drivers/staging/lustre/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 #include <asm/div64.h>
  43
  44 lnd_t the_o2iblnd = {
  45         .lnd_type       = O2IBLND,
  46         .lnd_startup    = kiblnd_startup,
  47         .lnd_shutdown   = kiblnd_shutdown,
  48         .lnd_ctl        = kiblnd_ctl,
  49         .lnd_query      = kiblnd_query,
  50         .lnd_send       = kiblnd_send,
  51         .lnd_recv       = kiblnd_recv,
  52 };
  53
  54 kib_data_t            kiblnd_data;
  55
  56 __u32
  57 kiblnd_cksum (void *ptr, int nob)
  58 {
  59         char  *c  = ptr;
  60         __u32  sum = 0;
  61
  62         while (nob-- > 0)
  63                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  64
  65         /* ensure I don't return 0 (== no checksum) */
  66         return (sum == 0) ? 1 : sum;
  67 }
  68
  69 static char *
  70 kiblnd_msgtype2str(int type)
  71 {
  72         switch (type) {
  73         case IBLND_MSG_CONNREQ:
  74                 return "CONNREQ";
  75
  76         case IBLND_MSG_CONNACK:
  77                 return "CONNACK";
  78
  79         case IBLND_MSG_NOOP:
  80                 return "NOOP";
  81
  82         case IBLND_MSG_IMMEDIATE:
  83                 return "IMMEDIATE";
  84
  85         case IBLND_MSG_PUT_REQ:
  86                 return "PUT_REQ";
  87
  88         case IBLND_MSG_PUT_NAK:
  89                 return "PUT_NAK";
  90
  91         case IBLND_MSG_PUT_ACK:
  92                 return "PUT_ACK";
  93
  94         case IBLND_MSG_PUT_DONE:
  95                 return "PUT_DONE";
  96
  97         case IBLND_MSG_GET_REQ:
  98                 return "GET_REQ";
  99
 100         case IBLND_MSG_GET_DONE:
 101                 return "GET_DONE";
 102
 103         default:
 104                 return "???";
 105         }
 106 }
 107
 108 static int
 109 kiblnd_msgtype2size(int type)
 110 {
 111         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 112
 113         switch (type) {
 114         case IBLND_MSG_CONNREQ:
 115         case IBLND_MSG_CONNACK:
 116                 return hdr_size + sizeof(kib_connparams_t);
 117
 118         case IBLND_MSG_NOOP:
 119                 return hdr_size;
 120
 121         case IBLND_MSG_IMMEDIATE:
 122                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 123
 124         case IBLND_MSG_PUT_REQ:
 125                 return hdr_size + sizeof(kib_putreq_msg_t);
 126
 127         case IBLND_MSG_PUT_ACK:
 128                 return hdr_size + sizeof(kib_putack_msg_t);
 129
 130         case IBLND_MSG_GET_REQ:
 131                 return hdr_size + sizeof(kib_get_msg_t);
 132
 133         case IBLND_MSG_PUT_NAK:
 134         case IBLND_MSG_PUT_DONE:
 135         case IBLND_MSG_GET_DONE:
 136                 return hdr_size + sizeof(kib_completion_msg_t);
 137         default:
 138                 return -1;
 139         }
 140 }
 141
 142 static int
 143 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 144 {
 145         kib_rdma_desc_t   *rd;
 146         int             nob;
 147         int             n;
 148         int             i;
 149
 150         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 151                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 152
 153         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 154                               &msg->ibm_u.get.ibgm_rd :
 155                               &msg->ibm_u.putack.ibpam_rd;
 156
 157         if (flip) {
 158                 __swab32s(&rd->rd_key);
 159                 __swab32s(&rd->rd_nfrags);
 160         }
 161
 162         n = rd->rd_nfrags;
 163
 164         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 165                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 166                        n, IBLND_MAX_RDMA_FRAGS);
 167                 return 1;
 168         }
 169
 170         nob = offsetof (kib_msg_t, ibm_u) +
 171               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 172
 173         if (msg->ibm_nob < nob) {
 174                 CERROR("Short %s: %d(%d)\n",
 175                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 176                 return 1;
 177         }
 178
 179         if (!flip)
 180                 return 0;
 181
 182         for (i = 0; i < n; i++) {
 183                 __swab32s(&rd->rd_frags[i].rf_nob);
 184                 __swab64s(&rd->rd_frags[i].rf_addr);
 185         }
 186
 187         return 0;
 188 }
 189
 190 void
 191 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
 192                  int credits, lnet_nid_t dstnid, __u64 dststamp)
 193 {
 194         kib_net_t *net = ni->ni_data;
 195
 196         /* CAVEAT EMPTOR! all message fields not set here should have been
 197          * initialised previously. */
 198         msg->ibm_magic    = IBLND_MSG_MAGIC;
 199         msg->ibm_version  = version;
 200         /*   ibm_type */
 201         msg->ibm_credits  = credits;
 202         /*   ibm_nob */
 203         msg->ibm_cksum    = 0;
 204         msg->ibm_srcnid   = ni->ni_nid;
 205         msg->ibm_srcstamp = net->ibn_incarnation;
 206         msg->ibm_dstnid   = dstnid;
 207         msg->ibm_dststamp = dststamp;
 208
 209         if (*kiblnd_tunables.kib_cksum) {
 210                 /* NB ibm_cksum zero while computing cksum */
 211                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 212         }
 213 }
 214
 215 int
 216 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 217 {
 218         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 219         __u32     msg_cksum;
 220         __u16     version;
 221         int       msg_nob;
 222         int       flip;
 223
 224         /* 6 bytes are enough to have received magic + version */
 225         if (nob < 6) {
 226                 CERROR("Short message: %d\n", nob);
 227                 return -EPROTO;
 228         }
 229
 230         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 231                 flip = 0;
 232         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 233                 flip = 1;
 234         } else {
 235                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 236                 return -EPROTO;
 237         }
 238
 239         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 240         if (version != IBLND_MSG_VERSION &&
 241             version != IBLND_MSG_VERSION_1) {
 242                 CERROR("Bad version: %x\n", version);
 243                 return -EPROTO;
 244         }
 245
 246         if (nob < hdr_size) {
 247                 CERROR("Short message: %d\n", nob);
 248                 return -EPROTO;
 249         }
 250
 251         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 252         if (msg_nob > nob) {
 253                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 254                 return -EPROTO;
 255         }
 256
 257         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 258          * gets flipped */
 259         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 260         msg->ibm_cksum = 0;
 261         if (msg_cksum != 0 &&
 262             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 263                 CERROR("Bad checksum\n");
 264                 return -EPROTO;
 265         }
 266
 267         msg->ibm_cksum = msg_cksum;
 268
 269         if (flip) {
 270                 /* leave magic unflipped as a clue to peer endianness */
 271                 msg->ibm_version = version;
 272                 CLASSERT (sizeof(msg->ibm_type) == 1);
 273                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 274                 msg->ibm_nob     = msg_nob;
 275                 __swab64s(&msg->ibm_srcnid);
 276                 __swab64s(&msg->ibm_srcstamp);
 277                 __swab64s(&msg->ibm_dstnid);
 278                 __swab64s(&msg->ibm_dststamp);
 279         }
 280
 281         if (msg->ibm_srcnid == LNET_NID_ANY) {
 282                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 283                 return -EPROTO;
 284         }
 285
 286         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 287                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 288                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 289                 return -EPROTO;
 290         }
 291
 292         switch (msg->ibm_type) {
 293         default:
 294                 CERROR("Unknown message type %x\n", msg->ibm_type);
 295                 return -EPROTO;
 296
 297         case IBLND_MSG_NOOP:
 298         case IBLND_MSG_IMMEDIATE:
 299         case IBLND_MSG_PUT_REQ:
 300                 break;
 301
 302         case IBLND_MSG_PUT_ACK:
 303         case IBLND_MSG_GET_REQ:
 304                 if (kiblnd_unpack_rd(msg, flip))
 305                         return -EPROTO;
 306                 break;
 307
 308         case IBLND_MSG_PUT_NAK:
 309         case IBLND_MSG_PUT_DONE:
