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 "o2iblnd.h"
  42
  43 static lnd_t the_o2iblnd;
  44
  45 kib_data_t              kiblnd_data;
  46
  47 static __u32
  48 kiblnd_cksum (void *ptr, int nob)
  49 {
  50         char  *c  = ptr;
  51         __u32  sum = 0;
  52
  53         while (nob-- > 0)
  54                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  55
  56         /* ensure I don't return 0 (== no checksum) */
  57         return (sum == 0) ? 1 : sum;
  58 }
  59
  60 static char *
  61 kiblnd_msgtype2str(int type)
  62 {
  63         switch (type) {
  64         case IBLND_MSG_CONNREQ:
  65                 return "CONNREQ";
  66
  67         case IBLND_MSG_CONNACK:
  68                 return "CONNACK";
  69
  70         case IBLND_MSG_NOOP:
  71                 return "NOOP";
  72
  73         case IBLND_MSG_IMMEDIATE:
  74                 return "IMMEDIATE";
  75
  76         case IBLND_MSG_PUT_REQ:
  77                 return "PUT_REQ";
  78
  79         case IBLND_MSG_PUT_NAK:
  80                 return "PUT_NAK";
  81
  82         case IBLND_MSG_PUT_ACK:
  83                 return "PUT_ACK";
  84
  85         case IBLND_MSG_PUT_DONE:
  86                 return "PUT_DONE";
  87
  88         case IBLND_MSG_GET_REQ:
  89                 return "GET_REQ";
  90
  91         case IBLND_MSG_GET_DONE:
  92                 return "GET_DONE";
  93
  94         default:
  95                 return "???";
  96         }
  97 }
  98
  99 static int
 100 kiblnd_msgtype2size(int type)
 101 {
 102         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 103
 104         switch (type) {
 105         case IBLND_MSG_CONNREQ:
 106         case IBLND_MSG_CONNACK:
 107                 return hdr_size + sizeof(kib_connparams_t);
 108
 109         case IBLND_MSG_NOOP:
 110                 return hdr_size;
 111
 112         case IBLND_MSG_IMMEDIATE:
 113                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 114
 115         case IBLND_MSG_PUT_REQ:
 116                 return hdr_size + sizeof(kib_putreq_msg_t);
 117
 118         case IBLND_MSG_PUT_ACK:
 119                 return hdr_size + sizeof(kib_putack_msg_t);
 120
 121         case IBLND_MSG_GET_REQ:
 122                 return hdr_size + sizeof(kib_get_msg_t);
 123
 124         case IBLND_MSG_PUT_NAK:
 125         case IBLND_MSG_PUT_DONE:
 126         case IBLND_MSG_GET_DONE:
 127                 return hdr_size + sizeof(kib_completion_msg_t);
 128         default:
 129                 return -1;
 130         }
 131 }
 132
 133 static int
 134 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 135 {
 136         kib_rdma_desc_t   *rd;
 137         int                nob;
 138         int                n;
 139         int                i;
 140
 141         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 142                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 143
 144         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 145                               &msg->ibm_u.get.ibgm_rd :
 146                               &msg->ibm_u.putack.ibpam_rd;
 147
 148         if (flip) {
 149                 __swab32s(&rd->rd_key);
 150                 __swab32s(&rd->rd_nfrags);
 151         }
 152
 153         n = rd->rd_nfrags;
 154
 155         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 156                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 157                        n, IBLND_MAX_RDMA_FRAGS);
 158                 return 1;
 159         }
 160
 161         nob = offsetof (kib_msg_t, ibm_u) +
 162               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 163
 164         if (msg->ibm_nob < nob) {
 165                 CERROR("Short %s: %d(%d)\n",
 166                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 167                 return 1;
 168         }
 169
 170         if (!flip)
 171                 return 0;
 172
 173         for (i = 0; i < n; i++) {
 174                 __swab32s(&rd->rd_frags[i].rf_nob);
 175                 __swab64s(&rd->rd_frags[i].rf_addr);
 176         }
 177
 178         return 0;
 179 }
 180
 181 void
 182 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 183                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 184 {
 185         kib_net_t *net = ni->ni_data;
 186
 187         /* CAVEAT EMPTOR! all message fields not set here should have been
 188          * initialised previously. */
 189         msg->ibm_magic    = IBLND_MSG_MAGIC;
 190         msg->ibm_version  = version;
 191         /*   ibm_type */
 192         msg->ibm_credits  = credits;
 193         /*   ibm_nob */
 194         msg->ibm_cksum    = 0;
 195         msg->ibm_srcnid   = ni->ni_nid;
 196         msg->ibm_srcstamp = net->ibn_incarnation;
 197         msg->ibm_dstnid   = dstnid;
 198         msg->ibm_dststamp = dststamp;
 199
 200         if (*kiblnd_tunables.kib_cksum) {
 201                 /* NB ibm_cksum zero while computing cksum */
 202                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 203         }
 204 }
 205
 206 int
 207 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 208 {
 209         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 210         __u32     msg_cksum;
 211         __u16     version;
 212         int       msg_nob;
 213         int       flip;
 214
 215         /* 6 bytes are enough to have received magic + version */
 216         if (nob < 6) {
 217                 CERROR("Short message: %d\n", nob);
 218                 return -EPROTO;
 219         }
 220
 221         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 222                 flip = 0;
 223         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 224                 flip = 1;
 225         } else {
 226                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 227                 return -EPROTO;
 228         }
 229
 230         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 231         if (version != IBLND_MSG_VERSION &&
 232             version != IBLND_MSG_VERSION_1) {
 233                 CERROR("Bad version: %x\n", version);
 234                 return -EPROTO;
 235         }
 236
 237         if (nob < hdr_size) {
 238                 CERROR("Short message: %d\n", nob);
 239                 return -EPROTO;
 240         }
 241
 242         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 243         if (msg_nob > nob) {
 244                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 245                 return -EPROTO;
 246         }
 247
 248         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 249          * gets flipped */
 250         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 251         msg->ibm_cksum = 0;
 252         if (msg_cksum != 0 &&
 253             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 254                 CERROR("Bad checksum\n");
 255                 return -EPROTO;
 256         }
 257
 258         msg->ibm_cksum = msg_cksum;
 259
 260         if (flip) {
 261                 /* leave magic unflipped as a clue to peer endianness */
 262                 msg->ibm_version = version;
 263                 CLASSERT (sizeof(msg->ibm_type) == 1);
 264                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 265                 msg->ibm_nob     = msg_nob;
 266                 __swab64s(&msg->ibm_srcnid);
 267                 __swab64s(&msg->ibm_srcstamp);
 268                 __swab64s(&msg->ibm_dstnid);
 269                 __swab64s(&msg->ibm_dststamp);
 270         }
 271
 272         if (msg->ibm_srcnid == LNET_NID_ANY) {
 273                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 274                 return -EPROTO;
 275         }
 276
 277         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 278                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 279                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 280                 return -EPROTO;
 281         }
 282
 283         switch (msg->ibm_type) {
 284         default:
 285                 CERROR("Unknown message type %x\n", msg->ibm_type);
 286                 return -EPROTO;
 287
 288         case IBLND_MSG_NOOP:
 289         case IBLND_MSG_IMMEDIATE:
 290         case IBLND_MSG_PUT_REQ:
 291                 break;
 292
 293         case IBLND_MSG_PUT_ACK:
 294         case IBLND_MSG_GET_REQ:
 295                 if (kiblnd_unpack_rd(msg, flip))
 296                         return -EPROTO;
 297                 break;
 298
 299         case IBLND_MSG_PUT_NAK:
 300         case IBLND_MSG_PUT_DONE:
 301         case IBLND_MSG_GET_DONE:
 302                 if (flip)
 303                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 304                 break;
 305
 306         case IBLND_MSG_CONNREQ:
 307         case IBLND_MSG_CONNACK:
 308                 if (flip) {
 309                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 310                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 311                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 312                 }
 313                 break;
 314         }
 315         return 0;
 316 }
 317
 318 int
 319 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 320 {
 321         kib_peer_t      *peer;
 322         kib_net_t       *net = ni->ni_data;
 323         int             cpt = lnet_cpt_of_nid(nid);
 324         unsigned long   flags;
 325
 326         LASSERT(net != NULL);
 327         LASSERT(nid != LNET_NID_ANY);
 328
 329         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 330         if (peer == NULL) {
 331                 CERROR("Cannot allocate peer\n");
 332                 return -ENOMEM;
 333         }
 334
 335         peer->ibp_ni = ni;
 336         peer->ibp_nid = nid;
 337         peer->ibp_error = 0;
 338         peer->ibp_last_alive = 0;
 339         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 340
 341         INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
 342         INIT_LIST_HEAD(&peer->ibp_conns);
 343         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 344
 345         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 346
 347         /* always called with a ref on ni, which prevents ni being shutdown */
 348         LASSERT(net->ibn_shutdown == 0);
 349
 350         /* npeers only grows with the global lock held */
 351         atomic_inc(&net->ibn_npeers);
 352
 353         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 354
 355         *peerp = peer;
 356         return 0;
 357 }
 358
 359 void
 360 kiblnd_destroy_peer (kib_peer_t *peer)
 361 {
 362         kib_net_t *net = peer->ibp_ni->ni_data;
 363
 364         LASSERT(net != NULL);
 365         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 366         LASSERT(!kiblnd_peer_active(peer));
 367         LASSERT(peer->ibp_connecting == 0);
 368         LASSERT(peer->ibp_accepting == 0);
 369         LASSERT(list_empty(&peer->ibp_conns));
 370         LASSERT(list_empty(&peer->ibp_tx_queue));
 371
 372         LIBCFS_FREE(peer, sizeof(*peer));
 373
 374         /* NB a peer's connections keep a reference on their peer until
 375          * they are destroyed, so we can be assured that _all_ state to do
 376          * with this peer has been cleaned up when its refcount drops to
 377          * zero. */
 378         atomic_dec(&net->ibn_npeers);
 379 }
 380
 381 kib_peer_t *
 382 kiblnd_find_peer_locked (lnet_nid_t nid)
 383 {
 384         /* the caller is responsible for accounting the additional reference
 385          * that this creates */
 386         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 387         struct list_head        *tmp;
 388         kib_peer_t              *peer;
 389
 390         list_for_each(tmp, peer_list) {
 391
 392                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 393
 394                 LASSERT(peer->ibp_connecting > 0 ||     /* creating conns */
 395                         peer->ibp_accepting > 0 ||
 396                         !list_empty(&peer->ibp_conns)); /* active conn */
 397
 398                 if (peer->ibp_nid != nid)
 399                         continue;
 400
 401                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 402                        peer, libcfs_nid2str(nid),
 403                        atomic_read(&peer->ibp_refcount),
 404                        peer->ibp_version);
 405                 return peer;
 406         }
 407         return NULL;
 408 }
 409
 410 void
 411 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 412 {
 413         LASSERT(list_empty(&peer->ibp_conns));
 414
 415         LASSERT (kiblnd_peer_active(peer));
 416         list_del_init(&peer->ibp_list);
 417         /* lose peerlist's ref */
 418         kiblnd_peer_decref(peer);
 419 }
 420
 421 static int
 422 kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 423                      lnet_nid_t *nidp, int *count)
 424 {
 425         kib_peer_t              *peer;
 426         struct list_head        *ptmp;
 427         int                      i;
 428         unsigned long            flags;
 429
 430         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 431
 432         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 433
 434                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 435
 436                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 437                         LASSERT(peer->ibp_connecting > 0 ||
 438                                 peer->ibp_accepting > 0 ||
 439                                 !list_empty(&peer->ibp_conns));
 440
 441                         if (peer->ibp_ni != ni)
 442                                 continue;
 443
 444                         if (index-- > 0)
 445                                 continue;
 446
 447                         *nidp = peer->ibp_nid;
 448                         *count = atomic_read(&peer->ibp_refcount);
 449
 450                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 451                                                flags);
 452                         return 0;
 453                 }
 454         }
 455
 456         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 457         return -ENOENT;
 458 }
 459
 460 static void
 461 kiblnd_del_peer_locked (kib_peer_t *peer)
 462 {
 463         struct list_head        *ctmp;
 464         struct list_head        *cnxt;
 465         kib_conn_t              *conn;
 466
 467         if (list_empty(&peer->ibp_conns)) {
 468                 kiblnd_unlink_peer_locked(peer);
 469         } else {
 470                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 471                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 472
 473                         kiblnd_close_conn_locked(conn, 0);
 474                 }
 475                 /* NB closing peer's last conn unlinked it. */
 476         }
 477         /* NB peer now unlinked; might even be freed if the peer table had the
 478          * last ref on it. */
 479 }
 480
 481 static int
 482 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 483 {
