kiblnd_pmr_pool_size(int ncpts)
1993 {
1994         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1995
1996         return max(IBLND_PMR_POOL, size);
1997 }
1998
1999 static int
2000 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2001 {
2002         struct kib_pmr_pool     *ppo;
2003         struct kib_pool         *pool;
2004         kib_phys_mr_t           *pmr;
2005         int                     i;
2006
2007         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2008                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2009         if (ppo == NULL) {
2010                 CERROR("Failed to allocate PMR pool\n");
2011                 return -ENOMEM;
2012         }
2013
2014         pool = &ppo->ppo_pool;
2015         kiblnd_init_pool(ps, pool, size);
2016
2017         for (i = 0; i < size; i++) {
2018                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2019                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2020                 if (pmr == NULL)
2021                         break;
2022
2023                 pmr->pmr_pool = ppo;
2024                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2025                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2026                 if (pmr->pmr_ipb == NULL)
2027                         break;
2028
2029                 list_add(&pmr->pmr_list, &pool->po_free_list);
2030         }
2031
2032         if (i < size) {
2033                 ps->ps_pool_destroy(pool);
2034                 return -ENOMEM;
2035         }
2036
2037         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2038         *pp_po = pool;
2039         return 0;
2040 }
2041
2042 static void
2043 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2044 {
2045         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2046         int             i;
2047
2048         LASSERT (pool->po_allocated == 0);
2049
2050         if (tpo->tpo_tx_pages != NULL) {
2051                 kiblnd_unmap_tx_pool(tpo);
2052                 kiblnd_free_pages(tpo->tpo_tx_pages);
2053         }
2054
2055         if (tpo->tpo_tx_descs == NULL)
2056                 goto out;
2057
2058         for (i = 0; i < pool->po_size; i++) {
2059                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2060
2061                 list_del(&tx->tx_list);
2062                 if (tx->tx_pages != NULL)
2063                         LIBCFS_FREE(tx->tx_pages,
2064                                     LNET_MAX_IOV *
2065                                     sizeof(*tx->tx_pages));
2066                 if (tx->tx_frags != NULL)
2067                         LIBCFS_FREE(tx->tx_frags,
2068                                     IBLND_MAX_RDMA_FRAGS *
2069                                             sizeof(*tx->tx_frags));
2070                 if (tx->tx_wrq != NULL)
2071                         LIBCFS_FREE(tx->tx_wrq,
2072                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2073                                     sizeof(*tx->tx_wrq));
2074                 if (tx->tx_sge != NULL)
2075                         LIBCFS_FREE(tx->tx_sge,
2076                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2077                                     sizeof(*tx->tx_sge));
2078                 if (tx->tx_rd != NULL)
2079                         LIBCFS_FREE(tx->tx_rd,
2080                                     offsetof(kib_rdma_desc_t,
2081                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2082         }
2083
2084         LIBCFS_FREE(tpo->tpo_tx_descs,
2085                     pool->po_size * sizeof(kib_tx_t));
2086 out:
2087         kiblnd_fini_pool(pool);
2088         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2089 }
2090
2091 static int kiblnd_tx_pool_size(int ncpts)
2092 {
2093         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2094
2095         return max(IBLND_TX_POOL, ntx);
2096 }
2097
2098 static int
2099 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2100 {
2101         int            i;
2102         int            npg;
2103         kib_pool_t    *pool;
2104         kib_tx_pool_t *tpo;
2105
2106         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2107         if (tpo == NULL) {
2108                 CERROR("Failed to allocate TX pool\n");
2109                 return -ENOMEM;
2110         }
2111
2112         pool = &tpo->tpo_pool;
2113         kiblnd_init_pool(ps, pool, size);
2114         tpo->tpo_tx_descs = NULL;
2115         tpo->tpo_tx_pages = NULL;
2116
2117         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2118         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2119                 CERROR("Can't allocate tx pages: %d\n", npg);
2120                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2121                 return -ENOMEM;
2122         }
2123
2124         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2125                          size * sizeof(kib_tx_t));
2126         if (tpo->tpo_tx_descs == NULL) {
2127                 CERROR("Can't allocate %d tx descriptors\n", size);
2128                 ps->ps_pool_destroy(pool);
2129                 return -ENOMEM;
2130         }
2131
2132         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2133
2134         for (i = 0; i < size; i++) {
2135                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2136
2137                 tx->tx_pool = tpo;
2138                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2139                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2140                                          lnet_cpt_table(), ps->ps_cpt,
2141                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2142                         if (tx->tx_pages == NULL)
2143                                 break;
2144                 }
2145