 310         case IBLND_MSG_GET_DONE:
 311                 if (flip)
 312                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 313                 break;
 314
 315         case IBLND_MSG_CONNREQ:
 316         case IBLND_MSG_CONNACK:
 317                 if (flip) {
 318                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 319                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 320                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 321                 }
 322                 break;
 323         }
 324         return 0;
 325 }
 326
 327 int
 328 kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 329 {
 330         kib_peer_t      *peer;
 331         kib_net_t       *net = ni->ni_data;
 332         int             cpt = lnet_cpt_of_nid(nid);
 333         unsigned long   flags;
 334
 335         LASSERT(net != NULL);
 336         LASSERT(nid != LNET_NID_ANY);
 337
 338         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 339         if (peer == NULL) {
 340                 CERROR("Cannot allocate peer\n");
 341                 return -ENOMEM;
 342         }
 343
 344         memset(peer, 0, sizeof(*peer));  /* zero flags etc */
 345
 346         peer->ibp_ni = ni;
 347         peer->ibp_nid = nid;
 348         peer->ibp_error = 0;
 349         peer->ibp_last_alive = 0;
 350         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 351
 352         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 353         INIT_LIST_HEAD(&peer->ibp_conns);
 354         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 355
 356         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 357
 358         /* always called with a ref on ni, which prevents ni being shutdown */
 359         LASSERT (net->ibn_shutdown == 0);
 360
 361         /* npeers only grows with the global lock held */
 362         atomic_inc(&net->ibn_npeers);
 363
 364         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 365
 366         *peerp = peer;
 367         return 0;
 368 }
 369
 370 void
 371 kiblnd_destroy_peer (kib_peer_t *peer)
 372 {
 373         kib_net_t *net = peer->ibp_ni->ni_data;
 374
 375         LASSERT (net != NULL);
 376         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 377         LASSERT (!kiblnd_peer_active(peer));
 378         LASSERT (peer->ibp_connecting == 0);
 379         LASSERT (peer->ibp_accepting == 0);
 380         LASSERT (list_empty(&peer->ibp_conns));
 381         LASSERT (list_empty(&peer->ibp_tx_queue));
 382
 383         LIBCFS_FREE(peer, sizeof(*peer));
 384
 385         /* NB a peer's connections keep a reference on their peer until
 386          * they are destroyed, so we can be assured that _all_ state to do
 387          * with this peer has been cleaned up when its refcount drops to
 388          * zero. */
 389         atomic_dec(&net->ibn_npeers);
 390 }
 391
 392 kib_peer_t *
 393 kiblnd_find_peer_locked (lnet_nid_t nid)
 394 {
 395         /* the caller is responsible for accounting the additional reference
 396          * that this creates */
 397         struct list_head       *peer_list = kiblnd_nid2peerlist(nid);
 398         struct list_head       *tmp;
 399         kib_peer_t       *peer;
 400
 401         list_for_each (tmp, peer_list) {
 402
 403                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 404
 405                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 406                          peer->ibp_accepting > 0 ||
 407                          !list_empty(&peer->ibp_conns));  /* active conn */
 408
 409                 if (peer->ibp_nid != nid)
 410                         continue;
 411
 412                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 413                        peer, libcfs_nid2str(nid),
 414                        atomic_read(&peer->ibp_refcount),
 415                        peer->ibp_version);
 416                 return peer;
 417         }
 418         return NULL;
 419 }
 420
 421 void
 422 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 423 {
 424         LASSERT (list_empty(&peer->ibp_conns));
 425
 426         LASSERT (kiblnd_peer_active(peer));
 427         list_del_init(&peer->ibp_list);
 428         /* lose peerlist's ref */
 429         kiblnd_peer_decref(peer);
 430 }
 431
 432 int
 433 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 434                       lnet_nid_t *nidp, int *count)
 435 {
 436         kib_peer_t          *peer;
 437         struct list_head            *ptmp;
 438         int                 i;
 439         unsigned long     flags;
 440
 441         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 442
 443         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 444
 445                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 446
 447                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 448                         LASSERT (peer->ibp_connecting > 0 ||
 449                                  peer->ibp_accepting > 0 ||
 450                                  !list_empty(&peer->ibp_conns));
 451
 452                         if (peer->ibp_ni != ni)
 453                                 continue;
 454
 455                         if (index-- > 0)
 456                                 continue;
 457
 458                         *nidp = peer->ibp_nid;
 459                         *count = atomic_read(&peer->ibp_refcount);
 460
 461                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 462                                                flags);
 463                         return 0;
 464                 }
 465         }
 466
 467         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 468         return -ENOENT;
 469 }
 470
 471 void
 472 kiblnd_del_peer_locked (kib_peer_t *peer)
 473 {
 474         struct list_head           *ctmp;
 475         struct list_head           *cnxt;
 476         kib_conn_t         *conn;
 477
 478         if (list_empty(&peer->ibp_conns)) {
 479                 kiblnd_unlink_peer_locked(peer);
 480         } else {
 481                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 482                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 483
 484                         kiblnd_close_conn_locked(conn, 0);
 485                 }
 486                 /* NB closing peer's last conn unlinked it. */
 487         }
 488         /* NB peer now unlinked; might even be freed if the peer table had the
 489          * last ref on it. */
 490 }
 491
 492 int
 493 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 494 {
 495         LIST_HEAD        (zombies);
 496         struct list_head            *ptmp;
 497         struct list_head            *pnxt;
 498         kib_peer_t          *peer;
 499         int                 lo;
 500         int                 hi;
 501         int                 i;
 502         unsigned long     flags;
 503         int                 rc = -ENOENT;
 504
 505         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 506
 507         if (nid != LNET_NID_ANY) {
 508                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 509         } else {
 510                 lo = 0;
 511                 hi = kiblnd_data.kib_peer_hash_size - 1;
 512         }
 513
 514         for (i = lo; i <= hi; i++) {
 515                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 516                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 517                         LASSERT (peer->ibp_connecting > 0 ||
 518                                  peer->ibp_accepting > 0 ||
 519                                  !list_empty(&peer->ibp_conns));
 520
 521                         if (peer->ibp_ni != ni)
 522                                 continue;
 523
 524                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 525                                 continue;
 526
 527                         if (!list_empty(&peer->ibp_tx_queue)) {
 528                                 LASSERT (list_empty(&peer->ibp_conns));
 529
 530                                 list_splice_init(&peer->ibp_tx_queue,
 531                                                      &zombies);
 532                         }
 533
 534                         kiblnd_del_peer_locked(peer);
 535                         rc = 0;  /* matched something */
 536                 }
 537         }
 538
 539         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 540
 541         kiblnd_txlist_done(ni, &zombies, -EIO);
 542
 543         return rc;
 544 }
 545
 546 kib_conn_t *
 547 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 548 {
 549         kib_peer_t          *peer;
 550         struct list_head            *ptmp;
 551         kib_conn_t          *conn;
 552         struct list_head            *ctmp;
 553         int                 i;
 554         unsigned long     flags;
 555