 484         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 485         struct list_head        *ptmp;
 486         struct list_head        *pnxt;
 487         kib_peer_t              *peer;
 488         int                     lo;
 489         int                     hi;
 490         int                     i;
 491         unsigned long           flags;
 492         int                     rc = -ENOENT;
 493
 494         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 495
 496         if (nid != LNET_NID_ANY) {
 497                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 498         } else {
 499                 lo = 0;
 500                 hi = kiblnd_data.kib_peer_hash_size - 1;
 501         }
 502
 503         for (i = lo; i <= hi; i++) {
 504                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 505                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 506                         LASSERT(peer->ibp_connecting > 0 ||
 507                                 peer->ibp_accepting > 0 ||
 508                                 !list_empty(&peer->ibp_conns));
 509
 510                         if (peer->ibp_ni != ni)
 511                                 continue;
 512
 513                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 514                                 continue;
 515
 516                         if (!list_empty(&peer->ibp_tx_queue)) {
 517                                 LASSERT(list_empty(&peer->ibp_conns));
 518
 519                                 list_splice_init(&peer->ibp_tx_queue,
 520                                                  &zombies);
 521                         }
 522
 523                         kiblnd_del_peer_locked(peer);
 524                         rc = 0;         /* matched something */
 525                 }
 526         }
 527
 528         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 529
 530         kiblnd_txlist_done(ni, &zombies, -EIO);
 531
 532         return rc;
 533 }
 534
 535 static kib_conn_t *
 536 kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 537 {
 538         kib_peer_t              *peer;
 539         struct list_head        *ptmp;
 540         kib_conn_t              *conn;
 541         struct list_head        *ctmp;
 542         int                     i;
 543         unsigned long           flags;
 544
 545         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 546
 547         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 548                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 549
 550                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 551                         LASSERT(peer->ibp_connecting > 0 ||
 552                                 peer->ibp_accepting > 0 ||
 553                                 !list_empty(&peer->ibp_conns));
 554
 555                         if (peer->ibp_ni != ni)
 556                                 continue;
 557
 558                         list_for_each(ctmp, &peer->ibp_conns) {
 559                                 if (index-- > 0)
 560                                         continue;
 561
 562                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 563                                 kiblnd_conn_addref(conn);
 564                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 565                                                        flags);
 566                                 return conn;
 567                         }
 568                 }
 569         }
 570
 571         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 572         return NULL;
 573 }
 574
 575 static void
 576 kiblnd_debug_rx (kib_rx_t *rx)
 577 {
 578         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 579                rx, rx->rx_status, rx->rx_msg->ibm_type,
 580                rx->rx_msg->ibm_credits);
 581 }
 582
 583 static void
 584 kiblnd_debug_tx (kib_tx_t *tx)
 585 {
 586         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 587                "cookie "LPX64" msg %s%s type %x cred %d\n",
 588                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 589                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 590                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 591                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 592                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 593 }
 594
 595 void
 596 kiblnd_debug_conn (kib_conn_t *conn)
 597 {
 598         struct list_head        *tmp;
 599         int                     i;
 600
 601         spin_lock(&conn->ibc_lock);
 602
 603         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 604                atomic_read(&conn->ibc_refcount), conn,
 605                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 606         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 607                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 608                conn->ibc_nsends_posted, conn->ibc_credits,
 609                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 610         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 611
 612         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 613         list_for_each(tmp, &conn->ibc_early_rxs)
 614                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 615
 616         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 617         list_for_each(tmp, &conn->ibc_tx_noops)
 618                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 621         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 622                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 623
 624         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 625         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 626                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 629         list_for_each(tmp, &conn->ibc_tx_queue)
 630                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 631
 632         CDEBUG(D_CONSOLE, "   active_txs:\n");
 633         list_for_each(tmp, &conn->ibc_active_txs)
 634                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 635
 636         CDEBUG(D_CONSOLE, "   rxs:\n");
 637         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 638                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 639
 640         spin_unlock(&conn->ibc_lock);
 641 }
 642
 643 int
 644 kiblnd_translate_mtu(int value)
 645 {
 646         switch (value) {
 647         default:
 648                 return -1;
 649         case 0:
 650                 return 0;
 651         case 256:
 652                 return IB_MTU_256;
 653         case 512:
 654                 return IB_MTU_512;
 655         case 1024:
 656                 return IB_MTU_1024;
 657         case 2048:
 658                 return IB_MTU_2048;
 659         case 4096:
 660                 return IB_MTU_4096;
 661         }
 662 }
 663
 664 static void
 665 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 666 {
 667         int           mtu;
 668
 669         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 670         if (cmid->route.path_rec == NULL)
 671                 return;
 672
 673         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 674         LASSERT (mtu >= 0);
 675         if (mtu != 0)
 676                 cmid->route.path_rec->mtu = mtu;
 677 }
 678
 679 #ifdef HAVE_OFED_IB_COMP_VECTOR
 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 % cpus_weight(*mask);
 696         for_each_cpu_mask(i, *mask) {
 697                 if (off-- == 0)
 698                         return i % vectors;
 699         }
 700
 701         LBUG();
 702         return 1;
 703 }
 704 #endif
 705
 706 kib_conn_t *
 707 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 708                    int state, int version)
 709 {
 710         /* CAVEAT EMPTOR:
 711          * If the new conn is created successfully it takes over the caller's
 712          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 713          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 714          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 715          * to destroy 'cmid' here since I'm called from the CM which still has
 716          * its ref on 'cmid'). */
 717         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 718         kib_net_t              *net = peer->ibp_ni->ni_data;
 719         kib_dev_t              *dev;
 720         struct ib_qp_init_attr *init_qp_attr;
 721         struct kib_sched_info   *sched;
 722         kib_conn_t              *conn;
 723         struct ib_cq            *cq;
 724         unsigned long           flags;
 725         int                     cpt;
 726         int                     rc;
 727         int                     i;
 728
 729         LASSERT(net != NULL);
 730         LASSERT(!in_interrupt());
 731
 732         dev = net->ibn_dev;
 733
 734         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 735         sched = kiblnd_data.kib_scheds[cpt];
 736
 737         LASSERT(sched->ibs_nthreads > 0);
 738
 739         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 740                          sizeof(*init_qp_attr));
 741         if (init_qp_attr == NULL) {
 742                 CERROR("Can't allocate qp_attr for %s\n",
 743                        libcfs_nid2str(peer->ibp_nid));
 744                 goto failed_0;
 745         }
 746
 747         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 748         if (conn == NULL) {
 749                 CERROR("Can't allocate connection for %s\n",
 750                        libcfs_nid2str(peer->ibp_nid));
 751                 goto failed_1;
 752         }
 753
 754         conn->ibc_state = IBLND_CONN_INIT;
 755         conn->ibc_version = version;
 756         conn->ibc_peer = peer;                  /* I take the caller's ref */
 757         cmid->context = conn;                   /* for future CM callbacks */
 758         conn->ibc_cmid = cmid;
 759
 760         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 761         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 762         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 763         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 764         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 765         INIT_LIST_HEAD(&conn->ibc_active_txs);
 766         spin_lock_init(&conn->ibc_lock);
 767
 768         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 769                          sizeof(*conn->ibc_connvars));
 770         if (conn->ibc_connvars == NULL) {
 771                 CERROR("Can't allocate in-progress connection state\n");
 772                 goto failed_2;
 773         }
 774
 775         write_lock_irqsave(glock, flags);
 776         if (dev->ibd_failover) {
 777                 write_unlock_irqrestore(glock, flags);
 778                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 779                 goto failed_2;
 780         }
 781
 782         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 783                 /* wakeup failover thread and teardown connection */
 784                 if (kiblnd_dev_can_failover(dev)) {
 785                         list_add_tail(&dev->ibd_fail_list,
 786                                       &kiblnd_data.kib_failed_devs);
 787                         wake_up(&kiblnd_data.kib_failover_waitq);
 788                 }
 789
 790                 write_unlock_irqrestore(glock, flags);
 791                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 792                        cmid->device->name, dev->ibd_ifname);
 793                 goto failed_2;
 794         }
 795
 796         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 797         conn->ibc_hdev = dev->ibd_hdev;
 798
 799         kiblnd_setup_mtu_locked(cmid);
 800
 801         write_unlock_irqrestore(glock, flags);
 802
 803         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 804                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 805         if (conn->ibc_rxs == NULL) {
 806                 CERROR("Cannot allocate RX buffers\n");
 807                 goto failed_2;
 808         }
 809
 810         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 811                                 IBLND_RX_MSG_PAGES(version));
 812         if (rc != 0)
 813                 goto failed_2;
 814
 815         kiblnd_map_rx_descs(conn);
 816
 817 #ifdef HAVE_OFED_IB_COMP_VECTOR
 818         cq = ib_create_cq(cmid->device,
 819                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 820                           IBLND_CQ_ENTRIES(version),
 821                           kiblnd_get_completion_vector(conn, cpt));
 822 #else
 823         cq = ib_create_cq(cmid->device,
 824                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 825                           IBLND_CQ_ENTRIES(version));
 826 #endif
 827         if (IS_ERR(cq)) {
 828                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 829                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 830                 goto failed_2;
 831         }
 832
 833         conn->ibc_cq = cq;
 834
 835         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 836         if (rc != 0) {
 837                 CERROR("Can't request completion notificiation: %d\n", rc);
 838                 goto failed_2;
 839         }
 840
 841         init_qp_attr->event_handler = kiblnd_qp_event;
 842         init_qp_attr->qp_context = conn;
 843         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 844         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 845         init_qp_attr->cap.max_send_sge = 1;
 846         init_qp_attr->cap.max_recv_sge = 1;
 847         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 848         init_qp_attr->qp_type = IB_QPT_RC;
 849         init_qp_attr->send_cq = cq;
 850         init_qp_attr->recv_cq = cq;
 851
 852         conn->ibc_sched = sched;
 853
 854         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 855         if (rc != 0) {
 856                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 857                        rc, init_qp_attr->cap.max_send_wr,
 858                        init_qp_attr->cap.max_recv_wr);
 859                 goto failed_2;
 860         }
 861
 862         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 863
 864         /* 1 ref for caller and each rxmsg */
 865         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 866         conn->ibc_nrx = IBLND_RX_MSGS(version);
 867
 868         /* post receives */
 869         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 870                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 871                                     IBLND_POSTRX_NO_CREDIT);
 872                 if (rc != 0) {
 873                         CERROR("Can't post rxmsg: %d\n", rc);
 874
 875                         /* Make posted receives complete */
 876                         kiblnd_abort_receives(conn);
 877
 878                         /* correct # of posted buffers
 879                          * NB locking needed now I'm racing with completion */