2146                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2147                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2148                 if (tx->tx_frags == NULL)
2149                         break;
2150
2151                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2152
2153                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2154                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2155                                  sizeof(*tx->tx_wrq));
2156                 if (tx->tx_wrq == NULL)
2157                         break;
2158
2159                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2160                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2161                                  sizeof(*tx->tx_sge));
2162                 if (tx->tx_sge == NULL)
2163                         break;
2164
2165                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2166                                  offsetof(kib_rdma_desc_t,
2167                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2168                 if (tx->tx_rd == NULL)
2169                         break;
2170         }
2171
2172         if (i == size) {
2173                 kiblnd_map_tx_pool(tpo);
2174                 *pp_po = pool;
2175                 return 0;
2176         }
2177
2178         ps->ps_pool_destroy(pool);
2179         return -ENOMEM;
2180 }
2181
2182 static void
2183 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2184 {
2185         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2186                                              tps_poolset);
2187         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2188
2189         tx->tx_cookie = tps->tps_next_tx_cookie++;
2190 }
2191
2192 static void
2193 kiblnd_net_fini_pools(kib_net_t *net)
2194 {
2195         int     i;
2196
2197         cfs_cpt_for_each(i, lnet_cpt_table()) {
2198                 kib_tx_poolset_t        *tps;
2199                 kib_fmr_poolset_t       *fps;
2200                 kib_pmr_poolset_t       *pps;
2201
2202                 if (net->ibn_tx_ps != NULL) {
2203                         tps = net->ibn_tx_ps[i];
2204                         kiblnd_fini_poolset(&tps->tps_poolset);
2205                 }
2206
2207                 if (net->ibn_fmr_ps != NULL) {
2208                         fps = net->ibn_fmr_ps[i];
2209                         kiblnd_fini_fmr_poolset(fps);
2210                 }
2211
2212                 if (net->ibn_pmr_ps != NULL) {
2213                         pps = net->ibn_pmr_ps[i];
2214                         kiblnd_fini_poolset(&pps->pps_poolset);
2215                 }
2216         }
2217
2218         if (net->ibn_tx_ps != NULL) {
2219                 cfs_percpt_free(net->ibn_tx_ps);
2220                 net->ibn_tx_ps = NULL;
2221         }
2222
2223         if (net->ibn_fmr_ps != NULL) {
2224                 cfs_percpt_free(net->ibn_fmr_ps);
2225                 net->ibn_fmr_ps = NULL;
2226         }
2227
2228         if (net->ibn_pmr_ps != NULL) {
2229                 cfs_percpt_free(net->ibn_pmr_ps);
2230                 net->ibn_pmr_ps = NULL;
2231         }
2232 }
2233
2234 static int
2235 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2236 {
2237         unsigned long   flags;
2238         int             cpt;
2239         int             rc;
2240         int             i;
2241
2242         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2243         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2244             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2245                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2246                                            flags);
2247                 goto create_tx_pool;
2248         }
2249
2250         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2251
2252         if (*kiblnd_tunables.kib_fmr_pool_size <
2253             *kiblnd_tunables.kib_ntx / 4) {
2254                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2255                        *kiblnd_tunables.kib_fmr_pool_size,
2256                        *kiblnd_tunables.kib_ntx / 4);
2257                 rc = -EINVAL;
2258                 goto failed;
2259         }
2260
2261         /* TX pool must be created later than FMR/PMR, see LU-2268
2262          * for details */
2263         LASSERT(net->ibn_tx_ps == NULL);
2264
2265         /* premapping can fail if ibd_nmr > 1, so we always create
2266          * FMR/PMR pool and map-on-demand if premapping failed */
2267
2268         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2269                                            sizeof(kib_fmr_poolset_t));
2270         if (net->ibn_fmr_ps == NULL) {
2271                 CERROR("Failed to allocate FMR pool array\n");
2272                 rc = -ENOMEM;
2273                 goto failed;
2274         }
2275
2276         for (i = 0; i < ncpts; i++) {
2277                 cpt = (cpts == NULL) ? i : cpts[i];
2278                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2279                                              kiblnd_fmr_pool_size(ncpts),
2280                                              kiblnd_fmr_flush_trigger(ncpts));
2281                 if (rc == -ENOSYS && i == 0) /* no FMR */
2282                         break; /* create PMR pool */
2283
2284                 if (rc != 0) { /* a real error */
2285                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2286                                cpt, rc);
2287                         goto failed;
2288                 }
2289         }
2290
2291         if (i > 0) {
2292                 LASSERT(i == ncpts);
2293                 goto create_tx_pool;
2294         }
2295
2296         cfs_percpt_free(net->ibn_fmr_ps);
2297         net->ibn_fmr_ps = NULL;
2298
2299         CWARN("Device does not support FMR, failing back to PMR\n");
2300
2301         if (*kiblnd_tunables.kib_pmr_pool_size <
2302             *kiblnd_tunables.kib_ntx / 4) {