 556         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 557
 558         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 559                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 560
 561                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 562                         LASSERT (peer->ibp_connecting > 0 ||
 563                                  peer->ibp_accepting > 0 ||
 564                                  !list_empty(&peer->ibp_conns));
 565
 566                         if (peer->ibp_ni != ni)
 567                                 continue;
 568
 569                         list_for_each (ctmp, &peer->ibp_conns) {
 570                                 if (index-- > 0)
 571                                         continue;
 572
 573                                 conn = list_entry(ctmp, kib_conn_t,
 574                                                       ibc_list);
 575                                 kiblnd_conn_addref(conn);
 576                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 577                                                        flags);
 578                                 return conn;
 579                         }
 580                 }
 581         }
 582
 583         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 584         return NULL;
 585 }
 586
 587 void
 588 kiblnd_debug_rx (kib_rx_t *rx)
 589 {
 590         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 591                rx, rx->rx_status, rx->rx_msg->ibm_type,
 592                rx->rx_msg->ibm_credits);
 593 }
 594
 595 void
 596 kiblnd_debug_tx (kib_tx_t *tx)
 597 {
 598         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 599                "cookie "LPX64" msg %s%s type %x cred %d\n",
 600                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 601                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 602                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 603                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 604                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 605 }
 606
 607 void
 608 kiblnd_debug_conn (kib_conn_t *conn)
 609 {
 610         struct list_head        *tmp;
 611         int             i;
 612
 613         spin_lock(&conn->ibc_lock);
 614
 615         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s: \n",
 616                atomic_read(&conn->ibc_refcount), conn,
 617                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 618         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d r_cred %d\n",
 619                conn->ibc_state, conn->ibc_noops_posted,
 620                conn->ibc_nsends_posted, conn->ibc_credits,
 621                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 622         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 623
 624         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 625         list_for_each(tmp, &conn->ibc_early_rxs)
 626                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 627
 628         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 629         list_for_each(tmp, &conn->ibc_tx_noops)
 630                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 631
 632         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 633         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 634                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 635
 636         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 637         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 638                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 639
 640         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 641         list_for_each(tmp, &conn->ibc_tx_queue)
 642                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 643
 644         CDEBUG(D_CONSOLE, "   active_txs:\n");
 645         list_for_each(tmp, &conn->ibc_active_txs)
 646                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 647
 648         CDEBUG(D_CONSOLE, "   rxs:\n");
 649         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++)
 650                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 651
 652         spin_unlock(&conn->ibc_lock);
 653 }
 654
 655 int
 656 kiblnd_translate_mtu(int value)
 657 {
 658         switch (value) {
 659         default:
 660                 return -1;
 661         case 0:
 662                 return 0;
 663         case 256:
 664                 return IB_MTU_256;
 665         case 512:
 666                 return IB_MTU_512;
 667         case 1024:
 668                 return IB_MTU_1024;
 669         case 2048:
 670                 return IB_MTU_2048;
 671         case 4096:
 672                 return IB_MTU_4096;
 673         }
 674 }
 675
 676 static void
 677 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 678 {
 679         int        mtu;
 680
 681         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 682         if (cmid->route.path_rec == NULL)
 683                 return;
 684
 685         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 686         LASSERT (mtu >= 0);
 687         if (mtu != 0)
 688                 cmid->route.path_rec->mtu = mtu;
 689 }
 690
 691 static int
 692 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 693 {
 694         cpumask_t       *mask;
 695         int             vectors;
 696         int             off;
 697         int             i;
 698         lnet_nid_t      nid = conn->ibc_peer->ibp_nid;
 699
 700         vectors = conn->ibc_cmid->device->num_comp_vectors;
 701         if (vectors <= 1)
 702                 return 0;
 703
 704         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 705
 706         /* hash NID to CPU id in this partition... */
 707         off = do_div(nid, cpus_weight(*mask));
 708         for_each_cpu_mask(i, *mask) {
 709                 if (off-- == 0)
 710                         return i % vectors;
 711         }
 712
 713         LBUG();
 714         return 1;
 715 }
 716
 717 kib_conn_t *
 718 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 719                    int state, int version)
 720 {
 721         /* CAVEAT EMPTOR:
 722          * If the new conn is created successfully it takes over the caller's
 723          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 724          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 725          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 726          * to destroy 'cmid' here since I'm called from the CM which still has
 727          * its ref on 'cmid'). */
 728         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
 729         kib_net_t             *net = peer->ibp_ni->ni_data;
 730         kib_dev_t             *dev;
 731         struct ib_qp_init_attr *init_qp_attr;
 732         struct kib_sched_info   *sched;
 733         kib_conn_t              *conn;
 734         struct ib_cq            *cq;
 735         unsigned long           flags;
 736         int                     cpt;
 737         int                     rc;
 738         int                     i;
 739
 740         LASSERT(net != NULL);
 741         LASSERT(!in_interrupt());
 742
 743         dev = net->ibn_dev;
 744
 745         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 746         sched = kiblnd_data.kib_scheds[cpt];
 747
 748         LASSERT(sched->ibs_nthreads > 0);
 749
 750         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 751                          sizeof(*init_qp_attr));
 752         if (init_qp_attr == NULL) {
 753                 CERROR("Can't allocate qp_attr for %s\n",
 754                        libcfs_nid2str(peer->ibp_nid));
 755                 goto failed_0;
 756         }
 757
 758         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 759         if (conn == NULL) {
 760                 CERROR("Can't allocate connection for %s\n",
 761                        libcfs_nid2str(peer->ibp_nid));
 762                 goto failed_1;
 763         }
 764
 765         conn->ibc_state = IBLND_CONN_INIT;
 766         conn->ibc_version = version;
 767         conn->ibc_peer = peer;            /* I take the caller's ref */
 768         cmid->context = conn;              /* for future CM callbacks */
 769         conn->ibc_cmid = cmid;
 770
 771         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 772         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 773         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 774         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 775         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 776         INIT_LIST_HEAD(&conn->ibc_active_txs);
 777         spin_lock_init(&conn->ibc_lock);
 778
 779         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 780                          sizeof(*conn->ibc_connvars));