 880                         spin_lock_irqsave(&sched->ibs_lock, flags);
 881                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 882                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 883
 884                         /* cmid will be destroyed by CM(ofed) after cm_callback
 885                          * returned, so we can't refer it anymore
 886                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 887                         rdma_destroy_qp(conn->ibc_cmid);
 888                         conn->ibc_cmid = NULL;
 889
 890                         /* Drop my own and unused rxbuffer refcounts */
 891                         while (i++ <= IBLND_RX_MSGS(version))
 892                                 kiblnd_conn_decref(conn);
 893
 894                         return NULL;
 895                 }
 896         }
 897
 898         /* Init successful! */
 899         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 900                  state == IBLND_CONN_PASSIVE_WAIT);
 901         conn->ibc_state = state;
 902
 903         /* 1 more conn */
 904         atomic_inc(&net->ibn_nconns);
 905         return conn;
 906
 907  failed_2:
 908         kiblnd_destroy_conn(conn);
 909  failed_1:
 910         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 911  failed_0:
 912         return NULL;
 913 }
 914
 915 void
 916 kiblnd_destroy_conn (kib_conn_t *conn)
 917 {
 918         struct rdma_cm_id *cmid = conn->ibc_cmid;
 919         kib_peer_t        *peer = conn->ibc_peer;
 920         int                rc;
 921
 922         LASSERT (!in_interrupt());
 923         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 924         LASSERT(list_empty(&conn->ibc_early_rxs));
 925         LASSERT(list_empty(&conn->ibc_tx_noops));
 926         LASSERT(list_empty(&conn->ibc_tx_queue));
 927         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 928         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 929         LASSERT(list_empty(&conn->ibc_active_txs));
 930         LASSERT (conn->ibc_noops_posted == 0);
 931         LASSERT (conn->ibc_nsends_posted == 0);
 932
 933         switch (conn->ibc_state) {
 934         default:
 935                 /* conn must be completely disengaged from the network */
 936                 LBUG();
 937
 938         case IBLND_CONN_DISCONNECTED:
 939                 /* connvars should have been freed already */
 940                 LASSERT (conn->ibc_connvars == NULL);
 941                 break;
 942
 943         case IBLND_CONN_INIT:
 944                 break;
 945         }
 946
 947         /* conn->ibc_cmid might be destroyed by CM already */
 948         if (cmid != NULL && cmid->qp != NULL)
 949                 rdma_destroy_qp(cmid);
 950
 951         if (conn->ibc_cq != NULL) {
 952                 rc = ib_destroy_cq(conn->ibc_cq);
 953                 if (rc != 0)
 954                         CWARN("Error destroying CQ: %d\n", rc);
 955         }
 956
 957         if (conn->ibc_rx_pages != NULL)
 958                 kiblnd_unmap_rx_descs(conn);
 959
 960         if (conn->ibc_rxs != NULL) {
 961                 LIBCFS_FREE(conn->ibc_rxs,
 962                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 963         }
 964
 965         if (conn->ibc_connvars != NULL)
 966                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 967
 968         if (conn->ibc_hdev != NULL)
 969                 kiblnd_hdev_decref(conn->ibc_hdev);
 970
 971         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 972         if (conn->ibc_state != IBLND_CONN_INIT) {
 973                 kib_net_t *net = peer->ibp_ni->ni_data;
 974
 975                 kiblnd_peer_decref(peer);
 976                 rdma_destroy_id(cmid);
 977                 atomic_dec(&net->ibn_nconns);
 978         }
 979
 980         LIBCFS_FREE(conn, sizeof(*conn));
 981 }
 982
 983 int
 984 kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 985 {
 986         kib_conn_t              *conn;
 987         struct list_head        *ctmp;
 988         struct list_head        *cnxt;
 989         int                     count = 0;
 990
 991         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 992                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 993
 994                 CDEBUG(D_NET, "Closing conn -> %s, "
 995                               "version: %x, reason: %d\n",
 996                        libcfs_nid2str(peer->ibp_nid),
 997                        conn->ibc_version, why);
 998
 999                 kiblnd_close_conn_locked(conn, why);
1000                 count++;
1001         }
1002
1003         return count;
1004 }
1005
1006 int
1007 kiblnd_close_stale_conns_locked(kib_peer_t *peer,
1008                                 int version, __u64 incarnation)
1009 {
1010         kib_conn_t              *conn;
1011         struct list_head        *ctmp;
1012         struct list_head        *cnxt;
1013         int                     count = 0;
1014
1015         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
1016                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1017
1018                 if (conn->ibc_version     == version &&
1019                     conn->ibc_incarnation == incarnation)
1020                         continue;
1021
1022                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1023                               "incarnation:"LPX64"(%x, "LPX64")\n",
1024                        libcfs_nid2str(peer->ibp_nid),
1025                        conn->ibc_version, conn->ibc_incarnation,
1026                        version, incarnation);
1027
1028                 kiblnd_close_conn_locked(conn, -ESTALE);
1029                 count++;
1030         }
1031
1032         return count;
1033 }
1034
1035 static int
1036 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
1037 {
1038         kib_peer_t              *peer;
1039         struct list_head        *ptmp;
1040         struct list_head        *pnxt;
1041         int                     lo;
1042         int                     hi;
1043         int                     i;
1044         unsigned long           flags;
1045         int                     count = 0;
1046
1047         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1048
1049         if (nid != LNET_NID_ANY)
1050                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1051         else {
1052                 lo = 0;
1053                 hi = kiblnd_data.kib_peer_hash_size - 1;
1054         }
1055
1056         for (i = lo; i <= hi; i++) {
1057                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1058
1059                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1060                         LASSERT(peer->ibp_connecting > 0 ||
1061                                 peer->ibp_accepting > 0 ||
1062                                 !list_empty(&peer->ibp_conns));
1063
1064                         if (peer->ibp_ni != ni)
1065                                 continue;
1066
1067                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1068                                 continue;
1069
1070                         count += kiblnd_close_peer_conns_locked(peer, 0);
1071                 }
1072         }
1073
1074         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1075
1076         /* wildcards always succeed */
1077         if (nid == LNET_NID_ANY)
1078                 return 0;
1079
1080         return (count == 0) ? -ENOENT : 0;
1081 }
1082
1083 static int
1084 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1085 {
1086         struct libcfs_ioctl_data *data = arg;
1087         int                       rc = -EINVAL;
1088
1089         switch(cmd) {
1090         case IOC_LIBCFS_GET_PEER: {
1091                 lnet_nid_t   nid = 0;
1092                 int          count = 0;
1093
1094                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1095                                           &nid, &count);
1096                 data->ioc_nid    = nid;
1097                 data->ioc_count  = count;
1098                 break;
1099         }
1100
1101         case IOC_LIBCFS_DEL_PEER: {
1102                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1103                 break;
1104         }
1105         case IOC_LIBCFS_GET_CONN: {
1106                 kib_conn_t *conn;
1107
1108                 rc = 0;
1109                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1110                 if (conn == NULL) {
1111                         rc = -ENOENT;
1112                         break;
1113                 }
1114
1115                 LASSERT (conn->ibc_cmid != NULL);
1116                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1117                 if (conn->ibc_cmid->route.path_rec == NULL)
1118                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1119                 else
1120                         data->ioc_u32[0] =
1121                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1122                 kiblnd_conn_decref(conn);
1123                 break;
1124         }
1125         case IOC_LIBCFS_CLOSE_CONNECTION: {
1126                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1127                 break;
1128         }
1129
1130         default:
1131                 break;
1132         }
1133
1134         return rc;
1135 }
1136
1137 static void
1138 kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1139 {
1140         cfs_time_t      last_alive = 0;
1141         cfs_time_t      now = cfs_time_current();
1142         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1143         kib_peer_t      *peer;
1144         unsigned long   flags;
1145
1146         read_lock_irqsave(glock, flags);
1147
1148         peer = kiblnd_find_peer_locked(nid);
1149         if (peer != NULL) {
1150                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1151                         peer->ibp_accepting > 0 ||
1152                         !list_empty(&peer->ibp_conns));  /* active conn */
1153                 last_alive = peer->ibp_last_alive;
1154         }
1155
1156         read_unlock_irqrestore(glock, flags);
1157
1158         if (last_alive != 0)
1159                 *when = last_alive;
1160
1161         /* peer is not persistent in hash, trigger peer creation
1162          * and connection establishment with a NULL tx */
1163         if (peer == NULL)
1164                 kiblnd_launch_tx(ni, NULL, nid);
1165
1166         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1167                libcfs_nid2str(nid), peer,
1168                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1169         return;
1170 }
1171
1172 static void
1173 kiblnd_free_pages(kib_pages_t *p)
1174 {
1175         int     npages = p->ibp_npages;
1176         int     i;
1177
1178         for (i = 0; i < npages; i++) {
1179                 if (p->ibp_pages[i] != NULL)
1180                         __free_page(p->ibp_pages[i]);
1181         }
1182
1183         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1184 }
1185
1186 int
1187 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1188 {
1189         kib_pages_t     *p;
1190         int             i;
1191
1192         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1193                          offsetof(kib_pages_t, ibp_pages[npages]));
1194         if (p == NULL) {
1195                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1196                 return -ENOMEM;
1197         }
1198
1199         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1200         p->ibp_npages = npages;
1201
1202         for (i = 0; i < npages; i++) {
1203                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1204                                                      GFP_NOFS);
1205                 if (p->ibp_pages[i] == NULL) {
1206                         CERROR("Can't allocate page %d of %d\n", i, npages);
1207                         kiblnd_free_pages(p);
1208                         return -ENOMEM;
1209                 }
1210         }
1211
1212         *pp = p;
1213         return 0;
1214 }
1215
1216 void
1217 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1218 {
1219         kib_rx_t *rx;
1220         int       i;
1221
1222         LASSERT (conn->ibc_rxs != NULL);
1223         LASSERT (conn->ibc_hdev != NULL);
1224
1225         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1226                 rx = &conn->ibc_rxs[i];
1227
1228                 LASSERT (rx->rx_nob >= 0); /* not posted */
1229
1230                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1231                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1232                                                           rx->rx_msgaddr),
1233                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1234         }
1235
1236         kiblnd_free_pages(conn->ibc_rx_pages);
1237
1238         conn->ibc_rx_pages = NULL;
1239 }
1240
1241 void
1242 kiblnd_map_rx_descs(kib_conn_t *conn)
1243 {
1244         kib_rx_t       *rx;
1245         struct page    *pg;
1246         int             pg_off;
1247         int             ipg;
1248         int             i;
1249
1250         for (pg_off = ipg = i = 0;
1251              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1252                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1253                 rx = &conn->ibc_rxs[i];
1254
1255                 rx->rx_conn = conn;
1256                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1257
1258                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1259                                                        rx->rx_msg, IBLND_MSG_SIZE,
1260                                                        DMA_FROM_DEVICE);
1261                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1262                                                    rx->rx_msgaddr));
1263                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1264
1265                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1266                        i, rx->rx_msg, rx->rx_msgaddr,
1267                        lnet_page2phys(pg) + pg_off);
1268
1269                 pg_off += IBLND_MSG_SIZE;
1270                 LASSERT (pg_off <= PAGE_SIZE);
1271
1272                 if (pg_off == PAGE_SIZE) {
1273                         pg_off = 0;
1274                         ipg++;
1275                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1276                 }
1277         }
1278 }
1279
1280 static void
1281 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1282 {
1283         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1284         kib_tx_t       *tx;
1285         int             i;
1286
1287         LASSERT (tpo->tpo_pool.po_allocated == 0);
1288
1289         if (hdev == NULL)
1290                 return;
1291
1292         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1293                 tx = &tpo->tpo_tx_descs[i];
1294                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1295                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1296                                                           tx->tx_msgaddr),