2303                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2304                        *kiblnd_tunables.kib_pmr_pool_size,
2305                        *kiblnd_tunables.kib_ntx / 4);
2306                 rc = -EINVAL;
2307                 goto failed;
2308         }
2309
2310         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2311                                            sizeof(kib_pmr_poolset_t));
2312         if (net->ibn_pmr_ps == NULL) {
2313                 CERROR("Failed to allocate PMR pool array\n");
2314                 rc = -ENOMEM;
2315                 goto failed;
2316         }
2317
2318         for (i = 0; i < ncpts; i++) {
2319                 cpt = (cpts == NULL) ? i : cpts[i];
2320                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2321                                          cpt, net, "PMR",
2322                                          kiblnd_pmr_pool_size(ncpts),
2323                                          kiblnd_create_pmr_pool,
2324                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2325                 if (rc != 0) {
2326                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2327                                cpt, rc);
2328                         goto failed;
2329                 }
2330         }
2331
2332  create_tx_pool:
2333         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2334                                           sizeof(kib_tx_poolset_t));
2335         if (net->ibn_tx_ps == NULL) {
2336                 CERROR("Failed to allocate tx pool array\n");
2337                 rc = -ENOMEM;
2338                 goto failed;
2339         }
2340
2341         for (i = 0; i < ncpts; i++) {
2342                 cpt = (cpts == NULL) ? i : cpts[i];
2343                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2344                                          cpt, net, "TX",
2345                                          kiblnd_tx_pool_size(ncpts),
2346                                          kiblnd_create_tx_pool,
2347                                          kiblnd_destroy_tx_pool,
2348                                          kiblnd_tx_init, NULL);
2349                 if (rc != 0) {
2350                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2351                                cpt, rc);
2352                         goto failed;
2353                 }
2354         }
2355
2356         return 0;
2357  failed:
2358         kiblnd_net_fini_pools(net);
2359         LASSERT(rc != 0);
2360         return rc;
2361 }
2362
2363 static int
2364 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2365 {
2366         struct ib_device_attr *attr;
2367         int                    rc;
2368
2369         /* It's safe to assume a HCA can handle a page size
2370          * matching that of the native system */
2371         hdev->ibh_page_shift = PAGE_SHIFT;
2372         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2373         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2374
2375         LIBCFS_ALLOC(attr, sizeof(*attr));
2376         if (attr == NULL) {
2377                 CERROR("Out of memory\n");
2378                 return -ENOMEM;
2379         }
2380
2381         rc = ib_query_device(hdev->ibh_ibdev, attr);
2382         if (rc == 0)
2383                 hdev->ibh_mr_size = attr->max_mr_size;
2384
2385         LIBCFS_FREE(attr, sizeof(*attr));
2386
2387         if (rc != 0) {
2388                 CERROR("Failed to query IB device: %d\n", rc);
2389                 return rc;
2390         }
2391
2392         if (hdev->ibh_mr_size == ~0ULL) {
2393                 hdev->ibh_mr_shift = 64;
2394                 return 0;
2395         }
2396
2397         for (hdev->ibh_mr_shift = 0;
2398              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2399                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2400                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2401                         return 0;
2402         }
2403
2404         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2405         return -EINVAL;
2406 }
2407
2408 static void
2409 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2410 {
2411         int     i;
2412
2413         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2414                 return;
2415
2416         for (i = 0; i < hdev->ibh_nmrs; i++) {
2417                 if (hdev->ibh_mrs[i] == NULL)
2418                         break;
2419
2420                 ib_dereg_mr(hdev->ibh_mrs[i]);
2421         }
2422
2423         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2424         hdev->ibh_mrs  = NULL;
2425         hdev->ibh_nmrs = 0;
2426 }
2427
2428 void
2429 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2430 {
2431         kiblnd_hdev_cleanup_mrs(hdev);
2432
2433         if (hdev->ibh_pd != NULL)
2434                 ib_dealloc_pd(hdev->ibh_pd);
2435
2436         if (hdev->ibh_cmid != NULL)
2437                 rdma_destroy_id(hdev->ibh_cmid);
2438
2439         LIBCFS_FREE(hdev, sizeof(*hdev));
2440 }
2441
2442 static int
2443 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2444 {
2445         struct ib_mr *mr;
2446         int           i;
2447         int           rc;
2448         __u64         mm_size;
2449         __u64         mr_size;
2450         int           acflags = IB_ACCESS_LOCAL_WRITE |
2451                                 IB_ACCESS_REMOTE_WRITE;
2452
2453         rc = kiblnd_hdev_get_attr(hdev);
2454         if (rc != 0)
2455                 return rc;
2456
2457         if (hdev->ibh_mr_shift == 64) {
2458                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2459                 if (hdev->ibh_mrs == NULL) {
2460                         CERROR("Failed to allocate MRs table\n");
2461                         return -ENOMEM;
2462                 }
2463
2464                 hdev->ibh_mrs[0] = NULL;