 781         if (conn->ibc_connvars == NULL) {
 782                 CERROR("Can't allocate in-progress connection state\n");
 783                 goto failed_2;
 784         }
 785
 786         write_lock_irqsave(glock, flags);
 787         if (dev->ibd_failover) {
 788                 write_unlock_irqrestore(glock, flags);
 789                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 790                 goto failed_2;
 791         }
 792
 793         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 794                 /* wakeup failover thread and teardown connection */
 795                 if (kiblnd_dev_can_failover(dev)) {
 796                         list_add_tail(&dev->ibd_fail_list,
 797                                       &kiblnd_data.kib_failed_devs);
 798                         wake_up(&kiblnd_data.kib_failover_waitq);
 799                 }
 800
 801                 write_unlock_irqrestore(glock, flags);
 802                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 803                        cmid->device->name, dev->ibd_ifname);
 804                 goto failed_2;
 805         }
 806
 807         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 808         conn->ibc_hdev = dev->ibd_hdev;
 809
 810         kiblnd_setup_mtu_locked(cmid);
 811
 812         write_unlock_irqrestore(glock, flags);
 813
 814         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 815                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 816         if (conn->ibc_rxs == NULL) {
 817                 CERROR("Cannot allocate RX buffers\n");
 818                 goto failed_2;
 819         }
 820
 821         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 822                                 IBLND_RX_MSG_PAGES(version));
 823         if (rc != 0)
 824                 goto failed_2;
 825
 826         kiblnd_map_rx_descs(conn);
 827
 828         cq = ib_create_cq(cmid->device,
 829                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 830                           IBLND_CQ_ENTRIES(version),
 831                           kiblnd_get_completion_vector(conn, cpt));
 832         if (IS_ERR(cq)) {
 833                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 834                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 835                 goto failed_2;
 836         }
 837
 838         conn->ibc_cq = cq;
 839
 840         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 841         if (rc != 0) {
 842                 CERROR("Can't request completion notificiation: %d\n", rc);
 843                 goto failed_2;
 844         }
 845
 846         init_qp_attr->event_handler = kiblnd_qp_event;
 847         init_qp_attr->qp_context = conn;
 848         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 849         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 850         init_qp_attr->cap.max_send_sge = 1;
 851         init_qp_attr->cap.max_recv_sge = 1;
 852         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 853         init_qp_attr->qp_type = IB_QPT_RC;
 854         init_qp_attr->send_cq = cq;
 855         init_qp_attr->recv_cq = cq;
 856
 857         conn->ibc_sched = sched;
 858
 859         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 860         if (rc != 0) {
 861                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 862                        rc, init_qp_attr->cap.max_send_wr,
 863                        init_qp_attr->cap.max_recv_wr);
 864                 goto failed_2;
 865         }
 866
 867         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 868
 869         /* 1 ref for caller and each rxmsg */
 870         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 871         conn->ibc_nrx = IBLND_RX_MSGS(version);
 872
 873         /* post receives */
 874         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 875                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 876                                     IBLND_POSTRX_NO_CREDIT);
 877                 if (rc != 0) {
 878                         CERROR("Can't post rxmsg: %d\n", rc);
 879
 880                         /* Make posted receives complete */
 881                         kiblnd_abort_receives(conn);
 882
 883                         /* correct # of posted buffers
 884                          * NB locking needed now I'm racing with completion */
 885                         spin_lock_irqsave(&sched->ibs_lock, flags);
 886                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 887                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 888
 889                         /* cmid will be destroyed by CM(ofed) after cm_callback
 890                          * returned, so we can't refer it anymore
 891                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 892                         rdma_destroy_qp(conn->ibc_cmid);
 893                         conn->ibc_cmid = NULL;
 894
 895                         /* Drop my own and unused rxbuffer refcounts */
 896                         while (i++ <= IBLND_RX_MSGS(version))
 897                                 kiblnd_conn_decref(conn);
 898
 899                         return NULL;
 900                 }
 901         }
 902
 903         /* Init successful! */
 904         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 905                  state == IBLND_CONN_PASSIVE_WAIT);
 906         conn->ibc_state = state;
 907
 908         /* 1 more conn */
 909         atomic_inc(&net->ibn_nconns);
 910         return conn;
 911
 912  failed_2:
 913         kiblnd_destroy_conn(conn);
 914  failed_1:
 915         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 916  failed_0:
 917         return NULL;
 918 }
 919
 920 void
 921 kiblnd_destroy_conn (kib_conn_t *conn)
 922 {
 923         struct rdma_cm_id *cmid = conn->ibc_cmid;
 924         kib_peer_t      *peer = conn->ibc_peer;
 925         int             rc;
 926
 927         LASSERT (!in_interrupt());
 928         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 929         LASSERT (list_empty(&conn->ibc_early_rxs));
 930         LASSERT (list_empty(&conn->ibc_tx_noops));
 931         LASSERT (list_empty(&conn->ibc_tx_queue));
 932         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 933         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 934         LASSERT (list_empty(&conn->ibc_active_txs));
 935         LASSERT (conn->ibc_noops_posted == 0);
 936         LASSERT (conn->ibc_nsends_posted == 0);
 937
 938         switch (conn->ibc_state) {
 939         default:
 940                 /* conn must be completely disengaged from the network */
 941                 LBUG();
 942
 943         case IBLND_CONN_DISCONNECTED:
 944                 /* connvars should have been freed already */
 945                 LASSERT (conn->ibc_connvars == NULL);
 946                 break;
 947
 948         case IBLND_CONN_INIT:
 949                 break;
 950         }
 951
 952         /* conn->ibc_cmid might be destroyed by CM already */
 953         if (cmid != NULL && cmid->qp != NULL)
 954                 rdma_destroy_qp(cmid);
 955
 956         if (conn->ibc_cq != NULL) {
 957                 rc = ib_destroy_cq(conn->ibc_cq);
 958                 if (rc != 0)
 959                         CWARN("Error destroying CQ: %d\n", rc);
 960         }
 961
 962         if (conn->ibc_rx_pages != NULL)
 963                 kiblnd_unmap_rx_descs(conn);
 964
 965         if (conn->ibc_rxs != NULL) {
 966                 LIBCFS_FREE(conn->ibc_rxs,
 967                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 968         }
 969
 970         if (conn->ibc_connvars != NULL)
 971                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 972
 973         if (conn->ibc_hdev != NULL)
 974                 kiblnd_hdev_decref(conn->ibc_hdev);
 975
 976         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 977         if (conn->ibc_state != IBLND_CONN_INIT) {
 978                 kib_net_t *net = peer->ibp_ni->ni_data;
 979
 980                 kiblnd_peer_decref(peer);
 981                 rdma_destroy_id(cmid);
 982                 atomic_dec(&net->ibn_nconns);
 983         }
 984
 985         LIBCFS_FREE(conn, sizeof(*conn));
 986 }
 987
 988 int
 989 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 990 {
 991         kib_conn_t           *conn;
 992         struct list_head             *ctmp;
 993         struct list_head             *cnxt;
 994         int                  count = 0;
 995
 996         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 997                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 998
 999                 CDEBUG(D_NET, "Closing conn -> %s, "
1000                               "version: %x, reason: %d\n",
1001                        libcfs_nid2str(peer->ibp_nid),
1002                        conn->ibc_version, why);
1003