1297                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1298         }
1299
1300         kiblnd_hdev_decref(hdev);
1301         tpo->tpo_hdev = NULL;
1302 }
1303
1304 static kib_hca_dev_t *
1305 kiblnd_current_hdev(kib_dev_t *dev)
1306 {
1307         kib_hca_dev_t *hdev;
1308         unsigned long  flags;
1309         int            i = 0;
1310
1311         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1312         while (dev->ibd_failover) {
1313                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1314                 if (i++ % 50 == 0)
1315                         CDEBUG(D_NET, "%s: Wait for failover\n",
1316                                dev->ibd_ifname);
1317                 schedule_timeout(cfs_time_seconds(1) / 100);
1318
1319                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1320         }
1321
1322         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1323         hdev = dev->ibd_hdev;
1324
1325         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1326
1327         return hdev;
1328 }
1329
1330 static void
1331 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1332 {
1333         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1334         kib_pool_t     *pool  = &tpo->tpo_pool;
1335         kib_net_t      *net   = pool->po_owner->ps_net;
1336         kib_dev_t      *dev;
1337         struct page    *page;
1338         kib_tx_t       *tx;
1339         int             page_offset;
1340         int             ipage;
1341         int             i;
1342
1343         LASSERT (net != NULL);
1344
1345         dev = net->ibn_dev;
1346
1347         /* pre-mapped messages are not bigger than 1 page */
1348         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1349
1350         /* No fancy arithmetic when we do the buffer calculations */
1351         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1352
1353         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1354
1355         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1356                 page = txpgs->ibp_pages[ipage];
1357                 tx = &tpo->tpo_tx_descs[i];
1358
1359                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1360                                            page_offset);
1361
1362                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1363                                                        tx->tx_msg,
1364                                                        IBLND_MSG_SIZE,
1365                                                        DMA_TO_DEVICE);
1366                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1367                                                   tx->tx_msgaddr));
1368                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1369
1370                 list_add(&tx->tx_list, &pool->po_free_list);
1371
1372                 page_offset += IBLND_MSG_SIZE;
1373                 LASSERT(page_offset <= PAGE_SIZE);
1374
1375                 if (page_offset == PAGE_SIZE) {
1376                         page_offset = 0;
1377                         ipage++;
1378                         LASSERT(ipage <= txpgs->ibp_npages);
1379                 }
1380         }
1381 }
1382
1383 struct ib_mr *
1384 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1385 {
1386         __u64   index;
1387
1388         LASSERT (hdev->ibh_mrs[0] != NULL);
1389
1390         if (hdev->ibh_nmrs == 1)
1391                 return hdev->ibh_mrs[0];
1392
1393         index = addr >> hdev->ibh_mr_shift;
1394
1395         if (index <  hdev->ibh_nmrs &&
1396             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1397                 return hdev->ibh_mrs[index];
1398
1399         return NULL;
1400 }
1401
1402 struct ib_mr *
1403 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1404 {
1405         struct ib_mr *prev_mr;
1406         struct ib_mr *mr;
1407         int           i;
1408
1409         LASSERT (hdev->ibh_mrs[0] != NULL);
1410
1411         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1412             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1413                 return NULL;
1414
1415         if (hdev->ibh_nmrs == 1)
1416                 return hdev->ibh_mrs[0];
1417
1418         for (i = 0, mr = prev_mr = NULL;
1419              i < rd->rd_nfrags; i++) {
1420                 mr = kiblnd_find_dma_mr(hdev,
1421                                         rd->rd_frags[i].rf_addr,
1422                                         rd->rd_frags[i].rf_nob);
1423                 if (prev_mr == NULL)
1424                         prev_mr = mr;
1425
1426                 if (mr == NULL || prev_mr != mr) {
1427                         /* Can't covered by one single MR */
1428                         mr = NULL;
1429                         break;
1430                 }
1431         }
1432
1433         return mr;
1434 }
1435
1436 static void
1437 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1438 {
1439         LASSERT (pool->fpo_map_count == 0);
1440
1441         if (pool->fpo_fmr_pool != NULL)
1442                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1443
1444         if (pool->fpo_hdev != NULL)
1445                 kiblnd_hdev_decref(pool->fpo_hdev);
1446
1447         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1448 }
1449
1450 static void
1451 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1452 {
1453         kib_fmr_pool_t *pool;
1454
1455         while (!list_empty(head)) {
1456                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1457                 list_del(&pool->fpo_list);
1458                 kiblnd_destroy_fmr_pool(pool);
1459         }
1460 }
1461
1462 static int kiblnd_fmr_pool_size(int ncpts)
1463 {
1464         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1465
1466         return max(IBLND_FMR_POOL, size);
1467 }
1468
1469 static int kiblnd_fmr_flush_trigger(int ncpts)
1470 {
1471         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1472
1473         return max(IBLND_FMR_POOL_FLUSH, size);
1474 }
1475
1476 static int
1477 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1478 {
1479         /* FMR pool for RDMA */
1480         kib_dev_t               *dev = fps->fps_net->ibn_dev;
1481         kib_fmr_pool_t          *fpo;
1482         struct ib_fmr_pool_param param = {
1483                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1484                 .page_shift        = PAGE_SHIFT,
1485                 .access            = (IB_ACCESS_LOCAL_WRITE |
1486                                       IB_ACCESS_REMOTE_WRITE),
1487                 .pool_size         = fps->fps_pool_size,
1488                 .dirty_watermark   = fps->fps_flush_trigger,
1489                 .flush_function    = NULL,
1490                 .flush_arg         = NULL,
1491                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1492         int rc;
1493
1494         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1495         if (fpo == NULL)
1496                 return -ENOMEM;
1497
1498         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1499
1500         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1501         if (IS_ERR(fpo->fpo_fmr_pool)) {
1502                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1503                 CERROR("Failed to create FMR pool: %d\n", rc);
1504
1505                 kiblnd_hdev_decref(fpo->fpo_hdev);
1506                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1507                 return rc;
1508         }
1509
1510         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1511         fpo->fpo_owner    = fps;
1512         *pp_fpo = fpo;
1513
1514         return 0;
1515 }
1516
1517 static void
1518 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1519 {
1520         if (fps->fps_net == NULL) /* intialized? */
1521                 return;
1522
1523         spin_lock(&fps->fps_lock);
1524
1525         while (!list_empty(&fps->fps_pool_list)) {
1526                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1527                                                  kib_fmr_pool_t, fpo_list);
1528                 fpo->fpo_failed = 1;
1529                 list_del(&fpo->fpo_list);
1530                 if (fpo->fpo_map_count == 0)
1531                         list_add(&fpo->fpo_list, zombies);
1532                 else
1533                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1534         }
1535
1536         spin_unlock(&fps->fps_lock);
1537 }
1538
1539 static void
1540 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1541 {
1542         if (fps->fps_net != NULL) { /* initialized? */
1543                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1544                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1545         }
1546 }
1547
1548 static int
1549 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1550                         int pool_size, int flush_trigger)
1551 {
1552         kib_fmr_pool_t *fpo;
1553         int             rc;
1554
1555         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1556
1557         fps->fps_net = net;
1558         fps->fps_cpt = cpt;
1559         fps->fps_pool_size = pool_size;
1560         fps->fps_flush_trigger = flush_trigger;
1561         spin_lock_init(&fps->fps_lock);
1562         INIT_LIST_HEAD(&fps->fps_pool_list);
1563         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1564
1565         rc = kiblnd_create_fmr_pool(fps, &fpo);
1566         if (rc == 0)
1567                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1568
1569         return rc;
1570 }
1571
1572 static int
1573 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1574 {
1575         if (fpo->fpo_map_count != 0) /* still in use */
1576                 return 0;
1577         if (fpo->fpo_failed)
1578                 return 1;
1579         return cfs_time_aftereq(now, fpo->fpo_deadline);
1580 }
1581
1582 void
1583 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1584 {
1585         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1586         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1587         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1588         cfs_time_t         now = cfs_time_current();
1589         kib_fmr_pool_t    *tmp;
1590         int                rc;
1591
1592         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1593         LASSERT(rc == 0);
1594
1595         if (status != 0) {
1596                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1597                 LASSERT(rc == 0);
1598         }
1599
1600         fmr->fmr_pool = NULL;
1601         fmr->fmr_pfmr = NULL;
1602
1603         spin_lock(&fps->fps_lock);
1604         fpo->fpo_map_count--;   /* decref the pool */
1605
1606         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1607                 /* the first pool is persistent */
1608                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1609                         continue;
1610
1611                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1612                         list_move(&fpo->fpo_list, &zombies);
1613                         fps->fps_version++;
1614                 }
1615         }
1616         spin_unlock(&fps->fps_lock);
1617
1618         if (!list_empty(&zombies))
1619                 kiblnd_destroy_fmr_pool_list(&zombies);
1620 }
1621
1622 int
1623 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1624                     __u64 iov, kib_fmr_t *fmr)
1625 {
1626         struct ib_pool_fmr *pfmr;
1627         kib_fmr_pool_t     *fpo;
1628         __u64               version;
1629         int                 rc;
1630
1631 again:
1632         spin_lock(&fps->fps_lock);
1633         version = fps->fps_version;
1634         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1635                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1636                 fpo->fpo_map_count++;
1637                 spin_unlock(&fps->fps_lock);
1638
1639                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1640                                             pages, npages, iov);
1641                 if (likely(!IS_ERR(pfmr))) {
1642                         fmr->fmr_pool = fpo;
1643                         fmr->fmr_pfmr = pfmr;
1644                         return 0;
1645                 }
1646
1647                 spin_lock(&fps->fps_lock);
1648                 fpo->fpo_map_count--;
1649                 if (PTR_ERR(pfmr) != -EAGAIN) {
1650                         spin_unlock(&fps->fps_lock);
1651                         return PTR_ERR(pfmr);
1652                 }
1653
1654                 /* EAGAIN and ... */
1655                 if (version != fps->fps_version) {
1656                         spin_unlock(&fps->fps_lock);
1657                         goto again;
1658                 }
1659         }
1660
1661         if (fps->fps_increasing) {
1662                 spin_unlock(&fps->fps_lock);
1663                 CDEBUG(D_NET, "Another thread is allocating new "
1664                        "FMR pool, waiting for her to complete\n");
1665                 schedule();
1666                 goto again;
1667
1668         }
1669
1670         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1671                 /* someone failed recently */
1672                 spin_unlock(&fps->fps_lock);
1673                 return -EAGAIN;
1674         }
1675
1676         fps->fps_increasing = 1;
1677         spin_unlock(&fps->fps_lock);
1678
1679         CDEBUG(D_NET, "Allocate new FMR pool\n");
1680         rc = kiblnd_create_fmr_pool(fps, &fpo);
1681         spin_lock(&fps->fps_lock);
1682         fps->fps_increasing = 0;
1683         if (rc == 0) {
1684                 fps->fps_version++;
1685                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1686         } else {
1687                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1688         }
1689         spin_unlock(&fps->fps_lock);
1690
1691         goto again;
1692 }
1693
1694 static void
1695 kiblnd_fini_pool(kib_pool_t *pool)
1696 {
1697         LASSERT(list_empty(&pool->po_free_list));
1698         LASSERT(pool->po_allocated == 0);
1699
1700         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1701 }
1702
1703 static void
1704 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1705 {
1706         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1707
1708         memset(pool, 0, sizeof(kib_pool_t));
1709         INIT_LIST_HEAD(&pool->po_free_list);
1710         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1711         pool->po_owner    = ps;
1712         pool->po_size     = size;
1713 }
1714
1715 static void
1716 kiblnd_destroy_pool_list(struct list_head *head)
1717 {
1718         kib_pool_t *pool;
1719
1720         while (!list_empty(head)) {
1721                 pool = list_entry(head->next, kib_pool_t, po_list);
1722                 list_del(&pool->po_list);
1723
1724                 LASSERT(pool->po_owner != NULL);
1725                 pool->po_owner->ps_pool_destroy(pool);
1726         }
1727 }
1728
1729 static void
