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         kib_conn_t              *conn;
 722         struct ib_cq            *cq;
 723         unsigned long           flags;
 724         int                     cpt;
 725         int                     rc;
 726         int                     i;
 727
 728         LASSERT(net != NULL);
 729         LASSERT(!in_interrupt());
 730
 731         dev = net->ibn_dev;
 732
 733         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 734         sched = kiblnd_data.kib_scheds[cpt];
 735
 736         LASSERT(sched->ibs_nthreads > 0);
 737
 738         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 739                          sizeof(*init_qp_attr));
 740         if (init_qp_attr == NULL) {
 741                 CERROR("Can't allocate qp_attr for %s\n",
 742                        libcfs_nid2str(peer->ibp_nid));
 743                 goto failed_0;
 744         }
 745
 746         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 747         if (conn == NULL) {
 748                 CERROR("Can't allocate connection for %s\n",
 749                        libcfs_nid2str(peer->ibp_nid));
 750                 goto failed_1;
 751         }
 752
 753         conn->ibc_state = IBLND_CONN_INIT;
 754         conn->ibc_version = version;
 755         conn->ibc_peer = peer;                  /* I take the caller's ref */
 756         cmid->context = conn;                   /* for future CM callbacks */
 757         conn->ibc_cmid = cmid;
 758
 759         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 760         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 761         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 762         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 763         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 764         INIT_LIST_HEAD(&conn->ibc_active_txs);
 765         spin_lock_init(&conn->ibc_lock);
 766
 767         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 768                          sizeof(*conn->ibc_connvars));
 769         if (conn->ibc_connvars == NULL) {
 770                 CERROR("Can't allocate in-progress connection state\n");
 771                 goto failed_2;
 772         }
 773
 774         write_lock_irqsave(glock, flags);
 775         if (dev->ibd_failover) {
 776                 write_unlock_irqrestore(glock, flags);
 777                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 778                 goto failed_2;
 779         }
 780
 781         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 782                 /* wakeup failover thread and teardown connection */
 783                 if (kiblnd_dev_can_failover(dev)) {
 784                         list_add_tail(&dev->ibd_fail_list,
 785                                       &kiblnd_data.kib_failed_devs);
 786                         wake_up(&kiblnd_data.kib_failover_waitq);
 787                 }
 788
 789                 write_unlock_irqrestore(glock, flags);
 790                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 791                        cmid->device->name, dev->ibd_ifname);
 792                 goto failed_2;
 793         }
 794
 795         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 796         conn->ibc_hdev = dev->ibd_hdev;
 797
 798         kiblnd_setup_mtu_locked(cmid);
 799
 800         write_unlock_irqrestore(glock, flags);
 801
 802         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 803                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 804         if (conn->ibc_rxs == NULL) {
 805                 CERROR("Cannot allocate RX buffers\n");
 806                 goto failed_2;
 807         }
 808
 809         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 810                                 IBLND_RX_MSG_PAGES(version));
 811         if (rc != 0)
 812                 goto failed_2;
 813
 814         kiblnd_map_rx_descs(conn);
 815
 816         cq = ib_create_cq(cmid->device,
 817                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 818                           IBLND_CQ_ENTRIES(version),
 819                           kiblnd_get_completion_vector(conn, cpt));
 820         if (IS_ERR(cq)) {
 821                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 822                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 823                 goto failed_2;
 824         }
 825
 826         conn->ibc_cq = cq;
 827
 828         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 829         if (rc != 0) {
 830                 CERROR("Can't request completion notification: %d\n", rc);
 831                 goto failed_2;
 832         }
 833
 834         init_qp_attr->event_handler = kiblnd_qp_event;
 835         init_qp_attr->qp_context = conn;
 836         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 837         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 838         init_qp_attr->cap.max_send_sge = 1;
 839         init_qp_attr->cap.max_recv_sge = 1;
 840         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 841         init_qp_attr->qp_type = IB_QPT_RC;
 842         init_qp_attr->send_cq = cq;
 843         init_qp_attr->recv_cq = cq;
 844
 845         conn->ibc_sched = sched;
 846
 847         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 848         if (rc != 0) {
 849                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 850                        rc, init_qp_attr->cap.max_send_wr,
 851                        init_qp_attr->cap.max_recv_wr);
 852                 goto failed_2;
 853         }
 854
 855         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 856
 857         /* 1 ref for caller and each rxmsg */
 858         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 859         conn->ibc_nrx = IBLND_RX_MSGS(version);
 860
 861         /* post receives */
 862         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 863                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 864                                     IBLND_POSTRX_NO_CREDIT);
 865                 if (rc != 0) {
 866                         CERROR("Can't post rxmsg: %d\n", rc);
 867
 868                         /* Make posted receives complete */
 869                         kiblnd_abort_receives(conn);
 870
 871                         /* correct # of posted buffers
 872                          * NB locking needed now I'm racing with completion */
 873                         spin_lock_irqsave(&sched->ibs_lock, flags);
 874                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 875                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 876
 877                         /* cmid will be destroyed by CM(ofed) after cm_callback
 878                          * returned, so we can't refer it anymore
 879                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 880                         rdma_destroy_qp(conn->ibc_cmid);
 881                         conn->ibc_cmid = NULL;
 882
 883                         /* Drop my own and unused rxbuffer refcounts */
 884                         while (i++ <= IBLND_RX_MSGS(version))
 885                                 kiblnd_conn_decref(conn);
 886
 887                         return NULL;
 888                 }
 889         }
 890
 891         /* Init successful! */
 892         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 893                  state == IBLND_CONN_PASSIVE_WAIT);
 894         conn->ibc_state = state;
 895
 896         /* 1 more conn */
 897         atomic_inc(&net->ibn_nconns);
 898         return conn;
 899
 900  failed_2:
 901         kiblnd_destroy_conn(conn);
 902  failed_1:
 903         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 904  failed_0:
 905         return NULL;
 906 }
 907
 908 void
 909 kiblnd_destroy_conn (kib_conn_t *conn)
 910 {
 911         struct rdma_cm_id *cmid = conn->ibc_cmid;
 912         kib_peer_t        *peer = conn->ibc_peer;
 913         int                rc;
 914
 915         LASSERT (!in_interrupt());
 916         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 917         LASSERT(list_empty(&conn->ibc_early_rxs));
 918         LASSERT(list_empty(&conn->ibc_tx_noops));
 919         LASSERT(list_empty(&conn->ibc_tx_queue));
 920         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 921         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 922         LASSERT(list_empty(&conn->ibc_active_txs));
 923         LASSERT (conn->ibc_noops_posted == 0);
 924         LASSERT (conn->ibc_nsends_posted == 0);
 925
 926         switch (conn->ibc_state) {
 927         default:
 928                 /* conn must be completely disengaged from the network */
 929                 LBUG();
 930
 931         case IBLND_CONN_DISCONNECTED:
 932                 /* connvars should have been freed already */
 933                 LASSERT (conn->ibc_connvars == NULL);
 934                 break;
 935
 936         case IBLND_CONN_INIT:
 937                 break;
 938         }
 939
 940         /* conn->ibc_cmid might be destroyed by CM already */
 941         if (cmid != NULL && cmid->qp != NULL)
 942                 rdma_destroy_qp(cmid);
 943
 944         if (conn->ibc_cq != NULL) {
 945                 rc = ib_destroy_cq(conn->ibc_cq);
 946                 if (rc != 0)
 947                         CWARN("Error destroying CQ: %d\n", rc);
 948         }
 949
 950         if (conn->ibc_rx_pages != NULL)
 951                 kiblnd_unmap_rx_descs(conn);
 952
 953         if (conn->ibc_rxs != NULL) {
 954                 LIBCFS_FREE(conn->ibc_rxs,
 955                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 956         }
 957
 958         if (conn->ibc_connvars != NULL)
 959                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 960
 961         if (conn->ibc_hdev != NULL)
 962                 kiblnd_hdev_decref(conn->ibc_hdev);
 963
 964         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 965         if (conn->ibc_state != IBLND_CONN_INIT) {
 966                 kib_net_t *net = peer->ibp_ni->ni_data;
 967
 968                 kiblnd_peer_decref(peer);
 969                 rdma_destroy_id(cmid);
 970                 atomic_dec(&net->ibn_nconns);
 971         }
 972
 973         LIBCFS_FREE(conn, sizeof(*conn));
 974 }
 975
 976 int
 977 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 978 {
 979         kib_conn_t              *conn;
 980         struct list_head        *ctmp;
 981         struct list_head        *cnxt;
 982         int                     count = 0;
 983
 984         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 985                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 986
 987                 CDEBUG(D_NET, "Closing conn -> %s, "
 988                               "version: %x, reason: %d\n",
 989                        libcfs_nid2str(peer->ibp_nid),
 990                        conn->ibc_version, why);
 991
 992                 kiblnd_close_conn_locked(conn, why);
 993                 count++;
 994         }
 995
 996         return count;
 997 }
 998
 999 int
1000 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1001                                 int version, __u64 incarnation)
1002 {
1003         kib_conn_t              *conn;
1004         struct list_head        *ctmp;
1005         struct list_head        *cnxt;
1006         int                     count = 0;
1007
1008         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1009                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1010
1011                 if (conn->ibc_version     == version &&