1004                 kiblnd_close_conn_locked(conn, why);
1005                 count++;
1006         }
1007
1008         return count;
1009 }
1010
1011 int
1012 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1013                                  int version, __u64 incarnation)
1014 {
1015         kib_conn_t           *conn;
1016         struct list_head             *ctmp;
1017         struct list_head             *cnxt;
1018         int                  count = 0;
1019
1020         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
1021                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1022
1023                 if (conn->ibc_version     == version &&
1024                     conn->ibc_incarnation == incarnation)
1025                         continue;
1026
1027                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1028                               "incarnation:"LPX64"(%x, "LPX64")\n",
1029                        libcfs_nid2str(peer->ibp_nid),
1030                        conn->ibc_version, conn->ibc_incarnation,
1031                        version, incarnation);
1032
1033                 kiblnd_close_conn_locked(conn, -ESTALE);
1034                 count++;
1035         }
1036
1037         return count;
1038 }
1039
1040 int
1041 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
1042 {
1043         kib_peer_t           *peer;
1044         struct list_head             *ptmp;
1045         struct list_head             *pnxt;
1046         int                  lo;
1047         int                  hi;
1048         int                  i;
1049         unsigned long      flags;
1050         int                  count = 0;
1051
1052         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1053
1054         if (nid != LNET_NID_ANY)
1055                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1056         else {
1057                 lo = 0;
1058                 hi = kiblnd_data.kib_peer_hash_size - 1;
1059         }
1060
1061         for (i = lo; i <= hi; i++) {
1062                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1063
1064                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1065                         LASSERT (peer->ibp_connecting > 0 ||
1066                                  peer->ibp_accepting > 0 ||
1067                                  !list_empty(&peer->ibp_conns));
1068
1069                         if (peer->ibp_ni != ni)
1070                                 continue;
1071
1072                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1073                                 continue;
1074
1075                         count += kiblnd_close_peer_conns_locked(peer, 0);
1076                 }
1077         }
1078
1079         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1080
1081         /* wildcards always succeed */
1082         if (nid == LNET_NID_ANY)
1083                 return 0;
1084
1085         return (count == 0) ? -ENOENT : 0;
1086 }
1087
1088 int
1089 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1090 {
1091         struct libcfs_ioctl_data *data = arg;
1092         int                    rc = -EINVAL;
1093
1094         switch(cmd) {
1095         case IOC_LIBCFS_GET_PEER: {
1096                 lnet_nid_t   nid = 0;
1097                 int       count = 0;
1098
1099                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1100                                           &nid, &count);
1101                 data->ioc_nid    = nid;
1102                 data->ioc_count  = count;
1103                 break;
1104         }
1105
1106         case IOC_LIBCFS_DEL_PEER: {
1107                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1108                 break;
1109         }
1110         case IOC_LIBCFS_GET_CONN: {
1111                 kib_conn_t *conn;
1112
1113                 rc = 0;
1114                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1115                 if (conn == NULL) {
1116                         rc = -ENOENT;
1117                         break;
1118                 }
1119
1120                 LASSERT (conn->ibc_cmid != NULL);
1121                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1122                 if (conn->ibc_cmid->route.path_rec == NULL)
1123                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1124                 else
1125                         data->ioc_u32[0] =
1126                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1127                 kiblnd_conn_decref(conn);
1128                 break;
1129         }
1130         case IOC_LIBCFS_CLOSE_CONNECTION: {
1131                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1132                 break;
1133         }
1134
1135         default:
1136                 break;
1137         }
1138
1139         return rc;
1140 }
1141
1142 void
1143 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1144 {
1145         cfs_time_t      last_alive = 0;
1146         cfs_time_t      now = cfs_time_current();
1147         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1148         kib_peer_t      *peer;
1149         unsigned long   flags;
1150
1151         read_lock_irqsave(glock, flags);
1152
1153         peer = kiblnd_find_peer_locked(nid);
1154         if (peer != NULL) {
1155                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1156                          peer->ibp_accepting > 0 ||
1157                          !list_empty(&peer->ibp_conns));  /* active conn */
1158                 last_alive = peer->ibp_last_alive;
1159         }
1160
1161         read_unlock_irqrestore(glock, flags);
1162
1163         if (last_alive != 0)
1164                 *when = last_alive;
1165
1166         /* peer is not persistent in hash, trigger peer creation
1167          * and connection establishment with a NULL tx */
1168         if (peer == NULL)
1169                 kiblnd_launch_tx(ni, NULL, nid);
1170
1171         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1172                libcfs_nid2str(nid), peer,
1173                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1174         return;
1175 }
1176
1177 void
1178 kiblnd_free_pages(kib_pages_t *p)
1179 {
1180         int     npages = p->ibp_npages;
1181         int     i;
1182
1183         for (i = 0; i < npages; i++) {
1184                 if (p->ibp_pages[i] != NULL)
1185                         __free_page(p->ibp_pages[i]);
1186         }
1187
1188         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1189 }
1190
1191 int
1192 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1193 {
1194         kib_pages_t     *p;
1195         int             i;
1196
1197         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1198                          offsetof(kib_pages_t, ibp_pages[npages]));
1199         if (p == NULL) {
1200                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1201                 return -ENOMEM;
1202         }
1203
1204         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1205         p->ibp_npages = npages;
1206
1207         for (i = 0; i < npages; i++) {
1208                 p->ibp_pages[i] = alloc_pages_node(
1209                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1210                                     __GFP_IO, 0);
1211                 if (p->ibp_pages[i] == NULL) {
1212                         CERROR("Can't allocate page %d of %d\n", i, npages);
1213                         kiblnd_free_pages(p);
1214                         return -ENOMEM;
1215                 }
1216         }
1217
1218         *pp = p;
1219         return 0;
1220 }
1221
1222 void
1223 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1224 {
1225         kib_rx_t *rx;
1226         int       i;
1227
1228         LASSERT (conn->ibc_rxs != NULL);
1229         LASSERT (conn->ibc_hdev != NULL);
1230
1231         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1232                 rx = &conn->ibc_rxs[i];
1233
1234                 LASSERT (rx->rx_nob >= 0); /* not posted */
1235
1236                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1237                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1238                                                           rx->rx_msgaddr),
1239                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1240         }
1241
1242         kiblnd_free_pages(conn->ibc_rx_pages);
1243
1244         conn->ibc_rx_pages = NULL;
1245 }
1246
1247 void
1248 kiblnd_map_rx_descs(kib_conn_t *conn)
1249 {
1250         kib_rx_t       *rx;
1251         struct page    *pg;
1252         int          pg_off;
1253         int          ipg;
1254         int          i;
1255
1256         for (pg_off = ipg = i = 0;
1257              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1258                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1259                 rx = &conn->ibc_rxs[i];