1730 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1731 {
1732         if (ps->ps_net == NULL) /* intialized? */
1733                 return;
1734
1735         spin_lock(&ps->ps_lock);
1736         while (!list_empty(&ps->ps_pool_list)) {
1737                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1738                                             kib_pool_t, po_list);
1739                 po->po_failed = 1;
1740                 list_del(&po->po_list);
1741                 if (po->po_allocated == 0)
1742                         list_add(&po->po_list, zombies);
1743                 else
1744                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1745         }
1746         spin_unlock(&ps->ps_lock);
1747 }
1748
1749 static void
1750 kiblnd_fini_poolset(kib_poolset_t *ps)
1751 {
1752         if (ps->ps_net != NULL) { /* initialized? */
1753                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1754                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1755         }
1756 }
1757
1758 static int
1759 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1760                     kib_net_t *net, char *name, int size,
1761                     kib_ps_pool_create_t po_create,
1762                     kib_ps_pool_destroy_t po_destroy,
1763                     kib_ps_node_init_t nd_init,
1764                     kib_ps_node_fini_t nd_fini)
1765 {
1766         kib_pool_t      *pool;
1767         int             rc;
1768
1769         memset(ps, 0, sizeof(kib_poolset_t));
1770
1771         ps->ps_cpt          = cpt;
1772         ps->ps_net          = net;
1773         ps->ps_pool_create  = po_create;
1774         ps->ps_pool_destroy = po_destroy;
1775         ps->ps_node_init    = nd_init;
1776         ps->ps_node_fini    = nd_fini;
1777         ps->ps_pool_size    = size;
1778         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1779             >= sizeof(ps->ps_name))
1780                 return -E2BIG;
1781         spin_lock_init(&ps->ps_lock);
1782         INIT_LIST_HEAD(&ps->ps_pool_list);
1783         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1784
1785         rc = ps->ps_pool_create(ps, size, &pool);
1786         if (rc == 0)
1787                 list_add(&pool->po_list, &ps->ps_pool_list);
1788         else
1789                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1790
1791         return rc;
1792 }
1793
1794 static int
1795 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1796 {
1797         if (pool->po_allocated != 0) /* still in use */
1798                 return 0;
1799         if (pool->po_failed)
1800                 return 1;
1801         return cfs_time_aftereq(now, pool->po_deadline);
1802 }
1803
1804 void
1805 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1806 {
1807         struct list_head zombies = LIST_HEAD_INIT(zombies);
1808         kib_poolset_t   *ps = pool->po_owner;
1809         kib_pool_t      *tmp;
1810         cfs_time_t       now = cfs_time_current();
1811
1812         spin_lock(&ps->ps_lock);
1813
1814         if (ps->ps_node_fini != NULL)
1815                 ps->ps_node_fini(pool, node);
1816
1817         LASSERT(pool->po_allocated > 0);
1818         list_add(node, &pool->po_free_list);
1819         pool->po_allocated--;
1820
1821         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1822                 /* the first pool is persistent */
1823                 if (ps->ps_pool_list.next == &pool->po_list)
1824                         continue;
1825
1826                 if (kiblnd_pool_is_idle(pool, now))
1827                         list_move(&pool->po_list, &zombies);
1828         }
1829         spin_unlock(&ps->ps_lock);
1830
1831         if (!list_empty(&zombies))
1832                 kiblnd_destroy_pool_list(&zombies);
1833 }
1834
1835 struct list_head *
1836 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1837 {
1838         struct list_head        *node;
1839         kib_pool_t              *pool;
1840         int                     rc;
1841
1842 again:
1843         spin_lock(&ps->ps_lock);
1844         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1845                 if (list_empty(&pool->po_free_list))
1846                         continue;
1847
1848                 pool->po_allocated++;
1849                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1850                 node = pool->po_free_list.next;
1851                 list_del(node);
1852
1853                 if (ps->ps_node_init != NULL) {
1854                         /* still hold the lock */
1855                         ps->ps_node_init(pool, node);
1856                 }
1857                 spin_unlock(&ps->ps_lock);
1858                 return node;
1859         }
1860
1861         /* no available tx pool and ... */
1862         if (ps->ps_increasing) {
1863                 /* another thread is allocating a new pool */
1864                 spin_unlock(&ps->ps_lock);
1865                 CDEBUG(D_NET, "Another thread is allocating new "
1866                        "%s pool, waiting for her to complete\n",
1867                        ps->ps_name);
1868                 schedule();
1869                 goto again;
1870         }
1871
1872         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1873                 /* someone failed recently */
1874                 spin_unlock(&ps->ps_lock);
1875                 return NULL;
1876         }
1877
1878         ps->ps_increasing = 1;
1879         spin_unlock(&ps->ps_lock);
1880
1881         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1882
1883         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1884
1885         spin_lock(&ps->ps_lock);
1886         ps->ps_increasing = 0;
1887         if (rc == 0) {
1888                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1889         } else {
1890                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1891                 CERROR("Can't allocate new %s pool because out of memory\n",
1892                        ps->ps_name);
1893         }
1894         spin_unlock(&ps->ps_lock);
1895
1896         goto again;
1897 }
1898
1899 void
1900 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1901 {
1902         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1903         struct ib_mr        *mr  = pmr->pmr_mr;
1904
1905         pmr->pmr_mr = NULL;
1906         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1907         if (mr != NULL)
1908                 ib_dereg_mr(mr);
1909 }
1910
1911 int
1912 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1913                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1914 {
1915         kib_phys_mr_t    *pmr;
1916         struct list_head *node;
1917         int               rc;
1918         int               i;
1919
1920         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1921         if (node == NULL) {
1922                 CERROR("Failed to allocate PMR descriptor\n");
1923                 return -ENOMEM;
1924         }
1925
1926         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1927         if (pmr->pmr_pool->ppo_hdev != hdev) {
1928                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1929                 return -EAGAIN;
1930         }
1931
1932         for (i = 0; i < rd->rd_nfrags; i ++) {
1933                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1934                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1935         }
1936
1937         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1938                                      pmr->pmr_ipb, rd->rd_nfrags,
1939                                      IB_ACCESS_LOCAL_WRITE |
1940                                      IB_ACCESS_REMOTE_WRITE,
1941                                      iova);
1942         if (!IS_ERR(pmr->pmr_mr)) {
1943                 pmr->pmr_iova = *iova;
1944                 *pp_pmr = pmr;
1945                 return 0;
1946         }
1947
1948         rc = PTR_ERR(pmr->pmr_mr);
1949         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1950
1951         pmr->pmr_mr = NULL;
1952         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1953
1954         return rc;
1955 }
1956
1957 static void
1958 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1959 {
1960         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1961         kib_phys_mr_t  *pmr;
1962
1963         LASSERT (pool->po_allocated == 0);
1964
1965         while (!list_empty(&pool->po_free_list)) {
1966                 pmr = list_entry(pool->po_free_list.next,
1967                                      kib_phys_mr_t, pmr_list);
1968
1969                 LASSERT (pmr->pmr_mr == NULL);
1970                 list_del(&pmr->pmr_list);
1971
1972                 if (pmr->pmr_ipb != NULL) {
1973                         LIBCFS_FREE(pmr->pmr_ipb,
1974                                     IBLND_MAX_RDMA_FRAGS *
1975                                     sizeof(struct ib_phys_buf));
1976                 }
1977
1978                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1979         }
1980
1981         kiblnd_fini_pool(pool);
1982         if (ppo->ppo_hdev != NULL)
1983                 kiblnd_hdev_decref(ppo->ppo_hdev);
1984
1985         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1986 }
1987
1988 static inline int kiblnd_pmr_pool_size(int ncpts)
1989 {
1990         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1991
1992         return max(IBLND_PMR_POOL, size);
1993 }
1994
1995 static int
1996 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1997 {
1998         struct kib_pmr_pool     *ppo;
1999         struct kib_pool         *pool;
2000         kib_phys_mr_t           *pmr;
2001         int                     i;
2002
2003         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2004                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2005         if (ppo == NULL) {
2006                 CERROR("Failed to allocate PMR pool\n");
2007                 return -ENOMEM;
2008         }
2009
2010         pool = &ppo->ppo_pool;
2011         kiblnd_init_pool(ps, pool, size);
2012
2013         for (i = 0; i < size; i++) {
2014                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2015                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2016                 if (pmr == NULL)
2017                         break;
2018
2019                 pmr->pmr_pool = ppo;
2020                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2021                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2022                 if (pmr->pmr_ipb == NULL)
2023                         break;
2024
2025                 list_add(&pmr->pmr_list, &pool->po_free_list);
2026         }
2027
2028         if (i < size) {
2029                 ps->ps_pool_destroy(pool);
2030                 return -ENOMEM;
2031         }
2032
2033         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2034         *pp_po = pool;
2035         return 0;
2036 }
2037
2038 static void
2039 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2040 {
2041         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2042         int             i;
2043
2044         LASSERT (pool->po_allocated == 0);
2045
2046         if (tpo->tpo_tx_pages != NULL) {
2047                 kiblnd_unmap_tx_pool(tpo);
2048                 kiblnd_free_pages(tpo->tpo_tx_pages);
2049         }
2050
2051         if (tpo->tpo_tx_descs == NULL)
2052                 goto out;
2053
2054         for (i = 0; i < pool->po_size; i++) {
2055                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2056
2057                 list_del(&tx->tx_list);
2058                 if (tx->tx_pages != NULL)
2059                         LIBCFS_FREE(tx->tx_pages,
2060                                     LNET_MAX_IOV *
2061                                     sizeof(*tx->tx_pages));
2062                 if (tx->tx_frags != NULL)
2063                         LIBCFS_FREE(tx->tx_frags,
2064                                     IBLND_MAX_RDMA_FRAGS *
2065                                             sizeof(*tx->tx_frags));
2066                 if (tx->tx_wrq != NULL)
2067                         LIBCFS_FREE(tx->tx_wrq,
2068                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2069                                     sizeof(*tx->tx_wrq));
2070                 if (tx->tx_sge != NULL)
2071                         LIBCFS_FREE(tx->tx_sge,
2072                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2073                                     sizeof(*tx->tx_sge));
2074                 if (tx->tx_rd != NULL)
2075                         LIBCFS_FREE(tx->tx_rd,
2076                                     offsetof(kib_rdma_desc_t,
2077                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2078         }
2079
2080         LIBCFS_FREE(tpo->tpo_tx_descs,
2081                     pool->po_size * sizeof(kib_tx_t));
2082 out:
2083         kiblnd_fini_pool(pool);
2084         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2085 }
2086
2087 static int kiblnd_tx_pool_size(int ncpts)
2088 {
2089         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2090
2091         return max(IBLND_TX_POOL, ntx);
2092 }
2093
2094 static int
2095 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2096 {
2097         int            i;
2098         int            npg;
2099         kib_pool_t    *pool;
2100         kib_tx_pool_t *tpo;
2101
2102         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2103         if (tpo == NULL) {
2104                 CERROR("Failed to allocate TX pool\n");
2105                 return -ENOMEM;
2106         }
2107
2108         pool = &tpo->tpo_pool;
2109         kiblnd_init_pool(ps, pool, size);
2110         tpo->tpo_tx_descs = NULL;
2111         tpo->tpo_tx_pages = NULL;
2112
2113         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2114         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2115                 CERROR("Can't allocate tx pages: %d\n", npg);
2116                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2117                 return -ENOMEM;
2118         }
2119
2120         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2121                          size * sizeof(kib_tx_t));
2122         if (tpo->tpo_tx_descs == NULL) {
2123                 CERROR("Can't allocate %d tx descriptors\n", size);
2124                 ps->ps_pool_destroy(pool);
2125                 return -ENOMEM;
2126         }
2127
2128         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2129
2130         for (i = 0; i < size; i++) {
2131                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2132
2133                 tx->tx_pool = tpo;
2134                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2135                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2136                                          lnet_cpt_table(), ps->ps_cpt,
2137                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2138                         if (tx->tx_pages == NULL)
2139                                 break;
2140                 }
2141