1012                     conn->ibc_incarnation == incarnation)
1013                         continue;
1014
1015                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1016                               "incarnation:"LPX64"(%x, "LPX64")\n",
1017                        libcfs_nid2str(peer->ibp_nid),
1018                        conn->ibc_version, conn->ibc_incarnation,
1019                        version, incarnation);
1020
1021                 kiblnd_close_conn_locked(conn, -ESTALE);
1022                 count++;
1023         }
1024
1025         return count;
1026 }
1027
1028 static int
1029 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1030 {
1031         kib_peer_t              *peer;
1032         struct list_head        *ptmp;
1033         struct list_head        *pnxt;
1034         int                     lo;
1035         int                     hi;
1036         int                     i;
1037         unsigned long           flags;
1038         int                     count = 0;
1039
1040         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1041
1042         if (nid != LNET_NID_ANY)
1043                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1044         else {
1045                 lo = 0;
1046                 hi = kiblnd_data.kib_peer_hash_size - 1;
1047         }
1048
1049         for (i = lo; i <= hi; i++) {
1050                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1051
1052                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1053                         LASSERT(peer->ibp_connecting > 0 ||
1054                                 peer->ibp_accepting > 0 ||
1055                                 !list_empty(&peer->ibp_conns));
1056
1057                         if (peer->ibp_ni != ni)
1058                                 continue;
1059
1060                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1061                                 continue;
1062
1063                         count += kiblnd_close_peer_conns_locked(peer, 0);
1064                 }
1065         }
1066
1067         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1068
1069         /* wildcards always succeed */
1070         if (nid == LNET_NID_ANY)
1071                 return 0;
1072
1073         return (count == 0) ? -ENOENT : 0;
1074 }
1075
1076 static int
1077 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1078 {
1079         struct libcfs_ioctl_data *data = arg;
1080         int                       rc = -EINVAL;
1081
1082         switch(cmd) {
1083         case IOC_LIBCFS_GET_PEER: {
1084                 lnet_nid_t   nid = 0;
1085                 int          count = 0;
1086
1087                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1088                                           &nid, &count);
1089                 data->ioc_nid    = nid;
1090                 data->ioc_count  = count;
1091                 break;
1092         }
1093
1094         case IOC_LIBCFS_DEL_PEER: {
1095                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1096                 break;
1097         }
1098         case IOC_LIBCFS_GET_CONN: {
1099                 kib_conn_t *conn;
1100
1101                 rc = 0;
1102                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1103                 if (conn == NULL) {
1104                         rc = -ENOENT;
1105                         break;
1106                 }
1107
1108                 LASSERT (conn->ibc_cmid != NULL);
1109                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1110                 if (conn->ibc_cmid->route.path_rec == NULL)
1111                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1112                 else
1113                         data->ioc_u32[0] =
1114                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1115                 kiblnd_conn_decref(conn);
1116                 break;
1117         }
1118         case IOC_LIBCFS_CLOSE_CONNECTION: {
1119                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1120                 break;
1121         }
1122
1123         default:
1124                 break;
1125         }
1126
1127         return rc;
1128 }
1129
1130 static void
1131 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1132 {
1133         cfs_time_t      last_alive = 0;
1134         cfs_time_t      now = cfs_time_current();
1135         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1136         kib_peer_t      *peer;
1137         unsigned long   flags;
1138
1139         read_lock_irqsave(glock, flags);
1140
1141         peer = kiblnd_find_peer_locked(nid);
1142         if (peer != NULL) {
1143                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1144                         peer->ibp_accepting > 0 ||
1145                         !list_empty(&peer->ibp_conns));  /* active conn */
1146                 last_alive = peer->ibp_last_alive;
1147         }
1148
1149         read_unlock_irqrestore(glock, flags);
1150
1151         if (last_alive != 0)
1152                 *when = last_alive;
1153
1154         /* peer is not persistent in hash, trigger peer creation
1155          * and connection establishment with a NULL tx */
1156         if (peer == NULL)
1157                 kiblnd_launch_tx(ni, NULL, nid);
1158
1159         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1160                libcfs_nid2str(nid), peer,
1161                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1162         return;
1163 }
1164
1165 static void
1166 kiblnd_free_pages(kib_pages_t *p)
1167 {
1168         int     npages = p->ibp_npages;
1169         int     i;
1170
1171         for (i = 0; i < npages; i++) {
1172                 if (p->ibp_pages[i] != NULL)
1173                         __free_page(p->ibp_pages[i]);
1174         }
1175
1176         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1177 }
1178
1179 int
1180 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1181 {
1182         kib_pages_t     *p;
1183         int             i;
1184
1185         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1186                          offsetof(kib_pages_t, ibp_pages[npages]));
1187         if (p == NULL) {
1188                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1189                 return -ENOMEM;
1190         }
1191
1192         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1193         p->ibp_npages = npages;
1194
1195         for (i = 0; i < npages; i++) {
1196                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1197                                                      GFP_NOFS);
1198                 if (p->ibp_pages[i] == NULL) {
1199                         CERROR("Can't allocate page %d of %d\n", i, npages);
1200                         kiblnd_free_pages(p);
1201                         return -ENOMEM;
1202                 }
1203         }
1204
1205         *pp = p;
1206         return 0;
1207 }
1208
1209 void
1210 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1211 {
1212         kib_rx_t *rx;
1213         int       i;
1214
1215         LASSERT (conn->ibc_rxs != NULL);
1216         LASSERT (conn->ibc_hdev != NULL);
1217
1218         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1219                 rx = &conn->ibc_rxs[i];
1220
1221                 LASSERT (rx->rx_nob >= 0); /* not posted */
1222
1223                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1224                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1225                                                           rx->rx_msgaddr),
1226                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1227         }
1228
1229         kiblnd_free_pages(conn->ibc_rx_pages);
1230
1231         conn->ibc_rx_pages = NULL;
1232 }
1233
1234 void
1235 kiblnd_map_rx_descs(kib_conn_t *conn)
1236 {
1237         kib_rx_t       *rx;
1238         struct page    *pg;
1239         int             pg_off;
1240         int             ipg;
1241         int             i;
1242
1243         for (pg_off = ipg = i = 0;
1244              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1245                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1246                 rx = &conn->ibc_rxs[i];
1247
1248                 rx->rx_conn = conn;
1249                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1250
1251                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1252                                                        rx->rx_msg, IBLND_MSG_SIZE,
1253                                                        DMA_FROM_DEVICE);
1254                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1255                                                    rx->rx_msgaddr));
1256                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1257
1258                 CDEBUG(D_NET, "rx %d: %p "LPX64"("LPX64")\n",
1259                        i, rx->rx_msg, rx->rx_msgaddr,
1260                        (__u64)(page_to_phys(pg) + pg_off));
1261
1262                 pg_off += IBLND_MSG_SIZE;
1263                 LASSERT (pg_off <= PAGE_SIZE);
1264
1265                 if (pg_off == PAGE_SIZE) {
1266                         pg_off = 0;
1267                         ipg++;
1268                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1269                 }
1270         }
1271 }
1272
1273 static void
1274 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1275 {
1276         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1277         kib_tx_t       *tx;
1278         int             i;
1279
1280         LASSERT (tpo->tpo_pool.po_allocated == 0);
1281
1282         if (hdev == NULL)
1283                 return;
1284
1285         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1286                 tx = &tpo->tpo_tx_descs[i];
1287                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1288                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1289                                                           tx->tx_msgaddr),
1290                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1291         }
1292
1293         kiblnd_hdev_decref(hdev);
1294         tpo->tpo_hdev = NULL;
1295 }
1296
1297 static kib_hca_dev_t *
1298 kiblnd_current_hdev(kib_dev_t *dev)
1299 {
1300         kib_hca_dev_t *hdev;
1301         unsigned long  flags;
1302         int            i = 0;
1303
1304         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1305         while (dev->ibd_failover) {
1306                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1307                 if (i++ % 50 == 0)
1308                         CDEBUG(D_NET, "%s: Wait for failover\n",
1309                                dev->ibd_ifname);
1310                 schedule_timeout(cfs_time_seconds(1) / 100);
1311
1312                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1313         }
1314
1315         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1316         hdev = dev->ibd_hdev;
1317
1318         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1319
1320         return hdev;
1321 }
1322
1323 static void
1324 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1325 {
1326         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1327         kib_pool_t     *pool  = &tpo->tpo_pool;
1328         kib_net_t      *net   = pool->po_owner->ps_net;
1329         kib_dev_t      *dev;
1330         struct page    *page;
1331         kib_tx_t       *tx;
1332         int             page_offset;
1333         int             ipage;
1334         int             i;
1335
1336         LASSERT (net != NULL);