1260
1261                 rx->rx_conn = conn;
1262                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1263
1264                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1265                                                        rx->rx_msg, IBLND_MSG_SIZE,
1266                                                        DMA_FROM_DEVICE);
1267                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1268                                                    rx->rx_msgaddr));
1269                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1270
1271                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
1272                        i, rx->rx_msg, rx->rx_msgaddr,
1273                        lnet_page2phys(pg) + pg_off);
1274
1275                 pg_off += IBLND_MSG_SIZE;
1276                 LASSERT (pg_off <= PAGE_SIZE);
1277
1278                 if (pg_off == PAGE_SIZE) {
1279                         pg_off = 0;
1280                         ipg++;
1281                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1282                 }
1283         }
1284 }
1285
1286 static void
1287 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1288 {
1289         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1290         kib_tx_t       *tx;
1291         int          i;
1292
1293         LASSERT (tpo->tpo_pool.po_allocated == 0);
1294
1295         if (hdev == NULL)
1296                 return;
1297
1298         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1299                 tx = &tpo->tpo_tx_descs[i];
1300                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1301                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1302                                                           tx->tx_msgaddr),
1303                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1304         }
1305
1306         kiblnd_hdev_decref(hdev);
1307         tpo->tpo_hdev = NULL;
1308 }
1309
1310 static kib_hca_dev_t *
1311 kiblnd_current_hdev(kib_dev_t *dev)
1312 {
1313         kib_hca_dev_t *hdev;
1314         unsigned long  flags;
1315         int         i = 0;
1316
1317         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1318         while (dev->ibd_failover) {
1319                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1320                 if (i++ % 50 == 0)
1321                         CDEBUG(D_NET, "%s: Wait for failover\n",
1322                                dev->ibd_ifname);
1323                 schedule_timeout(cfs_time_seconds(1) / 100);
1324
1325                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1326         }
1327
1328         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1329         hdev = dev->ibd_hdev;
1330
1331         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1332
1333         return hdev;
1334 }
1335
1336 static void
1337 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1338 {
1339         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1340         kib_pool_t     *pool  = &tpo->tpo_pool;
1341         kib_net_t      *net   = pool->po_owner->ps_net;
1342         kib_dev_t      *dev;
1343         struct page    *page;
1344         kib_tx_t       *tx;
1345         int          page_offset;
1346         int          ipage;
1347         int          i;
1348
1349         LASSERT (net != NULL);
1350
1351         dev = net->ibn_dev;
1352
1353         /* pre-mapped messages are not bigger than 1 page */
1354         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1355
1356         /* No fancy arithmetic when we do the buffer calculations */
1357         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1358
1359         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1360
1361         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1362                 page = txpgs->ibp_pages[ipage];
1363                 tx = &tpo->tpo_tx_descs[i];
1364
1365                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1366                                            page_offset);
1367
1368                 tx->tx_msgaddr = kiblnd_dma_map_single(
1369                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1370                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1371                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1372                                                    tx->tx_msgaddr));
1373                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1374
1375                 list_add(&tx->tx_list, &pool->po_free_list);
1376
1377                 page_offset += IBLND_MSG_SIZE;
1378                 LASSERT (page_offset <= PAGE_SIZE);
1379
1380                 if (page_offset == PAGE_SIZE) {
1381                         page_offset = 0;
1382                         ipage++;
1383                         LASSERT (ipage <= txpgs->ibp_npages);
1384                 }
1385         }
1386 }
1387
1388 struct ib_mr *
1389 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1390 {
1391         __u64   index;
1392
1393         LASSERT (hdev->ibh_mrs[0] != NULL);
1394
1395         if (hdev->ibh_nmrs == 1)
1396                 return hdev->ibh_mrs[0];
1397
1398         index = addr >> hdev->ibh_mr_shift;
1399
1400         if (index <  hdev->ibh_nmrs &&
1401             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1402                 return hdev->ibh_mrs[index];
1403
1404         return NULL;
1405 }
1406
1407 struct ib_mr *
1408 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1409 {
1410         struct ib_mr *prev_mr;
1411         struct ib_mr *mr;
1412         int        i;
1413
1414         LASSERT (hdev->ibh_mrs[0] != NULL);
1415
1416         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1417             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1418                 return NULL;
1419
1420         if (hdev->ibh_nmrs == 1)
1421                 return hdev->ibh_mrs[0];
1422
1423         for (i = 0, mr = prev_mr = NULL;
1424              i < rd->rd_nfrags; i++) {
1425                 mr = kiblnd_find_dma_mr(hdev,
1426                                         rd->rd_frags[i].rf_addr,
1427                                         rd->rd_frags[i].rf_nob);
1428                 if (prev_mr == NULL)
1429                         prev_mr = mr;
1430
1431                 if (mr == NULL || prev_mr != mr) {
1432                         /* Can't covered by one single MR */
1433                         mr = NULL;
1434                         break;
1435                 }
1436         }
1437
1438         return mr;
1439 }
1440
1441 void
1442 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1443 {
1444         LASSERT (pool->fpo_map_count == 0);
1445
1446         if (pool->fpo_fmr_pool != NULL)
1447                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1448
1449         if (pool->fpo_hdev != NULL)
1450                 kiblnd_hdev_decref(pool->fpo_hdev);
1451
1452         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1453 }
1454
1455 void
1456 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1457 {
1458         kib_fmr_pool_t *pool;
1459
1460         while (!list_empty(head)) {
1461                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1462                 list_del(&pool->fpo_list);
1463                 kiblnd_destroy_fmr_pool(pool);
1464         }
1465 }
1466
1467 static int kiblnd_fmr_pool_size(int ncpts)
1468 {
1469         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1470
1471         return max(IBLND_FMR_POOL, size);
1472 }
1473
1474 static int kiblnd_fmr_flush_trigger(int ncpts)
1475 {
1476         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1477
1478         return max(IBLND_FMR_POOL_FLUSH, size);
1479 }
1480
1481 int
1482 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1483 {
1484         /* FMR pool for RDMA */
1485         kib_dev_t              *dev = fps->fps_net->ibn_dev;
1486         kib_fmr_pool_t    *fpo;
1487         struct ib_fmr_pool_param param = {
1488                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1489                 .page_shift     = PAGE_SHIFT,
1490                 .access     = (IB_ACCESS_LOCAL_WRITE |
1491                                       IB_ACCESS_REMOTE_WRITE),
1492                 .pool_size         = fps->fps_pool_size,
1493                 .dirty_watermark   = fps->fps_flush_trigger,
1494                 .flush_function    = NULL,
1495                 .flush_arg       = NULL,
1496                 .cache       = !!*kiblnd_tunables.kib_fmr_cache};
1497         int rc;
1498
1499         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1500         if (fpo == NULL)
1501                 return -ENOMEM;
1502
1503         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1504
1505         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1506         if (IS_ERR(fpo->fpo_fmr_pool)) {