2142                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2143                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2144                 if (tx->tx_frags == NULL)
2145                         break;
2146
2147                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2148
2149                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2150                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2151                                  sizeof(*tx->tx_wrq));
2152                 if (tx->tx_wrq == NULL)
2153                         break;
2154
2155                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2156                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2157                                  sizeof(*tx->tx_sge));
2158                 if (tx->tx_sge == NULL)
2159                         break;
2160
2161                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2162                                  offsetof(kib_rdma_desc_t,
2163                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2164                 if (tx->tx_rd == NULL)
2165                         break;
2166         }
2167
2168         if (i == size) {
2169                 kiblnd_map_tx_pool(tpo);
2170                 *pp_po = pool;
2171                 return 0;
2172         }
2173
2174         ps->ps_pool_destroy(pool);
2175         return -ENOMEM;
2176 }
2177
2178 static void
2179 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2180 {
2181         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2182                                              tps_poolset);
2183         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2184
2185         tx->tx_cookie = tps->tps_next_tx_cookie++;
2186 }
2187
2188 static void
2189 kiblnd_net_fini_pools(kib_net_t *net)
2190 {
2191         int     i;
2192
2193         cfs_cpt_for_each(i, lnet_cpt_table()) {
2194                 kib_tx_poolset_t        *tps;
2195                 kib_fmr_poolset_t       *fps;
2196                 kib_pmr_poolset_t       *pps;
2197
2198                 if (net->ibn_tx_ps != NULL) {
2199                         tps = net->ibn_tx_ps[i];
2200                         kiblnd_fini_poolset(&tps->tps_poolset);
2201                 }
2202
2203                 if (net->ibn_fmr_ps != NULL) {
2204                         fps = net->ibn_fmr_ps[i];
2205                         kiblnd_fini_fmr_poolset(fps);
2206                 }
2207
2208                 if (net->ibn_pmr_ps != NULL) {
2209                         pps = net->ibn_pmr_ps[i];
2210                         kiblnd_fini_poolset(&pps->pps_poolset);
2211                 }
2212         }
2213
2214         if (net->ibn_tx_ps != NULL) {
2215                 cfs_percpt_free(net->ibn_tx_ps);
2216                 net->ibn_tx_ps = NULL;
2217         }
2218
2219         if (net->ibn_fmr_ps != NULL) {
2220                 cfs_percpt_free(net->ibn_fmr_ps);
2221                 net->ibn_fmr_ps = NULL;
2222         }
2223
2224         if (net->ibn_pmr_ps != NULL) {
2225                 cfs_percpt_free(net->ibn_pmr_ps);
2226                 net->ibn_pmr_ps = NULL;
2227         }
2228 }
2229
2230 static int
2231 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2232 {
2233         unsigned long   flags;
2234         int             cpt;
2235         int             rc;
2236         int             i;
2237
2238         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2239         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2240             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2241                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2242                                            flags);
2243                 goto create_tx_pool;
2244         }
2245
2246         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2247
2248         if (*kiblnd_tunables.kib_fmr_pool_size <
2249             *kiblnd_tunables.kib_ntx / 4) {
2250                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2251                        *kiblnd_tunables.kib_fmr_pool_size,
2252                        *kiblnd_tunables.kib_ntx / 4);
2253                 rc = -EINVAL;
2254                 goto failed;
2255         }
2256
2257         /* TX pool must be created later than FMR/PMR, see LU-2268
2258          * for details */
2259         LASSERT(net->ibn_tx_ps == NULL);
2260
2261         /* premapping can fail if ibd_nmr > 1, so we always create
2262          * FMR/PMR pool and map-on-demand if premapping failed */
2263
2264         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2265                                            sizeof(kib_fmr_poolset_t));
2266         if (net->ibn_fmr_ps == NULL) {
2267                 CERROR("Failed to allocate FMR pool array\n");
2268                 rc = -ENOMEM;
2269                 goto failed;
2270         }
2271
2272         for (i = 0; i < ncpts; i++) {
2273                 cpt = (cpts == NULL) ? i : cpts[i];
2274                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2275                                              kiblnd_fmr_pool_size(ncpts),
2276                                              kiblnd_fmr_flush_trigger(ncpts));
2277                 if (rc == -ENOSYS && i == 0) /* no FMR */
2278                         break; /* create PMR pool */
2279
2280                 if (rc != 0) { /* a real error */
2281                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2282                                cpt, rc);
2283                         goto failed;
2284                 }
2285         }
2286
2287         if (i > 0) {
2288                 LASSERT(i == ncpts);
2289                 goto create_tx_pool;
2290         }
2291
2292         cfs_percpt_free(net->ibn_fmr_ps);
2293         net->ibn_fmr_ps = NULL;
2294
2295         CWARN("Device does not support FMR, failing back to PMR\n");
2296
2297         if (*kiblnd_tunables.kib_pmr_pool_size <
2298             *kiblnd_tunables.kib_ntx / 4) {
2299                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2300                        *kiblnd_tunables.kib_pmr_pool_size,
2301                        *kiblnd_tunables.kib_ntx / 4);
2302                 rc = -EINVAL;
2303                 goto failed;
2304         }
2305
2306         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2307                                            sizeof(kib_pmr_poolset_t));
2308         if (net->ibn_pmr_ps == NULL) {
2309                 CERROR("Failed to allocate PMR pool array\n");
2310                 rc = -ENOMEM;
2311                 goto failed;
2312         }
2313
2314         for (i = 0; i < ncpts; i++) {
2315                 cpt = (cpts == NULL) ? i : cpts[i];
2316                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2317                                          cpt, net, "PMR",
2318                                          kiblnd_pmr_pool_size(ncpts),
2319                                          kiblnd_create_pmr_pool,
2320                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2321                 if (rc != 0) {
2322                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2323                                cpt, rc);
2324                         goto failed;
2325                 }
2326         }
2327
2328  create_tx_pool:
2329         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2330                                           sizeof(kib_tx_poolset_t));
2331         if (net->ibn_tx_ps == NULL) {
2332                 CERROR("Failed to allocate tx pool array\n");
2333                 rc = -ENOMEM;
2334                 goto failed;
2335         }
2336
2337         for (i = 0; i < ncpts; i++) {
2338                 cpt = (cpts == NULL) ? i : cpts[i];
2339                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2340                                          cpt, net, "TX",
2341                                          kiblnd_tx_pool_size(ncpts),
2342                                          kiblnd_create_tx_pool,
2343                                          kiblnd_destroy_tx_pool,
2344                                          kiblnd_tx_init, NULL);
2345                 if (rc != 0) {
2346                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2347                                cpt, rc);
2348                         goto failed;
2349                 }
2350         }
2351
2352         return 0;
2353  failed:
2354         kiblnd_net_fini_pools(net);
2355         LASSERT(rc != 0);
2356         return rc;
2357 }
2358
2359 static int
2360 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2361 {
2362         struct ib_device_attr *attr;
2363         int                    rc;
2364
2365         /* It's safe to assume a HCA can handle a page size
2366          * matching that of the native system */
2367         hdev->ibh_page_shift = PAGE_SHIFT;
2368         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2369         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2370
2371         LIBCFS_ALLOC(attr, sizeof(*attr));
2372         if (attr == NULL) {
2373                 CERROR("Out of memory\n");
2374                 return -ENOMEM;
2375         }
2376
2377         rc = ib_query_device(hdev->ibh_ibdev, attr);
2378         if (rc == 0)
2379                 hdev->ibh_mr_size = attr->max_mr_size;
2380
2381         LIBCFS_FREE(attr, sizeof(*attr));
2382
2383         if (rc != 0) {
2384                 CERROR("Failed to query IB device: %d\n", rc);
2385                 return rc;
2386         }
2387
2388         if (hdev->ibh_mr_size == ~0ULL) {
2389                 hdev->ibh_mr_shift = 64;
2390                 return 0;
2391         }
2392
2393         for (hdev->ibh_mr_shift = 0;
2394              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2395                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2396                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2397                         return 0;
2398         }
2399
2400         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2401         return -EINVAL;
2402 }
2403
2404 static void
2405 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2406 {
2407         int     i;
2408
2409         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2410                 return;
2411
2412         for (i = 0; i < hdev->ibh_nmrs; i++) {
2413                 if (hdev->ibh_mrs[i] == NULL)
2414                         break;
2415
2416                 ib_dereg_mr(hdev->ibh_mrs[i]);
2417         }
2418
2419         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2420         hdev->ibh_mrs  = NULL;
2421         hdev->ibh_nmrs = 0;
2422 }
2423
2424 void
2425 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2426 {
2427         kiblnd_hdev_cleanup_mrs(hdev);
2428
2429         if (hdev->ibh_pd != NULL)
2430                 ib_dealloc_pd(hdev->ibh_pd);
2431
2432         if (hdev->ibh_cmid != NULL)
2433                 rdma_destroy_id(hdev->ibh_cmid);
2434
2435         LIBCFS_FREE(hdev, sizeof(*hdev));
2436 }
2437
2438 static int
2439 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2440 {
2441         struct ib_mr *mr;
2442         int           i;
2443         int           rc;
2444         __u64         mm_size;
2445         __u64         mr_size;
2446         int           acflags = IB_ACCESS_LOCAL_WRITE |
2447                                 IB_ACCESS_REMOTE_WRITE;
2448
2449         rc = kiblnd_hdev_get_attr(hdev);
2450         if (rc != 0)
2451                 return rc;
2452
2453         if (hdev->ibh_mr_shift == 64) {
2454                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2455                 if (hdev->ibh_mrs == NULL) {
2456                         CERROR("Failed to allocate MRs table\n");
2457                         return -ENOMEM;
2458                 }
2459
2460                 hdev->ibh_mrs[0] = NULL;
2461                 hdev->ibh_nmrs   = 1;
2462
2463                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2464                 if (IS_ERR(mr)) {
2465                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2466                         kiblnd_hdev_cleanup_mrs(hdev);
2467                         return PTR_ERR(mr);
2468                 }
2469
2470                 hdev->ibh_mrs[0] = mr;
2471
2472                 goto out;
2473         }
2474
2475         mr_size = (1ULL << hdev->ibh_mr_shift);
2476         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2477
2478         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2479
2480         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2481                 /* it's 4T..., assume we will re-code at that time */
2482                 CERROR("Can't support memory size: x"LPX64
2483                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2484                 return -EINVAL;
2485         }
2486
2487         /* create an array of MRs to cover all memory */
2488         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2489         if (hdev->ibh_mrs == NULL) {
2490                 CERROR("Failed to allocate MRs' table\n");
2491                 return -ENOMEM;
2492         }
2493
2494         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2495
2496         for (i = 0; i < hdev->ibh_nmrs; i++) {
2497                 struct ib_phys_buf ipb;
2498                 __u64              iova;
2499
2500                 ipb.size = hdev->ibh_mr_size;
2501                 ipb.addr = i * mr_size;
2502                 iova     = ipb.addr;
2503
2504                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2505                 if (IS_ERR(mr)) {
2506                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2507                                " size "LPX64" : %ld\n",
2508                                ipb.addr, ipb.size, PTR_ERR(mr));
2509                         kiblnd_hdev_cleanup_mrs(hdev);
2510                         return PTR_ERR(mr);
2511                 }
2512
2513                 LASSERT (iova == ipb.addr);
2514
2515                 hdev->ibh_mrs[i] = mr;
2516         }
2517
2518 out:
2519         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2520                 LCONSOLE_INFO("Register global MR array, MR size: "
2521                               LPX64", array size: %d\n",
2522                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2523         return 0;
2524 }
2525
2526 static int
2527 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2528 {       /* DUMMY */
2529         return 0;
2530 }
2531
2532 static int
2533 kiblnd_dev_need_failover(kib_dev_t *dev)
2534 {
2535         struct rdma_cm_id  *cmid;
2536         struct sockaddr_in  srcaddr;
2537         struct sockaddr_in  dstaddr;
2538         int                 rc;
2539
2540         if (dev->ibd_hdev == NULL || /* initializing */
2541             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2542             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2543                 return 1;
2544
2545         /* XXX: it's UGLY, but I don't have better way to find
2546          * ib-bonding HCA failover because:
2547          *
2548          * a. no reliable CM event for HCA failover...