2465                 hdev->ibh_nmrs   = 1;
2466
2467                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2468                 if (IS_ERR(mr)) {
2469                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2470                         kiblnd_hdev_cleanup_mrs(hdev);
2471                         return PTR_ERR(mr);
2472                 }
2473
2474                 hdev->ibh_mrs[0] = mr;
2475
2476                 goto out;
2477         }
2478
2479         mr_size = (1ULL << hdev->ibh_mr_shift);
2480         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2481
2482         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2483
2484         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2485                 /* it's 4T..., assume we will re-code at that time */
2486                 CERROR("Can't support memory size: x"LPX64
2487                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2488                 return -EINVAL;
2489         }
2490
2491         /* create an array of MRs to cover all memory */
2492         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2493         if (hdev->ibh_mrs == NULL) {
2494                 CERROR("Failed to allocate MRs' table\n");
2495                 return -ENOMEM;
2496         }
2497
2498         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2499
2500         for (i = 0; i < hdev->ibh_nmrs; i++) {
2501                 struct ib_phys_buf ipb;
2502                 __u64              iova;
2503
2504                 ipb.size = hdev->ibh_mr_size;
2505                 ipb.addr = i * mr_size;
2506                 iova     = ipb.addr;
2507
2508                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2509                 if (IS_ERR(mr)) {
2510                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2511                                " size "LPX64" : %ld\n",
2512                                ipb.addr, ipb.size, PTR_ERR(mr));
2513                         kiblnd_hdev_cleanup_mrs(hdev);
2514                         return PTR_ERR(mr);
2515                 }
2516
2517                 LASSERT (iova == ipb.addr);
2518
2519                 hdev->ibh_mrs[i] = mr;
2520         }
2521
2522 out:
2523         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2524                 LCONSOLE_INFO("Register global MR array, MR size: "
2525                               LPX64", array size: %d\n",
2526                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2527         return 0;
2528 }
2529
2530 static int
2531 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2532 {       /* DUMMY */
2533         return 0;
2534 }
2535
2536 static int
2537 kiblnd_dev_need_failover(kib_dev_t *dev)
2538 {
2539         struct rdma_cm_id  *cmid;
2540         struct sockaddr_in  srcaddr;
2541         struct sockaddr_in  dstaddr;
2542         int                 rc;
2543
2544         if (dev->ibd_hdev == NULL || /* initializing */
2545             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2546             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2547                 return 1;
2548
2549         /* XXX: it's UGLY, but I don't have better way to find
2550          * ib-bonding HCA failover because:
2551          *
2552          * a. no reliable CM event for HCA failover...
2553          * b. no OFED API to get ib_device for current net_device...
2554          *
2555          * We have only two choices at this point:
2556          *
2557          * a. rdma_bind_addr(), it will conflict with listener cmid
2558          * b. rdma_resolve_addr() to zero addr */
2559         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2560                                      IB_QPT_RC);
2561         if (IS_ERR(cmid)) {
2562                 rc = PTR_ERR(cmid);
2563                 CERROR("Failed to create cmid for failover: %d\n", rc);
2564                 return rc;
2565         }
2566
2567         memset(&srcaddr, 0, sizeof(srcaddr));
2568         srcaddr.sin_family      = AF_INET;
2569         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2570
2571         memset(&dstaddr, 0, sizeof(dstaddr));
2572         dstaddr.sin_family = AF_INET;
2573         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2574                                (struct sockaddr *)&dstaddr, 1);
2575         if (rc != 0 || cmid->device == NULL) {
2576                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2577                        dev->ibd_ifname, &dev->ibd_ifip,
2578                        cmid->device, rc);
2579                 rdma_destroy_id(cmid);
2580                 return rc;
2581         }
2582
2583         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2584         rdma_destroy_id(cmid);
2585         return rc;
2586 }
2587
2588 int
2589 kiblnd_dev_failover(kib_dev_t *dev)
2590 {
2591         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2592         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2593         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2594         struct rdma_cm_id  *cmid  = NULL;
2595         kib_hca_dev_t      *hdev  = NULL;
2596         kib_hca_dev_t      *old;
2597         struct ib_pd       *pd;
2598         kib_net_t          *net;
2599         struct sockaddr_in  addr;
2600         unsigned long       flags;
2601         int                 rc = 0;
2602         int                 i;
2603
2604         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2605                  dev->ibd_can_failover ||
2606                  dev->ibd_hdev == NULL);
2607
2608         rc = kiblnd_dev_need_failover(dev);
2609         if (rc <= 0)
2610                 goto out;
2611
2612         if (dev->ibd_hdev != NULL &&
2613             dev->ibd_hdev->ibh_cmid != NULL) {
2614                 /* XXX it's not good to close old listener at here,
2615                  * because we can fail to create new listener.