1507                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1508                 CERROR("Failed to create FMR pool: %d\n", rc);
1509
1510                 kiblnd_hdev_decref(fpo->fpo_hdev);
1511                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1512                 return rc;
1513         }
1514
1515         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1516         fpo->fpo_owner    = fps;
1517         *pp_fpo = fpo;
1518
1519         return 0;
1520 }
1521
1522 static void
1523 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1524 {
1525         if (fps->fps_net == NULL) /* intialized? */
1526                 return;
1527
1528         spin_lock(&fps->fps_lock);
1529
1530         while (!list_empty(&fps->fps_pool_list)) {
1531                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1532                                                  kib_fmr_pool_t, fpo_list);
1533                 fpo->fpo_failed = 1;
1534                 list_del(&fpo->fpo_list);
1535                 if (fpo->fpo_map_count == 0)
1536                         list_add(&fpo->fpo_list, zombies);
1537                 else
1538                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1539         }
1540
1541         spin_unlock(&fps->fps_lock);
1542 }
1543
1544 static void
1545 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1546 {
1547         if (fps->fps_net != NULL) { /* initialized? */
1548                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1549                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1550         }
1551 }
1552
1553 static int
1554 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1555                         int pool_size, int flush_trigger)
1556 {
1557         kib_fmr_pool_t *fpo;
1558         int          rc;
1559
1560         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1561
1562         fps->fps_net = net;
1563         fps->fps_cpt = cpt;
1564         fps->fps_pool_size = pool_size;
1565         fps->fps_flush_trigger = flush_trigger;
1566         spin_lock_init(&fps->fps_lock);
1567         INIT_LIST_HEAD(&fps->fps_pool_list);
1568         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1569
1570         rc = kiblnd_create_fmr_pool(fps, &fpo);
1571         if (rc == 0)
1572                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1573
1574         return rc;
1575 }
1576
1577 static int
1578 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1579 {
1580         if (fpo->fpo_map_count != 0) /* still in use */
1581                 return 0;
1582         if (fpo->fpo_failed)
1583                 return 1;
1584         return cfs_time_aftereq(now, fpo->fpo_deadline);
1585 }
1586
1587 void
1588 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1589 {
1590         LIST_HEAD     (zombies);
1591         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1592         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1593         cfs_time_t       now = cfs_time_current();
1594         kib_fmr_pool_t    *tmp;
1595         int             rc;
1596
1597         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1598         LASSERT (rc == 0);
1599
1600         if (status != 0) {
1601                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1602                 LASSERT (rc == 0);
1603         }
1604
1605         fmr->fmr_pool = NULL;
1606         fmr->fmr_pfmr = NULL;
1607
1608         spin_lock(&fps->fps_lock);
1609         fpo->fpo_map_count --;  /* decref the pool */
1610
1611         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1612                 /* the first pool is persistent */
1613                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1614                         continue;
1615
1616                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1617                         list_move(&fpo->fpo_list, &zombies);
1618                         fps->fps_version ++;
1619                 }
1620         }
1621         spin_unlock(&fps->fps_lock);
1622
1623         if (!list_empty(&zombies))
1624                 kiblnd_destroy_fmr_pool_list(&zombies);
1625 }
1626
1627 int
1628 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1629                     __u64 iov, kib_fmr_t *fmr)
1630 {
1631         struct ib_pool_fmr *pfmr;
1632         kib_fmr_pool_t     *fpo;
1633         __u64          version;
1634         int              rc;
1635
1636  again:
1637         spin_lock(&fps->fps_lock);
1638         version = fps->fps_version;
1639         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1640                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1641                 fpo->fpo_map_count++;
1642                 spin_unlock(&fps->fps_lock);
1643
1644                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1645                                             pages, npages, iov);
1646                 if (likely(!IS_ERR(pfmr))) {
1647                         fmr->fmr_pool = fpo;
1648                         fmr->fmr_pfmr = pfmr;
1649                         return 0;
1650                 }
1651
1652                 spin_lock(&fps->fps_lock);
1653                 fpo->fpo_map_count--;
1654                 if (PTR_ERR(pfmr) != -EAGAIN) {
1655                         spin_unlock(&fps->fps_lock);
1656                         return PTR_ERR(pfmr);
1657                 }
1658
1659                 /* EAGAIN and ... */
1660                 if (version != fps->fps_version) {
1661                         spin_unlock(&fps->fps_lock);
1662                         goto again;
1663                 }
1664         }
1665
1666         if (fps->fps_increasing) {
1667                 spin_unlock(&fps->fps_lock);
1668                 CDEBUG(D_NET, "Another thread is allocating new "
1669                        "FMR pool, waiting for her to complete\n");
1670                 schedule();
1671                 goto again;
1672
1673         }
1674
1675         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1676                 /* someone failed recently */
1677                 spin_unlock(&fps->fps_lock);
1678                 return -EAGAIN;
1679         }
1680
1681         fps->fps_increasing = 1;
1682         spin_unlock(&fps->fps_lock);
1683
1684         CDEBUG(D_NET, "Allocate new FMR pool\n");
1685         rc = kiblnd_create_fmr_pool(fps, &fpo);
1686         spin_lock(&fps->fps_lock);
1687         fps->fps_increasing = 0;
1688         if (rc == 0) {
1689                 fps->fps_version++;
1690                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1691         } else {
1692                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1693         }
1694         spin_unlock(&fps->fps_lock);
1695
1696         goto again;
1697 }
1698
1699 static void
1700 kiblnd_fini_pool(kib_pool_t *pool)
1701 {
1702         LASSERT (list_empty(&pool->po_free_list));
1703         LASSERT (pool->po_allocated == 0);
1704
1705         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1706 }
1707
1708 static void
1709 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1710 {
1711         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1712
1713         memset(pool, 0, sizeof(kib_pool_t));
1714         INIT_LIST_HEAD(&pool->po_free_list);
1715         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1716         pool->po_owner    = ps;
1717         pool->po_size     = size;
1718 }
1719
1720 void
1721 kiblnd_destroy_pool_list(struct list_head *head)
1722 {
1723         kib_pool_t *pool;
1724
1725         while (!list_empty(head)) {
1726                 pool = list_entry(head->next, kib_pool_t, po_list);
1727                 list_del(&pool->po_list);
1728
1729                 LASSERT (pool->po_owner != NULL);
1730                 pool->po_owner->ps_pool_destroy(pool);
1731         }
1732 }
1733
1734 static void
1735 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1736 {
1737         if (ps->ps_net == NULL) /* intialized? */
1738                 return;
1739
1740         spin_lock(&ps->ps_lock);
1741         while (!list_empty(&ps->ps_pool_list)) {
1742                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1743                                             kib_pool_t, po_list);
1744                 po->po_failed = 1;
1745                 list_del(&po->po_list);
1746                 if (po->po_allocated == 0)
1747                         list_add(&po->po_list, zombies);
1748                 else
1749                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1750         }
1751         spin_unlock(&ps->ps_lock);
1752 }
1753
1754 static void
1755 kiblnd_fini_poolset(kib_poolset_t *ps)
1756 {
1757         if (ps->ps_net != NULL) { /* initialized? */
1758                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1759                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1760         }
1761 }
1762
1763 static int