1337
1338         dev = net->ibn_dev;
1339
1340         /* pre-mapped messages are not bigger than 1 page */
1341         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1342
1343         /* No fancy arithmetic when we do the buffer calculations */
1344         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1345
1346         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1347
1348         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1349                 page = txpgs->ibp_pages[ipage];
1350                 tx = &tpo->tpo_tx_descs[i];
1351
1352                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1353                                            page_offset);
1354
1355                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1356                                                        tx->tx_msg,
1357                                                        IBLND_MSG_SIZE,
1358                                                        DMA_TO_DEVICE);
1359                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1360                                                   tx->tx_msgaddr));
1361                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1362
1363                 list_add(&tx->tx_list, &pool->po_free_list);
1364
1365                 page_offset += IBLND_MSG_SIZE;
1366                 LASSERT(page_offset <= PAGE_SIZE);
1367
1368                 if (page_offset == PAGE_SIZE) {
1369                         page_offset = 0;
1370                         ipage++;
1371                         LASSERT(ipage <= txpgs->ibp_npages);
1372                 }
1373         }
1374 }
1375
1376 struct ib_mr *
1377 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1378 {
1379         __u64   index;
1380
1381         LASSERT (hdev->ibh_mrs[0] != NULL);
1382
1383         if (hdev->ibh_nmrs == 1)
1384                 return hdev->ibh_mrs[0];
1385
1386         index = addr >> hdev->ibh_mr_shift;
1387
1388         if (index <  hdev->ibh_nmrs &&
1389             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1390                 return hdev->ibh_mrs[index];
1391
1392         return NULL;
1393 }
1394
1395 struct ib_mr *
1396 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1397 {
1398         struct ib_mr *prev_mr;
1399         struct ib_mr *mr;
1400         int           i;
1401
1402         LASSERT (hdev->ibh_mrs[0] != NULL);
1403
1404         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1405             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1406                 return NULL;
1407
1408         if (hdev->ibh_nmrs == 1)
1409                 return hdev->ibh_mrs[0];
1410
1411         for (i = 0, mr = prev_mr = NULL;
1412              i < rd->rd_nfrags; i++) {
1413                 mr = kiblnd_find_dma_mr(hdev,
1414                                         rd->rd_frags[i].rf_addr,
1415                                         rd->rd_frags[i].rf_nob);
1416                 if (prev_mr == NULL)
1417                         prev_mr = mr;
1418
1419                 if (mr == NULL || prev_mr != mr) {
1420                         /* Can't covered by one single MR */
1421                         mr = NULL;
1422                         break;
1423                 }
1424         }
1425
1426         return mr;
1427 }
1428
1429 static void
1430 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1431 {
1432         LASSERT (pool->fpo_map_count == 0);
1433
1434         if (pool->fpo_fmr_pool != NULL)
1435                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1436
1437         if (pool->fpo_hdev != NULL)
1438                 kiblnd_hdev_decref(pool->fpo_hdev);
1439
1440         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1441 }
1442
1443 static void
1444 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1445 {
1446         kib_fmr_pool_t *pool;
1447
1448         while (!list_empty(head)) {
1449                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1450                 list_del(&pool->fpo_list);
1451                 kiblnd_destroy_fmr_pool(pool);
1452         }
1453 }
1454
1455 static int kiblnd_fmr_pool_size(int ncpts)
1456 {
1457         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1458
1459         return max(IBLND_FMR_POOL, size);
1460 }
1461
1462 static int kiblnd_fmr_flush_trigger(int ncpts)
1463 {
1464         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1465
1466         return max(IBLND_FMR_POOL_FLUSH, size);
1467 }
1468
1469 static int
1470 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1471 {
1472         /* FMR pool for RDMA */
1473         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1474         kib_fmr_pool_t          *fpo;
1475         struct ib_fmr_pool_param param = {
1476                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1477                 .page_shift        = PAGE_SHIFT,
1478                 .access            = (IB_ACCESS_LOCAL_WRITE |
1479                                       IB_ACCESS_REMOTE_WRITE),
1480                 .pool_size         = fps->fps_pool_size,
1481                 .dirty_watermark   = fps->fps_flush_trigger,
1482                 .flush_function    = NULL,
1483                 .flush_arg         = NULL,
1484                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1485         int rc;
1486
1487         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1488         if (fpo == NULL)
1489                 return -ENOMEM;
1490
1491         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1492
1493         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1494         if (IS_ERR(fpo->fpo_fmr_pool)) {
1495                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1496                 CERROR("Failed to create FMR pool: %d\n", rc);
1497
1498                 kiblnd_hdev_decref(fpo->fpo_hdev);
1499                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1500                 return rc;
1501         }
1502
1503         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1504         fpo->fpo_owner    = fps;
1505         *pp_fpo = fpo;
1506
1507         return 0;
1508 }
1509
1510 static void
1511 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1512 {
1513         if (fps->fps_net == NULL) /* intialized? */
1514                 return;
1515
1516         spin_lock(&fps->fps_lock);
1517
1518         while (!list_empty(&fps->fps_pool_list)) {
1519                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1520                                                  kib_fmr_pool_t, fpo_list);
1521                 fpo->fpo_failed = 1;
1522                 list_del(&fpo->fpo_list);
1523                 if (fpo->fpo_map_count == 0)
1524                         list_add(&fpo->fpo_list, zombies);
1525                 else
1526                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1527         }
1528
1529         spin_unlock(&fps->fps_lock);
1530 }
1531
1532 static void
1533 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1534 {
1535         if (fps->fps_net != NULL) { /* initialized? */
1536                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1537                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1538         }
1539 }
1540
1541 static int
1542 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1543                         int pool_size, int flush_trigger)
1544 {
1545         kib_fmr_pool_t *fpo;
1546         int             rc;
1547
1548         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1549
1550         fps->fps_net = net;
1551         fps->fps_cpt = cpt;
1552         fps->fps_pool_size = pool_size;
1553         fps->fps_flush_trigger = flush_trigger;
1554         spin_lock_init(&fps->fps_lock);
1555         INIT_LIST_HEAD(&fps->fps_pool_list);
1556         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1557
1558         rc = kiblnd_create_fmr_pool(fps, &fpo);
1559         if (rc == 0)
1560                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1561
1562         return rc;
1563 }
1564
1565 static int
1566 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1567 {
1568         if (fpo->fpo_map_count != 0) /* still in use */
1569                 return 0;
1570         if (fpo->fpo_failed)
1571                 return 1;
1572         return cfs_time_aftereq(now, fpo->fpo_deadline);
1573 }
1574
1575 void
1576 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1577 {
1578         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1579         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1580         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1581         cfs_time_t         now = cfs_time_current();
1582         kib_fmr_pool_t    *tmp;
1583         int                rc;
1584
1585         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1586         LASSERT(rc == 0);
1587
1588         if (status != 0) {
1589                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1590                 LASSERT(rc == 0);
1591         }
1592
1593         fmr->fmr_pool = NULL;
1594         fmr->fmr_pfmr = NULL;
1595
1596         spin_lock(&fps->fps_lock);
1597         fpo->fpo_map_count--;   /* decref the pool */
1598
1599         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1600                 /* the first pool is persistent */
1601                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1602                         continue;
1603
1604                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1605                         list_move(&fpo->fpo_list, &zombies);
1606                         fps->fps_version++;
1607                 }
1608         }
1609         spin_unlock(&fps->fps_lock);
1610
1611         if (!list_empty(&zombies))
1612                 kiblnd_destroy_fmr_pool_list(&zombies);
1613 }
1614
1615 int
1616 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1617                     __u64 iov, kib_fmr_t *fmr)
1618 {
1619         struct ib_pool_fmr *pfmr;
1620         kib_fmr_pool_t     *fpo;
1621         __u64               version;
1622         int                 rc;
1623
1624 again:
1625         spin_lock(&fps->fps_lock);
1626         version = fps->fps_version;
1627         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1628                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1629                 fpo->fpo_map_count++;
1630                 spin_unlock(&fps->fps_lock);
1631
1632                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1633                                             pages, npages, iov);
1634                 if (likely(!IS_ERR(pfmr))) {
1635                         fmr->fmr_pool = fpo;
1636                         fmr->fmr_pfmr = pfmr;
1637                         return 0;
1638                 }
1639
1640                 spin_lock(&fps->fps_lock);
1641                 fpo->fpo_map_count--;
1642                 if (PTR_ERR(pfmr) != -EAGAIN) {
1643                         spin_unlock(&fps->fps_lock);
1644                         return PTR_ERR(pfmr);
1645                 }
1646
1647                 /* EAGAIN and ... */
1648                 if (version != fps->fps_version) {
1649                         spin_unlock(&fps->fps_lock);
1650                         goto again;
1651                 }
1652         }
1653
1654         if (fps->fps_increasing) {
1655                 spin_unlock(&fps->fps_lock);
1656                 CDEBUG(D_NET, "Another thread is allocating new "
1657                        "FMR pool, waiting for her to complete\n");
1658                 schedule();
1659                 goto again;