2616                  * But we have to close it now, otherwise rdma_bind_addr
2617                  * will return EADDRINUSE... How crap! */
2618                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2619
2620                 cmid = dev->ibd_hdev->ibh_cmid;
2621                 /* make next schedule of kiblnd_dev_need_failover()
2622                  * return 1 for me */
2623                 dev->ibd_hdev->ibh_cmid  = NULL;
2624                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2625
2626                 rdma_destroy_id(cmid);
2627         }
2628
2629         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2630                                      IB_QPT_RC);
2631         if (IS_ERR(cmid)) {
2632                 rc = PTR_ERR(cmid);
2633                 CERROR("Failed to create cmid for failover: %d\n", rc);
2634                 goto out;
2635         }
2636
2637         memset(&addr, 0, sizeof(addr));
2638         addr.sin_family      = AF_INET;
2639         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2640         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2641
2642         /* Bind to failover device or port */
2643         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2644         if (rc != 0 || cmid->device == NULL) {
2645                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2646                        dev->ibd_ifname, &dev->ibd_ifip,
2647                        cmid->device, rc);
2648                 rdma_destroy_id(cmid);
2649                 goto out;
2650         }
2651
2652         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2653         if (hdev == NULL) {
2654                 CERROR("Failed to allocate kib_hca_dev\n");
2655                 rdma_destroy_id(cmid);
2656                 rc = -ENOMEM;
2657                 goto out;
2658         }
2659
2660         atomic_set(&hdev->ibh_ref, 1);
2661         hdev->ibh_dev   = dev;
2662         hdev->ibh_cmid  = cmid;
2663         hdev->ibh_ibdev = cmid->device;
2664
2665         pd = ib_alloc_pd(cmid->device);
2666         if (IS_ERR(pd)) {
2667                 rc = PTR_ERR(pd);
2668                 CERROR("Can't allocate PD: %d\n", rc);
2669                 goto out;
2670         }
2671
2672         hdev->ibh_pd = pd;
2673
2674         rc = rdma_listen(cmid, 0);
2675         if (rc != 0) {
2676                 CERROR("Can't start new listener: %d\n", rc);
2677                 goto out;
2678         }
2679
2680         rc = kiblnd_hdev_setup_mrs(hdev);
2681         if (rc != 0) {
2682                 CERROR("Can't setup device: %d\n", rc);
2683                 goto out;
2684         }
2685
2686         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2687
2688         old = dev->ibd_hdev;
2689         dev->ibd_hdev = hdev;   /* take over the refcount */
2690         hdev = old;
2691
2692         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2693                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2694                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2695                                             &zombie_tpo);
2696
2697                         if (net->ibn_fmr_ps != NULL) {
2698                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2699                                                         &zombie_fpo);
2700
2701                         } else if (net->ibn_pmr_ps != NULL) {
2702                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2703                                                     pps_poolset, &zombie_ppo);
2704                         }
2705                 }
2706         }
2707
2708         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2709  out:
2710         if (!list_empty(&zombie_tpo))
2711                 kiblnd_destroy_pool_list(&zombie_tpo);
2712         if (!list_empty(&zombie_ppo))
2713                 kiblnd_destroy_pool_list(&zombie_ppo);
2714         if (!list_empty(&zombie_fpo))
2715                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2716         if (hdev != NULL)
2717                 kiblnd_hdev_decref(hdev);
2718
2719         if (rc != 0)
2720                 dev->ibd_failed_failover++;
2721         else
2722                 dev->ibd_failed_failover = 0;
2723
2724         return rc;
2725 }
2726
2727 void
2728 kiblnd_destroy_dev (kib_dev_t *dev)
2729 {
2730         LASSERT (dev->ibd_nnets == 0);
2731         LASSERT(list_empty(&dev->ibd_nets));
2732
2733         list_del(&dev->ibd_fail_list);
2734         list_del(&dev->ibd_list);
2735
2736         if (dev->ibd_hdev != NULL)
2737                 kiblnd_hdev_decref(dev->ibd_hdev);
2738
2739         LIBCFS_FREE(dev, sizeof(*dev));
2740 }
2741
2742 static kib_dev_t *
2743 kiblnd_create_dev(char *ifname)
2744 {
2745         struct net_device *netdev;
2746         kib_dev_t         *dev;
2747         __u32              netmask;
2748         __u32              ip;
2749         int                up;
2750         int                rc;
2751
2752         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2753         if (rc != 0) {
2754                 CERROR("Can't query IPoIB interface %s: %d\n",
2755                        ifname, rc);
2756                 return NULL;
2757         }
2758
2759         if (!up) {
2760                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2761                 return NULL;
2762         }
2763
2764         LIBCFS_ALLOC(dev, sizeof(*dev));
2765         if (dev == NULL)
2766                 return NULL;
2767
2768         memset(dev, 0, sizeof(*dev));
2769         netdev = dev_get_by_name(&init_net, ifname);
2770         if (netdev == NULL) {
2771                 dev->ibd_can_failover = 0;
2772         } else {
2773                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2774                 dev_put(netdev);
2775         }
2776
2777         INIT_LIST_HEAD(&dev->ibd_nets);
2778         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2779         INIT_LIST_HEAD(&dev->ibd_fail_list);