2549          * b. no OFED API to get ib_device for current net_device...
2550          *
2551          * We have only two choices at this point:
2552          *
2553          * a. rdma_bind_addr(), it will conflict with listener cmid
2554          * b. rdma_resolve_addr() to zero addr */
2555         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2556                                      IB_QPT_RC);
2557         if (IS_ERR(cmid)) {
2558                 rc = PTR_ERR(cmid);
2559                 CERROR("Failed to create cmid for failover: %d\n", rc);
2560                 return rc;
2561         }
2562
2563         memset(&srcaddr, 0, sizeof(srcaddr));
2564         srcaddr.sin_family      = AF_INET;
2565         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2566
2567         memset(&dstaddr, 0, sizeof(dstaddr));
2568         dstaddr.sin_family = AF_INET;
2569         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2570                                (struct sockaddr *)&dstaddr, 1);
2571         if (rc != 0 || cmid->device == NULL) {
2572                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2573                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2574                        cmid->device, rc);
2575                 rdma_destroy_id(cmid);
2576                 return rc;
2577         }
2578
2579         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2580                 /* don't need device failover */
2581                 rdma_destroy_id(cmid);
2582                 return 0;
2583         }
2584
2585         return 1;
2586 }
2587
2588 int
2589 kiblnd_dev_failover(kib_dev_t *dev)
2590 {
2591         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2592         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2593         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2594         struct rdma_cm_id  *cmid  = NULL;
2595         kib_hca_dev_t      *hdev  = NULL;
2596         kib_hca_dev_t      *old;
2597         struct ib_pd       *pd;
2598         kib_net_t          *net;
2599         struct sockaddr_in  addr;
2600         unsigned long       flags;
2601         int                 rc = 0;
2602         int                 i;
2603
2604         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2605                  dev->ibd_can_failover ||
2606                  dev->ibd_hdev == NULL);
2607
2608         rc = kiblnd_dev_need_failover(dev);
2609         if (rc <= 0)
2610                 goto out;
2611
2612         if (dev->ibd_hdev != NULL &&
2613             dev->ibd_hdev->ibh_cmid != NULL) {
2614                 /* XXX it's not good to close old listener at here,
2615                  * because we can fail to create new listener.
2616                  * But we have to close it now, otherwise rdma_bind_addr
2617                  * will return EADDRINUSE... How crap! */
2618                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2619
2620                 cmid = dev->ibd_hdev->ibh_cmid;
2621                 /* make next schedule of kiblnd_dev_need_failover()
2622                  * return 1 for me */
2623                 dev->ibd_hdev->ibh_cmid  = NULL;
2624                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2625
2626                 rdma_destroy_id(cmid);
2627         }
2628
2629         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2630                                      IB_QPT_RC);
2631         if (IS_ERR(cmid)) {
2632                 rc = PTR_ERR(cmid);
2633                 CERROR("Failed to create cmid for failover: %d\n", rc);
2634                 goto out;
2635         }
2636
2637         memset(&addr, 0, sizeof(addr));
2638         addr.sin_family      = AF_INET;
2639         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2640         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2641
2642         /* Bind to failover device or port */
2643         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2644         if (rc != 0 || cmid->device == NULL) {
2645                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2646                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2647                        cmid->device, rc);
2648                 rdma_destroy_id(cmid);
2649                 goto out;
2650         }
2651
2652         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2653         if (hdev == NULL) {
2654                 CERROR("Failed to allocate kib_hca_dev\n");
2655                 rdma_destroy_id(cmid);
2656                 rc = -ENOMEM;
2657                 goto out;
2658         }
2659
2660         atomic_set(&hdev->ibh_ref, 1);
2661         hdev->ibh_dev   = dev;
2662         hdev->ibh_cmid  = cmid;
2663         hdev->ibh_ibdev = cmid->device;
2664
2665         pd = ib_alloc_pd(cmid->device);
2666         if (IS_ERR(pd)) {
2667                 rc = PTR_ERR(pd);
2668                 CERROR("Can't allocate PD: %d\n", rc);
2669                 goto out;
2670         }
2671
2672         hdev->ibh_pd = pd;
2673
2674         rc = rdma_listen(cmid, 0);
2675         if (rc != 0) {
2676                 CERROR("Can't start new listener: %d\n", rc);
2677                 goto out;
2678         }
2679
2680         rc = kiblnd_hdev_setup_mrs(hdev);
2681         if (rc != 0) {
2682                 CERROR("Can't setup device: %d\n", rc);
2683                 goto out;
2684         }
2685
2686         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2687
2688         old = dev->ibd_hdev;
2689         dev->ibd_hdev = hdev;   /* take over the refcount */
2690         hdev = old;
2691
2692         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2693                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2694                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2695                                             &zombie_tpo);
2696
2697                         if (net->ibn_fmr_ps != NULL) {
2698                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2699                                                         &zombie_fpo);
2700
2701                         } else if (net->ibn_pmr_ps != NULL) {
2702                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2703                                                     pps_poolset, &zombie_ppo);
2704                         }
2705                 }
2706         }
2707
2708         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2709  out:
2710         if (!list_empty(&zombie_tpo))
2711                 kiblnd_destroy_pool_list(&zombie_tpo);
2712         if (!list_empty(&zombie_ppo))
2713                 kiblnd_destroy_pool_list(&zombie_ppo);
2714         if (!list_empty(&zombie_fpo))
2715                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2716         if (hdev != NULL)
2717                 kiblnd_hdev_decref(hdev);
2718
2719         if (rc != 0)
2720                 dev->ibd_failed_failover++;
2721         else
2722                 dev->ibd_failed_failover = 0;
2723
2724         return rc;
2725 }
2726
2727 void
2728 kiblnd_destroy_dev (kib_dev_t *dev)
2729 {
2730         LASSERT (dev->ibd_nnets == 0);
2731         LASSERT(list_empty(&dev->ibd_nets));
2732
2733         list_del(&dev->ibd_fail_list);
2734         list_del(&dev->ibd_list);
2735
2736         if (dev->ibd_hdev != NULL)
2737                 kiblnd_hdev_decref(dev->ibd_hdev);
2738
2739         LIBCFS_FREE(dev, sizeof(*dev));
2740 }
2741
2742 static kib_dev_t *
2743 kiblnd_create_dev(char *ifname)
2744 {
2745         struct net_device *netdev;
2746         kib_dev_t         *dev;
2747         __u32              netmask;
2748         __u32              ip;
2749         int                up;
2750         int                rc;
2751
2752         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2753         if (rc != 0) {
2754                 CERROR("Can't query IPoIB interface %s: %d\n",
2755                        ifname, rc);
2756                 return NULL;
2757         }
2758
2759         if (!up) {
2760                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2761                 return NULL;
2762         }
2763
2764         LIBCFS_ALLOC(dev, sizeof(*dev));
2765         if (dev == NULL)
2766                 return NULL;
2767
2768         memset(dev, 0, sizeof(*dev));
2769 #ifdef HAVE_DEV_GET_BY_NAME_2ARG
2770         netdev = dev_get_by_name(&init_net, ifname);
2771 #else
2772         netdev = dev_get_by_name(ifname);
2773 #endif
2774         if (netdev == NULL) {
2775                 dev->ibd_can_failover = 0;
2776         } else {
2777                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2778                 dev_put(netdev);
2779         }
2780
2781         INIT_LIST_HEAD(&dev->ibd_nets);
2782         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2783         INIT_LIST_HEAD(&dev->ibd_fail_list);
2784         dev->ibd_ifip = ip;
2785         strcpy(&dev->ibd_ifname[0], ifname);
2786
2787         /* initialize the device */
2788         rc = kiblnd_dev_failover(dev);
2789         if (rc != 0) {
2790                 CERROR("Can't initialize device: %d\n", rc);
2791                 LIBCFS_FREE(dev, sizeof(*dev));
2792                 return NULL;
2793         }
2794
2795         list_add_tail(&dev->ibd_list,
2796                           &kiblnd_data.kib_devs);
2797         return dev;
2798 }
2799
2800 static void
2801 kiblnd_base_shutdown(void)
2802 {
2803         struct kib_sched_info   *sched;
2804         int                     i;
2805
2806         LASSERT(list_empty(&kiblnd_data.kib_devs));
2807
2808         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2809                atomic_read(&libcfs_kmemory));
2810
2811         switch (kiblnd_data.kib_init) {
2812         default:
2813                 LBUG();
2814
2815         case IBLND_INIT_ALL:
2816         case IBLND_INIT_DATA:
2817                 LASSERT (kiblnd_data.kib_peers != NULL);
2818                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2819                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2820                 }
2821                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2822                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2823
2824                 /* flag threads to terminate; wake and wait for them to die */
2825                 kiblnd_data.kib_shutdown = 1;
2826
2827                 /* NB: we really want to stop scheduler threads net by net
2828                  * instead of the whole module, this should be improved
2829                  * with dynamic configuration LNet */
2830                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2831                         wake_up_all(&sched->ibs_waitq);
2832
2833                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2834                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2835
2836                 i = 2;
2837                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2838                         i++;
2839                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2840                                "Waiting for %d threads to terminate\n",
2841                                atomic_read(&kiblnd_data.kib_nthreads));
2842                         cfs_pause(cfs_time_seconds(1));
2843                 }
2844
2845                 /* fall through */
2846
2847         case IBLND_INIT_NOTHING:
2848                 break;
2849         }
2850
2851         if (kiblnd_data.kib_peers != NULL) {
2852                 LIBCFS_FREE(kiblnd_data.kib_peers,
2853                             sizeof(struct list_head) *
2854                             kiblnd_data.kib_peer_hash_size);
2855         }
2856
2857         if (kiblnd_data.kib_scheds != NULL)
2858                 cfs_percpt_free(kiblnd_data.kib_scheds);
2859
2860         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2861                atomic_read(&libcfs_kmemory));
2862
2863         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2864         module_put(THIS_MODULE);
2865 }
2866
2867 static void
2868 kiblnd_shutdown (lnet_ni_t *ni)
2869 {
2870         kib_net_t        *net = ni->ni_data;
2871         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2872         int               i;
2873         unsigned long     flags;
2874
2875         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2876
2877         if (net == NULL)
2878                 goto out;
2879
2880         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2881                atomic_read(&libcfs_kmemory));
2882
2883         write_lock_irqsave(g_lock, flags);
2884         net->ibn_shutdown = 1;
2885         write_unlock_irqrestore(g_lock, flags);
2886
2887         switch (net->ibn_init) {
2888         default:
2889                 LBUG();
2890
2891         case IBLND_INIT_ALL:
2892                 /* nuke all existing peers within this net */
2893                 kiblnd_del_peer(ni, LNET_NID_ANY);
2894
2895                 /* Wait for all peer state to clean up */
2896                 i = 2;
2897                 while (atomic_read(&net->ibn_npeers) != 0) {
2898                         i++;
2899                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2900                                "%s: waiting for %d peers to disconnect\n",