1660
1661         }
1662
1663         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1664                 /* someone failed recently */
1665                 spin_unlock(&fps->fps_lock);
1666                 return -EAGAIN;
1667         }
1668
1669         fps->fps_increasing = 1;
1670         spin_unlock(&fps->fps_lock);
1671
1672         CDEBUG(D_NET, "Allocate new FMR pool\n");
1673         rc = kiblnd_create_fmr_pool(fps, &fpo);
1674         spin_lock(&fps->fps_lock);
1675         fps->fps_increasing = 0;
1676         if (rc == 0) {
1677                 fps->fps_version++;
1678                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1679         } else {
1680                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1681         }
1682         spin_unlock(&fps->fps_lock);
1683
1684         goto again;
1685 }
1686
1687 static void
1688 kiblnd_fini_pool(kib_pool_t *pool)
1689 {
1690         LASSERT(list_empty(&pool->po_free_list));
1691         LASSERT(pool->po_allocated == 0);
1692
1693         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1694 }
1695
1696 static void
1697 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1698 {
1699         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1700
1701         memset(pool, 0, sizeof(kib_pool_t));
1702         INIT_LIST_HEAD(&pool->po_free_list);
1703         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1704         pool->po_owner    = ps;
1705         pool->po_size     = size;
1706 }
1707
1708 static void
1709 kiblnd_destroy_pool_list(struct list_head *head)
1710 {
1711         kib_pool_t *pool;
1712
1713         while (!list_empty(head)) {
1714                 pool = list_entry(head->next, kib_pool_t, po_list);
1715                 list_del(&pool->po_list);
1716
1717                 LASSERT(pool->po_owner != NULL);
1718                 pool->po_owner->ps_pool_destroy(pool);
1719         }
1720 }
1721
1722 static void
1723 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1724 {
1725         if (ps->ps_net == NULL) /* intialized? */
1726                 return;
1727
1728         spin_lock(&ps->ps_lock);
1729         while (!list_empty(&ps->ps_pool_list)) {
1730                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1731                                             kib_pool_t, po_list);
1732                 po->po_failed = 1;
1733                 list_del(&po->po_list);
1734                 if (po->po_allocated == 0)
1735                         list_add(&po->po_list, zombies);
1736                 else
1737                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1738         }
1739         spin_unlock(&ps->ps_lock);
1740 }
1741
1742 static void
1743 kiblnd_fini_poolset(kib_poolset_t *ps)
1744 {
1745         if (ps->ps_net != NULL) { /* initialized? */
1746                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1747                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1748         }
1749 }
1750
1751 static int
1752 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1753                     kib_net_t *net, char *name, int size,
1754                     kib_ps_pool_create_t po_create,
1755                     kib_ps_pool_destroy_t po_destroy,
1756                     kib_ps_node_init_t nd_init,
1757                     kib_ps_node_fini_t nd_fini)
1758 {
1759         kib_pool_t      *pool;
1760         int             rc;
1761
1762         memset(ps, 0, sizeof(kib_poolset_t));
1763
1764         ps->ps_cpt          = cpt;
1765         ps->ps_net          = net;
1766         ps->ps_pool_create  = po_create;
1767         ps->ps_pool_destroy = po_destroy;
1768         ps->ps_node_init    = nd_init;
1769         ps->ps_node_fini    = nd_fini;
1770         ps->ps_pool_size    = size;
1771         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1772             >= sizeof(ps->ps_name))
1773                 return -E2BIG;
1774         spin_lock_init(&ps->ps_lock);
1775         INIT_LIST_HEAD(&ps->ps_pool_list);
1776         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1777
1778         rc = ps->ps_pool_create(ps, size, &pool);
1779         if (rc == 0)
1780                 list_add(&pool->po_list, &ps->ps_pool_list);
1781         else
1782                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1783
1784         return rc;
1785 }
1786
1787 static int
1788 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1789 {
1790         if (pool->po_allocated != 0) /* still in use */
1791                 return 0;
1792         if (pool->po_failed)
1793                 return 1;
1794         return cfs_time_aftereq(now, pool->po_deadline);
1795 }
1796
1797 void
1798 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1799 {
1800         struct list_head zombies = LIST_HEAD_INIT(zombies);
1801         kib_poolset_t   *ps = pool->po_owner;
1802         kib_pool_t      *tmp;
1803         cfs_time_t       now = cfs_time_current();
1804
1805         spin_lock(&ps->ps_lock);
1806
1807         if (ps->ps_node_fini != NULL)
1808                 ps->ps_node_fini(pool, node);
1809
1810         LASSERT(pool->po_allocated > 0);
1811         list_add(node, &pool->po_free_list);
1812         pool->po_allocated--;
1813
1814         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1815                 /* the first pool is persistent */
1816                 if (ps->ps_pool_list.next == &pool->po_list)
1817                         continue;
1818
1819                 if (kiblnd_pool_is_idle(pool, now))
1820                         list_move(&pool->po_list, &zombies);
1821         }
1822         spin_unlock(&ps->ps_lock);
1823
1824         if (!list_empty(&zombies))
1825                 kiblnd_destroy_pool_list(&zombies);
1826 }
1827
1828 struct list_head *
1829 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1830 {
1831         struct list_head        *node;
1832         kib_pool_t              *pool;
1833         int                     rc;
1834
1835 again:
1836         spin_lock(&ps->ps_lock);
1837         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1838                 if (list_empty(&pool->po_free_list))
1839                         continue;
1840
1841                 pool->po_allocated++;
1842                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1843                 node = pool->po_free_list.next;
1844                 list_del(node);
1845
1846                 if (ps->ps_node_init != NULL) {
1847                         /* still hold the lock */
1848                         ps->ps_node_init(pool, node);
1849                 }
1850                 spin_unlock(&ps->ps_lock);
1851                 return node;
1852         }
1853
1854         /* no available tx pool and ... */
1855         if (ps->ps_increasing) {
1856                 /* another thread is allocating a new pool */
1857                 spin_unlock(&ps->ps_lock);
1858                 CDEBUG(D_NET, "Another thread is allocating new "
1859                        "%s pool, waiting for her to complete\n",
1860                        ps->ps_name);
1861                 schedule();
1862                 goto again;
1863         }
1864
1865         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1866                 /* someone failed recently */
1867                 spin_unlock(&ps->ps_lock);
1868                 return NULL;
1869         }
1870
1871         ps->ps_increasing = 1;
1872         spin_unlock(&ps->ps_lock);
1873
1874         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1875
1876         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1877
1878         spin_lock(&ps->ps_lock);
1879         ps->ps_increasing = 0;
1880         if (rc == 0) {
1881                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1882         } else {
1883                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1884                 CERROR("Can't allocate new %s pool because out of memory\n",
1885                        ps->ps_name);
1886         }
1887         spin_unlock(&ps->ps_lock);
1888
1889         goto again;
1890 }
1891
1892 void
1893 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1894 {
1895         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1896         struct ib_mr        *mr  = pmr->pmr_mr;
1897
1898         pmr->pmr_mr = NULL;
1899         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1900         if (mr != NULL)
1901                 ib_dereg_mr(mr);
1902 }
1903
1904 int
1905 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1906                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1907 {
1908         kib_phys_mr_t    *pmr;
1909         struct list_head *node;
1910         int               rc;
1911         int               i;
1912
1913         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1914         if (node == NULL) {
1915                 CERROR("Failed to allocate PMR descriptor\n");
1916                 return -ENOMEM;
1917         }
1918
1919         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1920         if (pmr->pmr_pool->ppo_hdev != hdev) {
1921                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1922                 return -EAGAIN;
1923         }
1924
1925         for (i = 0; i < rd->rd_nfrags; i ++) {
1926                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1927                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1928         }
1929
1930         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1931                                      pmr->pmr_ipb, rd->rd_nfrags,
1932                                      IB_ACCESS_LOCAL_WRITE |
1933                                      IB_ACCESS_REMOTE_WRITE,
1934                                      iova);
1935         if (!IS_ERR(pmr->pmr_mr)) {
1936                 pmr->pmr_iova = *iova;
1937                 *pp_pmr = pmr;
1938                 return 0;
1939         }
1940
1941         rc = PTR_ERR(pmr->pmr_mr);
1942         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1943
1944         pmr->pmr_mr = NULL;
1945         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1946
1947         return rc;
1948 }
1949
1950 static void
1951 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1952 {
1953         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1954         kib_phys_mr_t  *pmr;
1955
1956         LASSERT (pool->po_allocated == 0);
1957
1958         while (!list_empty(&pool->po_free_list)) {
1959                 pmr = list_entry(pool->po_free_list.next,
1960                                      kib_phys_mr_t, pmr_list);
1961
1962                 LASSERT (pmr->pmr_mr == NULL);
1963                 list_del(&pmr->pmr_list);
1964
1965                 if (pmr->pmr_ipb != NULL) {
1966                         LIBCFS_FREE(pmr->pmr_ipb,
1967                                     IBLND_MAX_RDMA_FRAGS *
1968                                     sizeof(struct ib_phys_buf));
1969                 }
1970
1971                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1972         }
1973
1974         kiblnd_fini_pool(pool);
1975         if (ppo->ppo_hdev != NULL)
1976                 kiblnd_hdev_decref(ppo->ppo_hdev);
1977
1978         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1979 }
1980
1981 static inline int kiblnd_pmr_pool_size(int ncpts)
1982 {
1983         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1984
1985         return max(IBLND_PMR_POOL, size);
1986 }
1987
1988 static int
1989 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1990 {
1991         struct kib_pmr_pool     *ppo;
1992         struct kib_pool         *pool;
1993         kib_phys_mr_t           *pmr;