2780         dev->ibd_ifip = ip;
2781         strcpy(&dev->ibd_ifname[0], ifname);
2782
2783         /* initialize the device */
2784         rc = kiblnd_dev_failover(dev);
2785         if (rc != 0) {
2786                 CERROR("Can't initialize device: %d\n", rc);
2787                 LIBCFS_FREE(dev, sizeof(*dev));
2788                 return NULL;
2789         }
2790
2791         list_add_tail(&dev->ibd_list,
2792                           &kiblnd_data.kib_devs);
2793         return dev;
2794 }
2795
2796 static void
2797 kiblnd_base_shutdown(void)
2798 {
2799         struct kib_sched_info   *sched;
2800         int                     i;
2801
2802         LASSERT(list_empty(&kiblnd_data.kib_devs));
2803
2804         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2805                atomic_read(&libcfs_kmemory));
2806
2807         switch (kiblnd_data.kib_init) {
2808         default:
2809                 LBUG();
2810
2811         case IBLND_INIT_ALL:
2812         case IBLND_INIT_DATA:
2813                 LASSERT (kiblnd_data.kib_peers != NULL);
2814                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2815                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2816                 }
2817                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2818                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2819
2820                 /* flag threads to terminate; wake and wait for them to die */
2821                 kiblnd_data.kib_shutdown = 1;
2822
2823                 /* NB: we really want to stop scheduler threads net by net
2824                  * instead of the whole module, this should be improved
2825                  * with dynamic configuration LNet */
2826                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2827                         wake_up_all(&sched->ibs_waitq);
2828
2829                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2830                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2831
2832                 i = 2;
2833                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2834                         i++;
2835                         /* power of 2? */
2836                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2837                                "Waiting for %d threads to terminate\n",
2838                                atomic_read(&kiblnd_data.kib_nthreads));
2839                         set_current_state(TASK_UNINTERRUPTIBLE);
2840                         schedule_timeout(cfs_time_seconds(1));
2841                 }
2842
2843                 /* fall through */
2844
2845         case IBLND_INIT_NOTHING:
2846                 break;
2847         }
2848
2849         if (kiblnd_data.kib_peers != NULL) {
2850                 LIBCFS_FREE(kiblnd_data.kib_peers,
2851                             sizeof(struct list_head) *
2852                             kiblnd_data.kib_peer_hash_size);
2853         }
2854
2855         if (kiblnd_data.kib_scheds != NULL)
2856                 cfs_percpt_free(kiblnd_data.kib_scheds);
2857
2858         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2859                atomic_read(&libcfs_kmemory));
2860
2861         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2862         module_put(THIS_MODULE);
2863 }
2864
2865 static void
2866 kiblnd_shutdown (lnet_ni_t *ni)
2867 {
2868         kib_net_t        *net = ni->ni_data;
2869         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2870         int               i;
2871         unsigned long     flags;
2872
2873         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2874
2875         if (net == NULL)
2876                 goto out;
2877
2878         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2879                atomic_read(&libcfs_kmemory));
2880
2881         write_lock_irqsave(g_lock, flags);
2882         net->ibn_shutdown = 1;
2883         write_unlock_irqrestore(g_lock, flags);
2884
2885         switch (net->ibn_init) {
2886         default:
2887                 LBUG();
2888
2889         case IBLND_INIT_ALL:
2890                 /* nuke all existing peers within this net */
2891                 kiblnd_del_peer(ni, LNET_NID_ANY);
2892
2893                 /* Wait for all peer state to clean up */
2894                 i = 2;
2895                 while (atomic_read(&net->ibn_npeers) != 0) {
2896                         i++;
2897                         /* power of 2? */
2898                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2899                                "%s: waiting for %d peers to disconnect\n",
2900                                libcfs_nid2str(ni->ni_nid),
2901                                atomic_read(&net->ibn_npeers));
2902                         set_current_state(TASK_UNINTERRUPTIBLE);
2903                         schedule_timeout(cfs_time_seconds(1));
2904                 }
2905
2906                 kiblnd_net_fini_pools(net);
2907
2908                 write_lock_irqsave(g_lock, flags);
2909                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2910                 net->ibn_dev->ibd_nnets--;
2911                 list_del(&net->ibn_list);
2912                 write_unlock_irqrestore(g_lock, flags);
2913
2914                 /* fall through */
2915
2916         case IBLND_INIT_NOTHING:
2917                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2918
2919                 if (net->ibn_dev != NULL &&
2920                     net->ibn_dev->ibd_nnets == 0)
2921                         kiblnd_destroy_dev(net->ibn_dev);
2922
2923                 break;
2924         }
2925
2926         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2927                atomic_read(&libcfs_kmemory));
2928
2929         net->ibn_init = IBLND_INIT_NOTHING;
2930         ni->ni_data = NULL;
2931
2932         LIBCFS_FREE(net, sizeof(*net));
2933
2934 out:
2935         if (list_empty(&kiblnd_data.kib_devs))