2901                                libcfs_nid2str(ni->ni_nid),
2902                                atomic_read(&net->ibn_npeers));
2903                         cfs_pause(cfs_time_seconds(1));
2904                 }
2905
2906                 kiblnd_net_fini_pools(net);
2907
2908                 write_lock_irqsave(g_lock, flags);
2909                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2910                 net->ibn_dev->ibd_nnets--;
2911                 list_del(&net->ibn_list);
2912                 write_unlock_irqrestore(g_lock, flags);
2913
2914                 /* fall through */
2915
2916         case IBLND_INIT_NOTHING:
2917                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2918
2919                 if (net->ibn_dev != NULL &&
2920                     net->ibn_dev->ibd_nnets == 0)
2921                         kiblnd_destroy_dev(net->ibn_dev);
2922
2923                 break;
2924         }
2925
2926         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2927                atomic_read(&libcfs_kmemory));
2928
2929         net->ibn_init = IBLND_INIT_NOTHING;
2930         ni->ni_data = NULL;
2931
2932         LIBCFS_FREE(net, sizeof(*net));
2933
2934 out:
2935         if (list_empty(&kiblnd_data.kib_devs))
2936                 kiblnd_base_shutdown();
2937         return;
2938 }
2939
2940 static int
2941 kiblnd_base_startup(void)
2942 {
2943         struct kib_sched_info   *sched;
2944         int                     rc;
2945         int                     i;
2946
2947         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2948
2949         try_module_get(THIS_MODULE);
2950         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2951
2952         rwlock_init(&kiblnd_data.kib_global_lock);
2953
2954         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2955         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2956
2957         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2958         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2959                      sizeof(struct list_head) *
2960                      kiblnd_data.kib_peer_hash_size);
2961         if (kiblnd_data.kib_peers == NULL)
2962                 goto failed;
2963
2964         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2965                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2966
2967         spin_lock_init(&kiblnd_data.kib_connd_lock);
2968         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2969         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2970         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2971         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2972
2973         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2974                                                   sizeof(*sched));
2975         if (kiblnd_data.kib_scheds == NULL)
2976                 goto failed;
2977
2978         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2979                 int     nthrs;
2980
2981                 spin_lock_init(&sched->ibs_lock);
2982                 INIT_LIST_HEAD(&sched->ibs_conns);
2983                 init_waitqueue_head(&sched->ibs_waitq);
2984
2985                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2986                 if (*kiblnd_tunables.kib_nscheds > 0) {
2987                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2988                 } else {
2989                         /* max to half of CPUs, another half is reserved for
2990                          * upper layer modules */
2991                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2992                 }
2993
2994                 sched->ibs_nthreads_max = nthrs;
2995                 sched->ibs_cpt = i;
2996         }
2997
2998         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2999
3000         /* lists/ptrs/locks initialised */
3001         kiblnd_data.kib_init = IBLND_INIT_DATA;
3002         /*****************************************************/
3003
3004         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3005         if (rc != 0) {
3006                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3007                 goto failed;
3008         }
3009
3010         if (*kiblnd_tunables.kib_dev_failover != 0)
3011                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3012                                          "kiblnd_failover");
3013
3014         if (rc != 0) {
3015                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3016                 goto failed;
3017         }
3018
3019         /* flag everything initialised */
3020         kiblnd_data.kib_init = IBLND_INIT_ALL;
3021         /*****************************************************/
3022
3023         return 0;
3024
3025  failed:
3026         kiblnd_base_shutdown();
3027         return -ENETDOWN;
3028 }
3029
3030 static int
3031 kiblnd_start_schedulers(struct kib_sched_info *sched)
3032 {
3033         int     rc = 0;
3034         int     nthrs;
3035         int     i;
3036
3037         if (sched->ibs_nthreads == 0) {
3038                 if (*kiblnd_tunables.kib_nscheds > 0) {
3039                         nthrs = sched->ibs_nthreads_max;
3040                 } else {
3041                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3042                                                sched->ibs_cpt);
3043                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3044                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3045                 }
3046         } else {
3047                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3048                 /* increase one thread if there is new interface */
3049                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3050         }
3051
3052         for (i = 0; i < nthrs; i++) {
3053                 long    id;
3054                 char    name[20];
3055                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3056                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3057                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3058                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3059                 if (rc == 0)
3060                         continue;
3061
3062                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3063                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3064                 break;
3065         }
3066
3067         sched->ibs_nthreads += i;
3068         return rc;
3069 }
3070
3071 static int
3072 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3073 {
3074         int     cpt;
3075         int     rc;
3076         int     i;
3077
3078         for (i = 0; i < ncpts; i++) {
3079                 struct kib_sched_info *sched;
3080
3081                 cpt = (cpts == NULL) ? i : cpts[i];
3082                 sched = kiblnd_data.kib_scheds[cpt];
3083
3084                 if (!newdev && sched->ibs_nthreads > 0)
3085                         continue;
3086
3087                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3088                 if (rc != 0) {
3089                         CERROR("Failed to start scheduler threads for %s\n",
3090                                dev->ibd_ifname);
3091                         return rc;
3092                 }
3093         }
3094         return 0;
3095 }
3096
3097 static kib_dev_t *
3098 kiblnd_dev_search(char *ifname)
3099 {
3100         kib_dev_t       *alias = NULL;
3101         kib_dev_t       *dev;
3102         char            *colon;
3103         char            *colon2;
3104
3105         colon = strchr(ifname, ':');
3106         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3107                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3108                         return dev;
3109
3110                 if (alias != NULL)
3111                         continue;
3112
3113                 colon2 = strchr(dev->ibd_ifname, ':');
3114                 if (colon != NULL)
3115                         *colon = 0;
3116                 if (colon2 != NULL)
3117                         *colon2 = 0;
3118
3119                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3120                         alias = dev;
3121
3122                 if (colon != NULL)
3123                         *colon = ':';
3124                 if (colon2 != NULL)
3125                         *colon2 = ':';
3126         }
3127         return alias;
3128 }
3129
3130 static int
3131 kiblnd_startup (lnet_ni_t *ni)
3132 {
3133         char                     *ifname;
3134         kib_dev_t                *ibdev = NULL;
3135         kib_net_t                *net;
3136         struct timeval            tv;
3137         unsigned long             flags;
3138         int                       rc;
3139         int                       newdev;
3140
3141         LASSERT (ni->ni_lnd == &the_o2iblnd);
3142
3143         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3144                 rc = kiblnd_base_startup();
3145                 if (rc != 0)
3146                         return rc;
3147         }
3148
3149         LIBCFS_ALLOC(net, sizeof(*net));
3150         ni->ni_data = net;
3151         if (net == NULL)
3152                 goto failed;
3153
3154         memset(net, 0, sizeof(*net));
3155
3156         do_gettimeofday(&tv);
3157         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3158
3159         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3160         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3161         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3162         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3163
3164         if (ni->ni_interfaces[0] != NULL) {
3165                 /* Use the IPoIB interface specified in 'networks=' */
3166
3167                 CLASSERT (LNET_MAX_INTERFACES > 1);
3168                 if (ni->ni_interfaces[1] != NULL) {
3169                         CERROR("Multiple interfaces not supported\n");
3170                         goto failed;
3171                 }
3172
3173                 ifname = ni->ni_interfaces[0];
3174         } else {
3175                 ifname = *kiblnd_tunables.kib_default_ipif;
3176         }
3177
3178         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3179                 CERROR("IPoIB interface name too long: %s\n", ifname);
3180                 goto failed;
3181         }
3182
3183         ibdev = kiblnd_dev_search(ifname);
3184
3185         newdev = ibdev == NULL;
3186         /* hmm...create kib_dev even for alias */
3187         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3188                 ibdev = kiblnd_create_dev(ifname);
3189
3190         if (ibdev == NULL)
3191                 goto failed;
3192
3193         net->ibn_dev = ibdev;
3194         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3195
3196         rc = kiblnd_dev_start_threads(ibdev, newdev,
3197                                       ni->ni_cpts, ni->ni_ncpts);
3198         if (rc != 0)
3199                 goto failed;
3200
3201         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3202         if (rc != 0) {
3203                 CERROR("Failed to initialize NI pools: %d\n", rc);
3204                 goto failed;
3205         }
3206
3207         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3208         ibdev->ibd_nnets++;
3209         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3210         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3211
3212         net->ibn_init = IBLND_INIT_ALL;
3213
3214         return 0;
3215
3216 failed:
3217         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3218                 kiblnd_destroy_dev(ibdev);
3219
3220         kiblnd_shutdown(ni);
3221
3222         CDEBUG(D_NET, "kiblnd_startup failed\n");
3223         return -ENETDOWN;
3224 }
3225
3226 static lnd_t the_o2iblnd = {
3227         .lnd_type       = O2IBLND,
3228         .lnd_startup    = kiblnd_startup,
3229         .lnd_shutdown   = kiblnd_shutdown,
3230         .lnd_ctl        = kiblnd_ctl,
3231         .lnd_query      = kiblnd_query,
3232         .lnd_send       = kiblnd_send,
3233         .lnd_recv       = kiblnd_recv,
3234 };
3235
3236 static void __exit
3237 kiblnd_module_fini (void)
3238 {
3239         lnet_unregister_lnd(&the_o2iblnd);
3240         kiblnd_tunables_fini();
3241 }
3242
3243 static int __init
3244 kiblnd_module_init (void)
3245 {
3246         int    rc;
3247
3248         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3249         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3250                   <= IBLND_MSG_SIZE);
3251         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3252                   <= IBLND_MSG_SIZE);
3253
3254         rc = kiblnd_tunables_init();
3255         if (rc != 0)
3256                 return rc;
3257
3258         lnet_register_lnd(&the_o2iblnd);
3259
3260         return 0;
3261 }
3262
3263 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3264 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3265 MODULE_LICENSE("GPL");
3266
3267 module_init(kiblnd_module_init);
3268 module_exit(kiblnd_module_fini);