1994         int                     i;
1995
1996         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
1997                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
1998         if (ppo == NULL) {
1999                 CERROR("Failed to allocate PMR pool\n");
2000                 return -ENOMEM;
2001         }
2002
2003         pool = &ppo->ppo_pool;
2004         kiblnd_init_pool(ps, pool, size);
2005
2006         for (i = 0; i < size; i++) {
2007                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2008                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2009                 if (pmr == NULL)
2010                         break;
2011
2012                 pmr->pmr_pool = ppo;
2013                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2014                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2015                 if (pmr->pmr_ipb == NULL)
2016                         break;
2017
2018                 list_add(&pmr->pmr_list, &pool->po_free_list);
2019         }
2020
2021         if (i < size) {
2022                 ps->ps_pool_destroy(pool);
2023                 return -ENOMEM;
2024         }
2025
2026         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2027         *pp_po = pool;
2028         return 0;
2029 }
2030
2031 static void
2032 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2033 {
2034         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2035         int             i;
2036
2037         LASSERT (pool->po_allocated == 0);
2038
2039         if (tpo->tpo_tx_pages != NULL) {
2040                 kiblnd_unmap_tx_pool(tpo);
2041                 kiblnd_free_pages(tpo->tpo_tx_pages);
2042         }
2043
2044         if (tpo->tpo_tx_descs == NULL)
2045                 goto out;
2046
2047         for (i = 0; i < pool->po_size; i++) {
2048                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2049
2050                 list_del(&tx->tx_list);
2051                 if (tx->tx_pages != NULL)
2052                         LIBCFS_FREE(tx->tx_pages,
2053                                     LNET_MAX_IOV *
2054                                     sizeof(*tx->tx_pages));
2055                 if (tx->tx_frags != NULL)
2056                         LIBCFS_FREE(tx->tx_frags,
2057                                     IBLND_MAX_RDMA_FRAGS *
2058                                             sizeof(*tx->tx_frags));
2059                 if (tx->tx_wrq != NULL)
2060                         LIBCFS_FREE(tx->tx_wrq,
2061                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2062                                     sizeof(*tx->tx_wrq));
2063                 if (tx->tx_sge != NULL)
2064                         LIBCFS_FREE(tx->tx_sge,
2065                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2066                                     sizeof(*tx->tx_sge));
2067                 if (tx->tx_rd != NULL)
2068                         LIBCFS_FREE(tx->tx_rd,
2069                                     offsetof(kib_rdma_desc_t,
2070                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2071         }
2072
2073         LIBCFS_FREE(tpo->tpo_tx_descs,
2074                     pool->po_size * sizeof(kib_tx_t));
2075 out:
2076         kiblnd_fini_pool(pool);
2077         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2078 }
2079
2080 static int kiblnd_tx_pool_size(int ncpts)
2081 {
2082         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2083
2084         return max(IBLND_TX_POOL, ntx);
2085 }
2086
2087 static int
2088 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2089 {
2090         int            i;
2091         int            npg;
2092         kib_pool_t    *pool;
2093         kib_tx_pool_t *tpo;
2094
2095         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2096         if (tpo == NULL) {
2097                 CERROR("Failed to allocate TX pool\n");
2098                 return -ENOMEM;
2099         }
2100
2101         pool = &tpo->tpo_pool;
2102         kiblnd_init_pool(ps, pool, size);
2103         tpo->tpo_tx_descs = NULL;
2104         tpo->tpo_tx_pages = NULL;
2105
2106         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2107         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2108                 CERROR("Can't allocate tx pages: %d\n", npg);
2109                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2110                 return -ENOMEM;
2111         }
2112
2113         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2114                          size * sizeof(kib_tx_t));
2115         if (tpo->tpo_tx_descs == NULL) {
2116                 CERROR("Can't allocate %d tx descriptors\n", size);
2117                 ps->ps_pool_destroy(pool);
2118                 return -ENOMEM;
2119         }
2120
2121         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2122
2123         for (i = 0; i < size; i++) {
2124                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2125
2126                 tx->tx_pool = tpo;
2127                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2128                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2129                                          lnet_cpt_table(), ps->ps_cpt,
2130                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2131                         if (tx->tx_pages == NULL)
2132                                 break;
2133                 }
2134
2135                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2136                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2137                 if (tx->tx_frags == NULL)
2138                         break;
2139
2140                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2141
2142                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2143                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2144                                  sizeof(*tx->tx_wrq));
2145                 if (tx->tx_wrq == NULL)
2146                         break;
2147
2148                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2149                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2150                                  sizeof(*tx->tx_sge));
2151                 if (tx->tx_sge == NULL)
2152                         break;
2153
2154                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2155                                  offsetof(kib_rdma_desc_t,
2156                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2157                 if (tx->tx_rd == NULL)
2158                         break;
2159         }
2160
2161         if (i == size) {
2162                 kiblnd_map_tx_pool(tpo);
2163                 *pp_po = pool;
2164                 return 0;
2165         }
2166
2167         ps->ps_pool_destroy(pool);
2168         return -ENOMEM;
2169 }
2170
2171 static void
2172 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2173 {
2174         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2175                                              tps_poolset);
2176         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2177
2178         tx->tx_cookie = tps->tps_next_tx_cookie++;
2179 }
2180
2181 static void
2182 kiblnd_net_fini_pools(kib_net_t *net)
2183 {
2184         int     i;
2185
2186         cfs_cpt_for_each(i, lnet_cpt_table()) {
2187                 kib_tx_poolset_t        *tps;
2188                 kib_fmr_poolset_t       *fps;
2189                 kib_pmr_poolset_t       *pps;
2190
2191                 if (net->ibn_tx_ps != NULL) {
2192                         tps = net->ibn_tx_ps[i];
2193                         kiblnd_fini_poolset(&tps->tps_poolset);
2194                 }
2195
2196                 if (net->ibn_fmr_ps != NULL) {
2197                         fps = net->ibn_fmr_ps[i];
2198                         kiblnd_fini_fmr_poolset(fps);
2199                 }
2200
2201                 if (net->ibn_pmr_ps != NULL) {
2202                         pps = net->ibn_pmr_ps[i];
2203                         kiblnd_fini_poolset(&pps->pps_poolset);
2204                 }
2205         }
2206
2207         if (net->ibn_tx_ps != NULL) {
2208                 cfs_percpt_free(net->ibn_tx_ps);
2209                 net->ibn_tx_ps = NULL;
2210         }
2211
2212         if (net->ibn_fmr_ps != NULL) {
2213                 cfs_percpt_free(net->ibn_fmr_ps);
2214                 net->ibn_fmr_ps = NULL;
2215         }
2216
2217         if (net->ibn_pmr_ps != NULL) {
2218                 cfs_percpt_free(net->ibn_pmr_ps);
2219                 net->ibn_pmr_ps = NULL;
2220         }
2221 }
2222
2223 static int
2224 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2225 {
2226         unsigned long   flags;
2227         int             cpt;
2228         int             rc;
2229         int             i;
2230
2231         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2232         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2233             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2234                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2235                                            flags);
2236                 goto create_tx_pool;
2237         }
2238
2239         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2240
2241         if (*kiblnd_tunables.kib_fmr_pool_size <
2242             *kiblnd_tunables.kib_ntx / 4) {
2243                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2244                        *kiblnd_tunables.kib_fmr_pool_size,
2245                        *kiblnd_tunables.kib_ntx / 4);
2246                 rc = -EINVAL;
2247                 goto failed;
2248         }
2249
2250         /* TX pool must be created later than FMR/PMR, see LU-2268
2251          * for details */
2252         LASSERT(net->ibn_tx_ps == NULL);
2253
2254         /* premapping can fail if ibd_nmr > 1, so we always create
2255          * FMR/PMR pool and map-on-demand if premapping failed */
2256
2257         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2258                                            sizeof(kib_fmr_poolset_t));
2259         if (net->ibn_fmr_ps == NULL) {
2260                 CERROR("Failed to allocate FMR pool array\n");
2261                 rc = -ENOMEM;
2262                 goto failed;
2263         }
2264
2265         for (i = 0; i < ncpts; i++) {
2266                 cpt = (cpts == NULL) ? i : cpts[i];
2267                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2268                                              kiblnd_fmr_pool_size(ncpts),
2269                                              kiblnd_fmr_flush_trigger(ncpts));
2270                 if (rc == -ENOSYS && i == 0) /* no FMR */
2271                         break; /* create PMR pool */
2272
2273                 if (rc != 0) { /* a real error */
2274                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2275                                cpt, rc);
2276                         goto failed;
2277                 }
2278         }
2279
2280         if (i > 0) {
2281                 LASSERT(i == ncpts);
2282                 goto create_tx_pool;
2283         }
2284
2285         cfs_percpt_free(net->ibn_fmr_ps);
2286         net->ibn_fmr_ps = NULL;
2287
2288         CWARN("Device does not support FMR, failing back to PMR\n");
2289
2290         if (*kiblnd_tunables.kib_pmr_pool_size <
2291             *kiblnd_tunables.kib_ntx / 4) {
2292                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2293                        *kiblnd_tunables.kib_pmr_pool_size,