2936                 kiblnd_base_shutdown();
2937         return;
2938 }
2939
2940 static int
2941 kiblnd_base_startup(void)
2942 {
2943         struct kib_sched_info   *sched;
2944         int                     rc;
2945         int                     i;
2946
2947         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2948
2949         try_module_get(THIS_MODULE);
2950         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2951
2952         rwlock_init(&kiblnd_data.kib_global_lock);
2953
2954         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2955         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2956
2957         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2958         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2959                      sizeof(struct list_head) *
2960                      kiblnd_data.kib_peer_hash_size);
2961         if (kiblnd_data.kib_peers == NULL)
2962                 goto failed;
2963
2964         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2965                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2966
2967         spin_lock_init(&kiblnd_data.kib_connd_lock);
2968         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2969         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2970         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2971         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2972
2973         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2974                                                   sizeof(*sched));
2975         if (kiblnd_data.kib_scheds == NULL)
2976                 goto failed;
2977
2978         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2979                 int     nthrs;
2980
2981                 spin_lock_init(&sched->ibs_lock);
2982                 INIT_LIST_HEAD(&sched->ibs_conns);
2983                 init_waitqueue_head(&sched->ibs_waitq);
2984
2985                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2986                 if (*kiblnd_tunables.kib_nscheds > 0) {
2987                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2988                 } else {
2989                         /* max to half of CPUs, another half is reserved for
2990                          * upper layer modules */
2991                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2992                 }
2993
2994                 sched->ibs_nthreads_max = nthrs;
2995                 sched->ibs_cpt = i;
2996         }
2997
2998         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2999
3000         /* lists/ptrs/locks initialised */
3001         kiblnd_data.kib_init = IBLND_INIT_DATA;
3002         /*****************************************************/
3003
3004         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3005         if (rc != 0) {
3006                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3007                 goto failed;
3008         }
3009
3010         if (*kiblnd_tunables.kib_dev_failover != 0)
3011                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3012                                          "kiblnd_failover");
3013
3014         if (rc != 0) {
3015                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3016                 goto failed;
3017         }
3018
3019         /* flag everything initialised */
3020         kiblnd_data.kib_init = IBLND_INIT_ALL;
3021         /*****************************************************/
3022
3023         return 0;
3024
3025  failed:
3026         kiblnd_base_shutdown();
3027         return -ENETDOWN;
3028 }
3029
3030 static int
3031 kiblnd_start_schedulers(struct kib_sched_info *sched)
3032 {
3033         int     rc = 0;
3034         int     nthrs;
3035         int     i;
3036
3037         if (sched->ibs_nthreads == 0) {
3038                 if (*kiblnd_tunables.kib_nscheds > 0) {
3039                         nthrs = sched->ibs_nthreads_max;
3040                 } else {
3041                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3042                                                sched->ibs_cpt);
3043                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3044                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3045                 }
3046         } else {
3047                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3048                 /* increase one thread if there is new interface */
3049                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3050         }
3051
3052         for (i = 0; i < nthrs; i++) {
3053                 long    id;
3054                 char    name[20];
3055                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3056                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3057                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3058                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3059                 if (rc == 0)
3060                         continue;
3061
3062                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3063                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3064                 break;
3065         }
3066
3067         sched->ibs_nthreads += i;
3068         return rc;
3069 }
3070
3071 static int
3072 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3073 {
3074         int     cpt;
3075         int     rc;
3076         int     i;
3077
3078         for (i = 0; i < ncpts; i++) {
3079                 struct kib_sched_info *sched;
3080
3081                 cpt = (cpts == NULL) ? i : cpts[i];
3082                 sched = kiblnd_data.kib_scheds[cpt];
3083
3084                 if (!newdev && sched->ibs_nthreads > 0)
3085                         continue;
3086
3087                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3088                 if (rc != 0) {
3089                         CERROR("Failed to start scheduler threads for %s\n",
3090                                dev->ibd_ifname);
3091                         return rc;
3092                 }
3093         }
3094         return 0;
3095 }
3096
3097 static kib_dev_t *
3098 kiblnd_dev_search(char *ifname)
3099 {
3100         kib_dev_t       *alias = NULL;
3101         kib_dev_t       *dev;
3102         char            *colon;
3103         char            *colon2;
3104
3105         colon = strchr(ifname, ':');
3106         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3107                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3108                         return dev;
3109
3110                 if (alias != NULL)
3111                         continue;
3112
3113                 colon2 = strchr(dev->ibd_ifname, ':');
3114                 if (colon != NULL)
3115                         *colon = 0;
3116                 if (colon2 != NULL)
3117                         *colon2 = 0;
3118
3119                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3120                         alias = dev;
3121
3122                 if (colon != NULL)
3123                         *colon = ':';
3124                 if (colon2 != NULL)
3125                         *colon2 = ':';
3126         }
3127         return alias;
3128 }
3129
3130 static int
3131 kiblnd_startup (lnet_ni_t *ni)
3132 {
3133         char                     *ifname;
3134         kib_dev_t                *ibdev = NULL;
3135         kib_net_t                *net;
3136         struct timeval            tv;
3137         unsigned long             flags;
3138         int                       rc;
3139         int                       newdev;
3140
3141         LASSERT (ni->ni_lnd == &the_o2iblnd);
3142
3143         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3144                 rc = kiblnd_base_startup();
3145                 if (rc != 0)
3146                         return rc;
3147         }
3148
3149         LIBCFS_ALLOC(net, sizeof(*net));
3150         ni->ni_data = net;
3151         if (net == NULL)
3152                 goto failed;
3153
3154         memset(net, 0, sizeof(*net));
3155
3156         do_gettimeofday(&tv);
3157         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3158
3159         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3160         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3161         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3162         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3163
3164         if (ni->ni_interfaces[0] != NULL) {
3165                 /* Use the IPoIB interface specified in 'networks=' */
3166
3167                 CLASSERT (LNET_MAX_INTERFACES > 1);
3168                 if (ni->ni_interfaces[1] != NULL) {
3169                         CERROR("Multiple interfaces not supported\n");
3170                         goto failed;
3171                 }
3172
3173                 ifname = ni->ni_interfaces[0];
3174         } else {
3175                 ifname = *kiblnd_tunables.kib_default_ipif;
3176         }
3177
3178         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3179                 CERROR("IPoIB interface name too long: %s\n", ifname);
3180                 goto failed;
3181         }
3182
3183         ibdev = kiblnd_dev_search(ifname);
3184
3185         newdev = ibdev == NULL;
3186         /* hmm...create kib_dev even for alias */
3187         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3188                 ibdev = kiblnd_create_dev(ifname);
3189
3190         if (ibdev == NULL)
3191                 goto failed;
3192
3193         net->ibn_dev = ibdev;
3194         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3195
3196         rc = kiblnd_dev_start_threads(ibdev, newdev,
3197                                       ni->ni_cpts, ni->ni_ncpts);
3198         if (rc != 0)
3199                 goto failed;
3200
3201         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3202         if (rc != 0) {
3203                 CERROR("Failed to initialize NI pools: %d\n", rc);
3204                 goto failed;
3205         }
3206
3207         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3208         ibdev->ibd_nnets++;
3209         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3210         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3211
3212         net->ibn_init = IBLND_INIT_ALL;
3213
3214         return 0;
3215
3216 failed:
3217         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3218                 kiblnd_destroy_dev(ibdev);
3219
3220         kiblnd_shutdown(ni);
3221
3222         CDEBUG(D_NET, "kiblnd_startup failed\n");
3223         return -ENETDOWN;
3224 }
3225
3226 static lnd_t the_o2iblnd = {
3227         .lnd_type       = O2IBLND,
3228         .lnd_startup    = kiblnd_startup,
3229         .lnd_shutdown   = kiblnd_shutdown,
3230         .lnd_ctl        = kiblnd_ctl,
3231         .lnd_query      = kiblnd_query,
3232         .lnd_send       = kiblnd_send,
3233         .lnd_recv       = kiblnd_recv,
3234 };
3235
3236 static void __exit
3237 kiblnd_module_fini (void)
3238 {
3239         lnet_unregister_lnd(&the_o2iblnd);
3240         kiblnd_tunables_fini();
3241 }
3242
3243 static int __init
3244 kiblnd_module_init (void)
3245 {
3246         int    rc;
3247
3248         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3249         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3250                   <= IBLND_MSG_SIZE);
3251         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3252                   <= IBLND_MSG_SIZE);
3253
3254         rc = kiblnd_tunables_init();
3255         if (rc != 0)
3256                 return rc;
3257
3258         lnet_register_lnd(&the_o2iblnd);
3259
3260         return 0;
3261 }
3262
3263 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3264 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3265 MODULE_LICENSE("GPL");
3266
3267 module_init(kiblnd_module_init);
3268 module_exit(kiblnd_module_fini);