2294                        *kiblnd_tunables.kib_ntx / 4);
2295                 rc = -EINVAL;
2296                 goto failed;
2297         }
2298
2299         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2300                                            sizeof(kib_pmr_poolset_t));
2301         if (net->ibn_pmr_ps == NULL) {
2302                 CERROR("Failed to allocate PMR pool array\n");
2303                 rc = -ENOMEM;
2304                 goto failed;
2305         }
2306
2307         for (i = 0; i < ncpts; i++) {
2308                 cpt = (cpts == NULL) ? i : cpts[i];
2309                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2310                                          cpt, net, "PMR",
2311                                          kiblnd_pmr_pool_size(ncpts),
2312                                          kiblnd_create_pmr_pool,
2313                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2314                 if (rc != 0) {
2315                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2316                                cpt, rc);
2317                         goto failed;
2318                 }
2319         }
2320
2321  create_tx_pool:
2322         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2323                                           sizeof(kib_tx_poolset_t));
2324         if (net->ibn_tx_ps == NULL) {
2325                 CERROR("Failed to allocate tx pool array\n");
2326                 rc = -ENOMEM;
2327                 goto failed;
2328         }
2329
2330         for (i = 0; i < ncpts; i++) {
2331                 cpt = (cpts == NULL) ? i : cpts[i];
2332                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2333                                          cpt, net, "TX",
2334                                          kiblnd_tx_pool_size(ncpts),
2335                                          kiblnd_create_tx_pool,
2336                                          kiblnd_destroy_tx_pool,
2337                                          kiblnd_tx_init, NULL);
2338                 if (rc != 0) {
2339                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2340                                cpt, rc);
2341                         goto failed;
2342                 }
2343         }
2344
2345         return 0;
2346  failed:
2347         kiblnd_net_fini_pools(net);
2348         LASSERT(rc != 0);
2349         return rc;
2350 }
2351
2352 static int
2353 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2354 {
2355         struct ib_device_attr *attr;
2356         int                    rc;
2357
2358         /* It's safe to assume a HCA can handle a page size
2359          * matching that of the native system */
2360         hdev->ibh_page_shift = PAGE_SHIFT;
2361         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2362         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2363
2364         LIBCFS_ALLOC(attr, sizeof(*attr));
2365         if (attr == NULL) {
2366                 CERROR("Out of memory\n");
2367                 return -ENOMEM;
2368         }
2369
2370         rc = ib_query_device(hdev->ibh_ibdev, attr);
2371         if (rc == 0)
2372                 hdev->ibh_mr_size = attr->max_mr_size;
2373
2374         LIBCFS_FREE(attr, sizeof(*attr));
2375
2376         if (rc != 0) {
2377                 CERROR("Failed to query IB device: %d\n", rc);
2378                 return rc;
2379         }
2380
2381         if (hdev->ibh_mr_size == ~0ULL) {
2382                 hdev->ibh_mr_shift = 64;
2383                 return 0;
2384         }
2385
2386         for (hdev->ibh_mr_shift = 0;
2387              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2388                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2389                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2390                         return 0;
2391         }
2392
2393         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2394         return -EINVAL;
2395 }
2396
2397 static void
2398 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2399 {
2400         int     i;
2401
2402         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2403                 return;
2404
2405         for (i = 0; i < hdev->ibh_nmrs; i++) {
2406                 if (hdev->ibh_mrs[i] == NULL)
2407                         break;
2408
2409                 ib_dereg_mr(hdev->ibh_mrs[i]);
2410         }
2411
2412         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2413         hdev->ibh_mrs  = NULL;
2414         hdev->ibh_nmrs = 0;
2415 }
2416
2417 void
2418 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2419 {
2420         kiblnd_hdev_cleanup_mrs(hdev);
2421
2422         if (hdev->ibh_pd != NULL)
2423                 ib_dealloc_pd(hdev->ibh_pd);
2424
2425         if (hdev->ibh_cmid != NULL)
2426                 rdma_destroy_id(hdev->ibh_cmid);
2427
2428         LIBCFS_FREE(hdev, sizeof(*hdev));
2429 }
2430
2431 static int
2432 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2433 {
2434         struct ib_mr *mr;
2435         int           i;
2436         int           rc;
2437         __u64         mm_size;
2438         __u64         mr_size;
2439         int           acflags = IB_ACCESS_LOCAL_WRITE |
2440                                 IB_ACCESS_REMOTE_WRITE;
2441
2442         rc = kiblnd_hdev_get_attr(hdev);
2443         if (rc != 0)
2444                 return rc;
2445
2446         if (hdev->ibh_mr_shift == 64) {
2447                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2448                 if (hdev->ibh_mrs == NULL) {
2449                         CERROR("Failed to allocate MRs table\n");
2450                         return -ENOMEM;
2451                 }
2452
2453                 hdev->ibh_mrs[0] = NULL;
2454                 hdev->ibh_nmrs   = 1;
2455
2456                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2457                 if (IS_ERR(mr)) {
2458                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2459                         kiblnd_hdev_cleanup_mrs(hdev);
2460                         return PTR_ERR(mr);
2461                 }
2462
2463                 hdev->ibh_mrs[0] = mr;
2464
2465                 goto out;
2466         }
2467
2468         mr_size = (1ULL << hdev->ibh_mr_shift);
2469         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2470
2471         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2472
2473         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2474                 /* it's 4T..., assume we will re-code at that time */
2475                 CERROR("Can't support memory size: x"LPX64
2476                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2477                 return -EINVAL;
2478         }
2479
2480         /* create an array of MRs to cover all memory */
2481         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2482         if (hdev->ibh_mrs == NULL) {
2483                 CERROR("Failed to allocate MRs' table\n");
2484                 return -ENOMEM;
2485         }
2486
2487         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2488
2489         for (i = 0; i < hdev->ibh_nmrs; i++) {
2490                 struct ib_phys_buf ipb;
2491                 __u64              iova;
2492
2493                 ipb.size = hdev->ibh_mr_size;
2494                 ipb.addr = i * mr_size;
2495                 iova     = ipb.addr;
2496
2497                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2498                 if (IS_ERR(mr)) {
2499                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2500                                " size "LPX64" : %ld\n",
2501                                ipb.addr, ipb.size, PTR_ERR(mr));
2502                         kiblnd_hdev_cleanup_mrs(hdev);
2503                         return PTR_ERR(mr);
2504                 }
2505
2506                 LASSERT (iova == ipb.addr);
2507
2508                 hdev->ibh_mrs[i] = mr;
2509         }
2510
2511 out:
2512         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2513                 LCONSOLE_INFO("Register global MR array, MR size: "
2514                               LPX64", array size: %d\n",
2515                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2516         return 0;
2517 }
2518
2519 static int
2520 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2521 {       /* DUMMY */
2522         return 0;
2523 }
2524
2525 static int
2526 kiblnd_dev_need_failover(kib_dev_t *dev)
2527 {
2528         struct rdma_cm_id  *cmid;
2529         struct sockaddr_in  srcaddr;
2530         struct sockaddr_in  dstaddr;
2531         int                 rc;
2532
2533         if (dev->ibd_hdev == NULL || /* initializing */
2534             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2535             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2536                 return 1;
2537
2538         /* XXX: it's UGLY, but I don't have better way to find
2539          * ib-bonding HCA failover because:
2540          *
2541          * a. no reliable CM event for HCA failover...
2542          * b. no OFED API to get ib_device for current net_device...
2543          *
2544          * We have only two choices at this point:
2545          *
2546          * a. rdma_bind_addr(), it will conflict with listener cmid
2547          * b. rdma_resolve_addr() to zero addr */
2548         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2549                                      IB_QPT_RC);
2550         if (IS_ERR(cmid)) {
2551                 rc = PTR_ERR(cmid);
2552                 CERROR("Failed to create cmid for failover: %d\n", rc);
2553                 return rc;
2554         }
2555
2556         memset(&srcaddr, 0, sizeof(srcaddr));
2557         srcaddr.sin_family      = AF_INET;
2558         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2559
2560         memset(&dstaddr, 0, sizeof(dstaddr));
2561         dstaddr.sin_family = AF_INET;
2562         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2563                                (struct sockaddr *)&dstaddr, 1);
2564         if (rc != 0 || cmid->device == NULL) {
2565                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2566                        dev->ibd_ifname, &dev->ibd_ifip,
2567                        cmid->device, rc);
2568                 rdma_destroy_id(cmid);
2569                 return rc;
2570         }
2571
2572         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2573         rdma_destroy_id(cmid);
2574         return rc;
2575 }
2576
2577 int
2578 kiblnd_dev_failover(kib_dev_t *dev)
2579 {
2580         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2581         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2582         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2583         struct rdma_cm_id  *cmid  = NULL;
2584         kib_hca_dev_t      *hdev  = NULL;
2585         kib_hca_dev_t      *old;
2586         struct ib_pd       *pd;
2587         kib_net_t          *net;
2588         struct sockaddr_in  addr;
2589         unsigned long       flags;
2590         int                 rc = 0;
2591         int                 i;
2592
2593         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2594                  dev->ibd_can_failover ||
2595                  dev->ibd_hdev == NULL);
2596
2597         rc = kiblnd_dev_need_failover(dev);
2598         if (rc <= 0)
2599                 goto out;
2600
2601         if (dev->ibd_hdev != NULL &&
2602             dev->ibd_hdev->ibh_cmid != NULL) {
2603                 /* XXX it's not good to close old listener at here,
2604                  * because we can fail to create new listener.
2605                  * But we have to close it now, otherwise rdma_bind_addr
2606                  * will return EADDRINUSE... How crap! */
2607                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2608
2609                 cmid = dev->ibd_hdev->ibh_cmid;
2610                 /* make next schedule of kiblnd_dev_need_failover()
2611                  * return 1 for me */
2612                 dev->ibd_hdev->ibh_cmid  = NULL;
2613                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2614
2615                 rdma_destroy_id(cmid);
2616         }
2617
2618         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2619                                      IB_QPT_RC);
2620         if (IS_ERR(cmid)) {
2621                 rc = PTR_ERR(cmid);
2622                 CERROR("Failed to create cmid for failover: %d\n", rc);
2623                 goto out;
2624         }
2625
2626         memset(&addr, 0, sizeof(addr));
2627         addr.sin_family      = AF_INET;
2628         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2629         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2630
2631         /* Bind to failover device or port */
2632         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2633         if (rc != 0 || cmid->device == NULL) {
2634                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2635                        dev->ibd_ifname, &dev->ibd_ifip,
2636                        cmid->device, rc);
2637                 rdma_destroy_id(cmid);
2638                 goto out;
2639         }
2640
2641         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2642         if (hdev == NULL) {
2643                 CERROR("Failed to allocate kib_hca_dev\n");
2644                 rdma_destroy_id(cmid);
2645                 rc = -ENOMEM;
2646                 goto out;
2647         }
2648
2649         atomic_set(&hdev->ibh_ref, 1);
2650         hdev->ibh_dev   = dev;
2651         hdev->ibh_cmid  = cmid;
2652         hdev->ibh_ibdev = cmid->device;
2653
2654         pd = ib_alloc_pd(cmid->device);
2655         if (IS_ERR(pd)) {
2656                 rc = PTR_ERR(pd);
2657                 CERROR("Can't allocate PD: %d\n", rc);
2658                 goto out;
2659         }
2660
2661         hdev->ibh_pd = pd;
2662
2663         rc = rdma_listen(cmid, 0);
2664         if (rc != 0) {
2665                 CERROR("Can't start new listener: %d\n", rc);
2666                 goto out;
2667         }
2668
2669         rc = kiblnd_hdev_setup_mrs(hdev);
2670         if (rc != 0) {
2671                 CERROR("Can't setup device: %d\n", rc);
2672                 goto out;
2673         }
2674
2675         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2676
2677         old = dev->ibd_hdev;
2678         dev->ibd_hdev = hdev;   /* take over the refcount */
2679         hdev = old;
2680
2681         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2682                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2683                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2684                                             &zombie_tpo);
2685
2686                         if (net->ibn_fmr_ps != NULL) {
2687                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2688                                                         &zombie_fpo);
2689
2690                         } else if (net->ibn_pmr_ps != NULL) {
2691                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2692                                                     pps_poolset, &zombie_ppo);
2693                         }
2694                 }
2695         }
2696
2697         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2698  out:
2699         if (!list_empty(&zombie_tpo))
2700                 kiblnd_destroy_pool_list(&zombie_tpo);
2701         if (!list_empty(&zombie_ppo))
2702                 kiblnd_destroy_pool_list(&zombie_ppo);
2703         if (!list_empty(&zombie_fpo))
2704                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2705         if (hdev != NULL)
2706                 kiblnd_hdev_decref(hdev);
2707
2708         if (rc != 0)
2709                 dev->ibd_failed_failover++;
2710         else
2711                 dev->ibd_failed_failover = 0;
2712
2713         return rc;
2714 }
2715
2716 void
2717 kiblnd_destroy_dev (kib_dev_t *dev)
2718 {
2719         LASSERT (dev->ibd_nnets == 0);
2720         LASSERT(list_empty(&dev->ibd_nets));
2721
2722         list_del(&dev->ibd_fail_list);
2723         list_del(&dev->ibd_list);
2724
2725         if (dev->ibd_hdev != NULL)
2726                 kiblnd_hdev_decref(dev->ibd_hdev);
2727
2728         LIBCFS_FREE(dev, sizeof(*dev));
2729 }
2730
2731 static kib_dev_t *
2732 kiblnd_create_dev(char *ifname)
2733 {
2734         struct net_device *netdev;
2735         kib_dev_t         *dev;
2736         __u32              netmask;
2737         __u32              ip;
2738         int                up;
2739         int                rc;
2740
2741         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2742         if (rc != 0) {
2743                 CERROR("Can't query IPoIB interface %s: %d\n",
2744                        ifname, rc);
2745                 return NULL;
2746         }
2747
2748         if (!up) {
2749                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2750                 return NULL;
2751         }
2752
2753         LIBCFS_ALLOC(dev, sizeof(*dev));
2754         if (dev == NULL)
2755                 return NULL;
2756
2757         memset(dev, 0, sizeof(*dev));
2758         netdev = dev_get_by_name(&init_net, ifname);
2759         if (netdev == NULL) {
2760                 dev->ibd_can_failover = 0;
2761         } else {
2762                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2763                 dev_put(netdev);
2764         }
2765
2766         INIT_LIST_HEAD(&dev->ibd_nets);
2767         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2768         INIT_LIST_HEAD(&dev->ibd_fail_list);
2769         dev->ibd_ifip = ip;
2770         strcpy(&dev->ibd_ifname[0], ifname);
2771
2772         /* initialize the device */
2773         rc = kiblnd_dev_failover(dev);
2774         if (rc != 0) {
2775                 CERROR("Can't initialize device: %d\n", rc);
2776                 LIBCFS_FREE(dev, sizeof(*dev));
2777                 return NULL;
2778         }
2779
2780         list_add_tail(&dev->ibd_list,
2781                           &kiblnd_data.kib_devs);
2782         return dev;
2783 }
2784
2785 static void
2786 kiblnd_base_shutdown(void)
2787 {
2788         struct kib_sched_info   *sched;
2789         int                     i;
2790
2791         LASSERT(list_empty(&kiblnd_data.kib_devs));
2792
2793         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2794                atomic_read(&libcfs_kmemory));
2795
2796         switch (kiblnd_data.kib_init) {
2797         default:
2798                 LBUG();
2799
2800         case IBLND_INIT_ALL:
2801         case IBLND_INIT_DATA:
2802                 LASSERT (kiblnd_data.kib_peers != NULL);
2803                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2804                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2805                 }
2806                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2807                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2808
2809                 /* flag threads to terminate; wake and wait for them to die */
2810                 kiblnd_data.kib_shutdown = 1;
2811
2812                 /* NB: we really want to stop scheduler threads net by net
2813                  * instead of the whole module, this should be improved
2814                  * with dynamic configuration LNet */
2815                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2816                         wake_up_all(&sched->ibs_waitq);
2817
2818                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2819                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2820
2821                 i = 2;
2822                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2823                         i++;
2824                         /* power of 2? */
2825                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2826                                "Waiting for %d threads to terminate\n",
2827                                atomic_read(&kiblnd_data.kib_nthreads));
2828                         set_current_state(TASK_UNINTERRUPTIBLE);
2829                         schedule_timeout(cfs_time_seconds(1));
2830                 }
2831
2832                 /* fall through */
2833
2834         case IBLND_INIT_NOTHING:
2835                 break;
2836         }
2837
2838         if (kiblnd_data.kib_peers != NULL) {
2839                 LIBCFS_FREE(kiblnd_data.kib_peers,
2840                             sizeof(struct list_head) *
2841                             kiblnd_data.kib_peer_hash_size);
2842         }
2843
2844         if (kiblnd_data.kib_scheds != NULL)
2845                 cfs_percpt_free(kiblnd_data.kib_scheds);
2846
2847         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2848                atomic_read(&libcfs_kmemory));
2849
2850         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2851         module_put(THIS_MODULE);
2852 }
2853
2854 static void
2855 kiblnd_shutdown (lnet_ni_t *ni)
2856 {
2857         kib_net_t        *net = ni->ni_data;
2858         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2859         int               i;
2860         unsigned long     flags;
2861
2862         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2863
2864         if (net == NULL)
2865                 goto out;
2866
2867         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2868                atomic_read(&libcfs_kmemory));
2869
2870         write_lock_irqsave(g_lock, flags);
2871         net->ibn_shutdown = 1;
2872         write_unlock_irqrestore(g_lock, flags);
2873
2874         switch (net->ibn_init) {
2875         default:
2876                 LBUG();
2877
2878         case IBLND_INIT_ALL:
2879                 /* nuke all existing peers within this net */
2880                 kiblnd_del_peer(ni, LNET_NID_ANY);
2881
2882                 /* Wait for all peer state to clean up */
2883                 i = 2;
2884                 while (atomic_read(&net->ibn_npeers) != 0) {
2885                         i++;
2886                         /* power of 2? */
2887                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2888                                "%s: waiting for %d peers to disconnect\n",
2889                                libcfs_nid2str(ni->ni_nid),
2890                                atomic_read(&net->ibn_npeers));
2891                         set_current_state(TASK_UNINTERRUPTIBLE);
2892                         schedule_timeout(cfs_time_seconds(1));
2893                 }
2894
2895                 kiblnd_net_fini_pools(net);
2896
2897                 write_lock_irqsave(g_lock, flags);
2898                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2899                 net->ibn_dev->ibd_nnets--;
2900                 list_del(&net->ibn_list);
2901                 write_unlock_irqrestore(g_lock, flags);
2902
2903                 /* fall through */
2904
2905         case IBLND_INIT_NOTHING:
2906                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2907
2908                 if (net->ibn_dev != NULL &&
2909                     net->ibn_dev->ibd_nnets == 0)
2910                         kiblnd_destroy_dev(net->ibn_dev);
2911
2912                 break;
2913         }
2914
2915         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2916                atomic_read(&libcfs_kmemory));
2917
2918         net->ibn_init = IBLND_INIT_NOTHING;
2919         ni->ni_data = NULL;
2920
2921         LIBCFS_FREE(net, sizeof(*net));
2922
2923 out:
2924         if (list_empty(&kiblnd_data.kib_devs))
2925                 kiblnd_base_shutdown();
2926         return;
2927 }
2928
2929 static int
2930 kiblnd_base_startup(void)
2931 {
2932         struct kib_sched_info   *sched;
2933         int                     rc;
2934         int                     i;
2935
2936         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2937
2938         try_module_get(THIS_MODULE);
2939         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2940
2941         rwlock_init(&kiblnd_data.kib_global_lock);
2942
2943         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2944         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2945
2946         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2947         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2948                      sizeof(struct list_head) *
2949                      kiblnd_data.kib_peer_hash_size);
2950         if (kiblnd_data.kib_peers == NULL)
2951                 goto failed;
2952
2953         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2954                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2955
2956         spin_lock_init(&kiblnd_data.kib_connd_lock);
2957         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2958         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2959         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2960         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2961
2962         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2963                                                   sizeof(*sched));
2964         if (kiblnd_data.kib_scheds == NULL)
2965                 goto failed;
2966
2967         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2968                 int     nthrs;
2969
2970                 spin_lock_init(&sched->ibs_lock);
2971                 INIT_LIST_HEAD(&sched->ibs_conns);
2972                 init_waitqueue_head(&sched->ibs_waitq);
2973
2974                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2975                 if (*kiblnd_tunables.kib_nscheds > 0) {
2976                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2977                 } else {
2978                         /* max to half of CPUs, another half is reserved for
2979                          * upper layer modules */
2980                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2981                 }
2982
2983                 sched->ibs_nthreads_max = nthrs;
2984                 sched->ibs_cpt = i;
2985         }
2986
2987         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2988
2989         /* lists/ptrs/locks initialised */
2990         kiblnd_data.kib_init = IBLND_INIT_DATA;
2991         /*****************************************************/
2992
2993         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2994         if (rc != 0) {
2995                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2996                 goto failed;
2997         }
2998
2999         if (*kiblnd_tunables.kib_dev_failover != 0)
3000                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3001                                          "kiblnd_failover");
3002
3003         if (rc != 0) {
3004                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3005                 goto failed;
3006         }
3007
3008         /* flag everything initialised */
3009         kiblnd_data.kib_init = IBLND_INIT_ALL;
3010         /*****************************************************/
3011
3012         return 0;
3013
3014  failed:
3015         kiblnd_base_shutdown();
3016         return -ENETDOWN;
3017 }
3018
3019 static int
3020 kiblnd_start_schedulers(struct kib_sched_info *sched)
3021 {
3022         int     rc = 0;
3023         int     nthrs;
3024         int     i;
3025
3026         if (sched->ibs_nthreads == 0) {
3027                 if (*kiblnd_tunables.kib_nscheds > 0) {
3028                         nthrs = sched->ibs_nthreads_max;
3029                 } else {
3030                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3031                                                sched->ibs_cpt);
3032                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3033                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3034                 }
3035         } else {
3036                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3037                 /* increase one thread if there is new interface */
3038                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3039         }
3040
3041         for (i = 0; i < nthrs; i++) {
3042                 long    id;
3043                 char    name[20];
3044                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3045                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3046                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3047                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3048                 if (rc == 0)
3049                         continue;
3050
3051                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3052                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3053                 break;
3054         }
3055
3056         sched->ibs_nthreads += i;
3057         return rc;
3058 }
3059
3060 static int
3061 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3062 {
3063         int     cpt;
3064         int     rc;
3065         int     i;
3066
3067         for (i = 0; i < ncpts; i++) {
3068                 struct kib_sched_info *sched;
3069
3070                 cpt = (cpts == NULL) ? i : cpts[i];
3071                 sched = kiblnd_data.kib_scheds[cpt];
3072
3073                 if (!newdev && sched->ibs_nthreads > 0)
3074                         continue;
3075
3076                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3077                 if (rc != 0) {
3078                         CERROR("Failed to start scheduler threads for %s\n",
3079                                dev->ibd_ifname);
3080                         return rc;
3081                 }
3082         }
3083         return 0;
3084 }
3085
3086 static kib_dev_t *
3087 kiblnd_dev_search(char *ifname)
3088 {
3089         kib_dev_t       *alias = NULL;
3090         kib_dev_t       *dev;
3091         char            *colon;
3092         char            *colon2;
3093
3094         colon = strchr(ifname, ':');
3095         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3096                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3097                         return dev;
3098
3099                 if (alias != NULL)
3100                         continue;
3101
3102                 colon2 = strchr(dev->ibd_ifname, ':');
3103                 if (colon != NULL)
3104                         *colon = 0;
3105                 if (colon2 != NULL)
3106                         *colon2 = 0;
3107
3108                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3109                         alias = dev;
3110
3111                 if (colon != NULL)
3112                         *colon = ':';
3113                 if (colon2 != NULL)
3114                         *colon2 = ':';
3115         }
3116         return alias;
3117 }
3118
3119 static int
3120 kiblnd_startup (lnet_ni_t *ni)
3121 {
3122         char                     *ifname;
3123         kib_dev_t                *ibdev = NULL;
3124         kib_net_t                *net;
3125         struct timeval            tv;
3126         unsigned long             flags;
3127         int                       rc;
3128         int                       newdev;
3129
3130         LASSERT (ni->ni_lnd == &the_o2iblnd);
3131
3132         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3133                 rc = kiblnd_base_startup();
3134                 if (rc != 0)
3135                         return rc;
3136         }
3137
3138         LIBCFS_ALLOC(net, sizeof(*net));
3139         ni->ni_data = net;
3140         if (net == NULL)
3141                 goto failed;
3142
3143         memset(net, 0, sizeof(*net));
3144
3145         do_gettimeofday(&tv);
3146         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3147
3148         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3149         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3150         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3151         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3152
3153         if (ni->ni_interfaces[0] != NULL) {
3154                 /* Use the IPoIB interface specified in 'networks=' */
3155
3156                 CLASSERT (LNET_MAX_INTERFACES > 1);
3157                 if (ni->ni_interfaces[1] != NULL) {
3158                         CERROR("Multiple interfaces not supported\n");
3159                         goto failed;
3160                 }
3161
3162                 ifname = ni->ni_interfaces[0];
3163         } else {
3164                 ifname = *kiblnd_tunables.kib_default_ipif;
3165         }
3166
3167         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3168                 CERROR("IPoIB interface name too long: %s\n", ifname);
3169                 goto failed;
3170         }
3171
3172         ibdev = kiblnd_dev_search(ifname);
3173
3174         newdev = ibdev == NULL;
3175         /* hmm...create kib_dev even for alias */
3176         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3177                 ibdev = kiblnd_create_dev(ifname);
3178
3179         if (ibdev == NULL)
3180                 goto failed;
3181
3182         net->ibn_dev = ibdev;
3183         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3184
3185         rc = kiblnd_dev_start_threads(ibdev, newdev,
3186                                       ni->ni_cpts, ni->ni_ncpts);
3187         if (rc != 0)
3188                 goto failed;
3189
3190         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3191         if (rc != 0) {
3192                 CERROR("Failed to initialize NI pools: %d\n", rc);
3193                 goto failed;
3194         }
3195
3196         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3197         ibdev->ibd_nnets++;
3198         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3199         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3200
3201         net->ibn_init = IBLND_INIT_ALL;
3202
3203         return 0;
3204
3205 failed:
3206         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3207                 kiblnd_destroy_dev(ibdev);
3208
3209         kiblnd_shutdown(ni);
3210
3211         CDEBUG(D_NET, "kiblnd_startup failed\n");
3212         return -ENETDOWN;
3213 }
3214
3215 static lnd_t the_o2iblnd = {
3216         .lnd_type       = O2IBLND,
3217         .lnd_startup    = kiblnd_startup,
3218         .lnd_shutdown   = kiblnd_shutdown,
3219         .lnd_ctl        = kiblnd_ctl,
3220         .lnd_query      = kiblnd_query,
3221         .lnd_send       = kiblnd_send,
3222         .lnd_recv       = kiblnd_recv,
3223 };
3224
3225 static void __exit
3226 kiblnd_module_fini (void)
3227 {
3228         lnet_unregister_lnd(&the_o2iblnd);
3229         kiblnd_tunables_fini();
3230 }
3231
3232 static int __init
3233 kiblnd_module_init (void)
3234 {
3235         int    rc;
3236
3237         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3238         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3239                   <= IBLND_MSG_SIZE);
3240         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3241                   <= IBLND_MSG_SIZE);
3242
3243         rc = kiblnd_tunables_init();
3244         if (rc != 0)
3245                 return rc;
3246
3247         lnet_register_lnd(&the_o2iblnd);
3248
3249         return 0;
3250 }
3251
3252 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3253 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3254 MODULE_LICENSE("GPL");
3255
3256 module_init(kiblnd_module_init);
3257 module_exit(kiblnd_module_fini);