1764 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1765                     kib_net_t *net, char *name, int size,
1766                     kib_ps_pool_create_t po_create,
1767                     kib_ps_pool_destroy_t po_destroy,
1768                     kib_ps_node_init_t nd_init,
1769                     kib_ps_node_fini_t nd_fini)
1770 {
1771         kib_pool_t      *pool;
1772         int             rc;
1773
1774         memset(ps, 0, sizeof(kib_poolset_t));
1775
1776         ps->ps_cpt          = cpt;
1777         ps->ps_net        = net;
1778         ps->ps_pool_create  = po_create;
1779         ps->ps_pool_destroy = po_destroy;
1780         ps->ps_node_init    = nd_init;
1781         ps->ps_node_fini    = nd_fini;
1782         ps->ps_pool_size    = size;
1783         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1784             >= sizeof(ps->ps_name))
1785                 return -E2BIG;
1786         spin_lock_init(&ps->ps_lock);
1787         INIT_LIST_HEAD(&ps->ps_pool_list);
1788         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1789
1790         rc = ps->ps_pool_create(ps, size, &pool);
1791         if (rc == 0)
1792                 list_add(&pool->po_list, &ps->ps_pool_list);
1793         else
1794                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1795
1796         return rc;
1797 }
1798
1799 static int
1800 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
1801 {
1802         if (pool->po_allocated != 0) /* still in use */
1803                 return 0;
1804         if (pool->po_failed)
1805                 return 1;
1806         return cfs_time_aftereq(now, pool->po_deadline);
1807 }
1808
1809 void
1810 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1811 {
1812         LIST_HEAD  (zombies);
1813         kib_poolset_t  *ps = pool->po_owner;
1814         kib_pool_t     *tmp;
1815         cfs_time_t      now = cfs_time_current();
1816
1817         spin_lock(&ps->ps_lock);
1818
1819         if (ps->ps_node_fini != NULL)
1820                 ps->ps_node_fini(pool, node);
1821
1822         LASSERT (pool->po_allocated > 0);
1823         list_add(node, &pool->po_free_list);
1824         pool->po_allocated --;
1825
1826         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1827                 /* the first pool is persistent */
1828                 if (ps->ps_pool_list.next == &pool->po_list)
1829                         continue;
1830
1831                 if (kiblnd_pool_is_idle(pool, now))
1832                         list_move(&pool->po_list, &zombies);
1833         }
1834         spin_unlock(&ps->ps_lock);
1835
1836         if (!list_empty(&zombies))
1837                 kiblnd_destroy_pool_list(&zombies);
1838 }
1839
1840 struct list_head *
1841 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1842 {
1843         struct list_head            *node;
1844         kib_pool_t          *pool;
1845         int                 rc;
1846
1847  again:
1848         spin_lock(&ps->ps_lock);
1849         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1850                 if (list_empty(&pool->po_free_list))
1851                         continue;
1852
1853                 pool->po_allocated ++;
1854                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1855                 node = pool->po_free_list.next;
1856                 list_del(node);
1857
1858                 if (ps->ps_node_init != NULL) {
1859                         /* still hold the lock */
1860                         ps->ps_node_init(pool, node);
1861                 }
1862                 spin_unlock(&ps->ps_lock);
1863                 return node;
1864         }
1865
1866         /* no available tx pool and ... */
1867         if (ps->ps_increasing) {
1868                 /* another thread is allocating a new pool */
1869                 spin_unlock(&ps->ps_lock);
1870                 CDEBUG(D_NET, "Another thread is allocating new "
1871                        "%s pool, waiting for her to complete\n",
1872                        ps->ps_name);
1873                 schedule();
1874                 goto again;
1875         }
1876
1877         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
1878                 /* someone failed recently */
1879                 spin_unlock(&ps->ps_lock);
1880                 return NULL;
1881         }
1882
1883         ps->ps_increasing = 1;
1884         spin_unlock(&ps->ps_lock);
1885
1886         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1887
1888         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1889
1890         spin_lock(&ps->ps_lock);
1891         ps->ps_increasing = 0;
1892         if (rc == 0) {
1893                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1894         } else {
1895                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1896                 CERROR("Can't allocate new %s pool because out of memory\n",
1897                        ps->ps_name);
1898         }
1899         spin_unlock(&ps->ps_lock);
1900
1901         goto again;
1902 }
1903
1904 void
1905 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1906 {
1907         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1908         struct ib_mr    *mr  = pmr->pmr_mr;
1909
1910         pmr->pmr_mr = NULL;
1911         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1912         if (mr != NULL)
1913                 ib_dereg_mr(mr);
1914 }
1915
1916 int
1917 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1918                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1919 {
1920         kib_phys_mr_t *pmr;
1921         struct list_head    *node;
1922         int         rc;
1923         int         i;
1924
1925         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1926         if (node == NULL) {
1927                 CERROR("Failed to allocate PMR descriptor\n");
1928                 return -ENOMEM;
1929         }
1930
1931         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1932         if (pmr->pmr_pool->ppo_hdev != hdev) {
1933                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1934                 return -EAGAIN;
1935         }
1936
1937         for (i = 0; i < rd->rd_nfrags; i ++) {
1938                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1939                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1940         }
1941
1942         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1943                                      pmr->pmr_ipb, rd->rd_nfrags,
1944                                      IB_ACCESS_LOCAL_WRITE |
1945                                      IB_ACCESS_REMOTE_WRITE,
1946                                      iova);
1947         if (!IS_ERR(pmr->pmr_mr)) {
1948                 pmr->pmr_iova = *iova;
1949                 *pp_pmr = pmr;
1950                 return 0;
1951         }
1952
1953         rc = PTR_ERR(pmr->pmr_mr);
1954         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1955
1956         pmr->pmr_mr = NULL;
1957         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1958
1959         return rc;
1960 }
1961
1962 static void
1963 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1964 {
1965         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1966         kib_phys_mr_t  *pmr;
1967
1968         LASSERT (pool->po_allocated == 0);
1969
1970         while (!list_empty(&pool->po_free_list)) {
1971                 pmr = list_entry(pool->po_free_list.next,
1972                                      kib_phys_mr_t, pmr_list);
1973
1974                 LASSERT (pmr->pmr_mr == NULL);
1975                 list_del(&pmr->pmr_list);
1976
1977                 if (pmr->pmr_ipb != NULL) {
1978                         LIBCFS_FREE(pmr->pmr_ipb,
1979                                     IBLND_MAX_RDMA_FRAGS *
1980                                     sizeof(struct ib_phys_buf));
1981                 }
1982
1983                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1984         }
1985
1986         kiblnd_fini_pool(pool);
1987         if (ppo->ppo_hdev != NULL)
1988                 kiblnd_hdev_decref(ppo->ppo_hdev);
1989
1990         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1991 }
1992
1993 static inline int kiblnd_pmr_pool_size(int ncpts)
1994 {
1995         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1996
1997         return max(IBLND_PMR_POOL, size);
1998 }
1999
2000 static int
2001 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2002 {
2003         struct kib_pmr_pool     *ppo;
2004         struct kib_pool         *pool;
2005         kib_phys_mr_t           *pmr;
2006         int                     i;
2007
2008         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
2009                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
2010         if (ppo == NULL) {
2011                 CERROR("Failed to allocate PMR pool\n");
2012                 return -ENOMEM;
2013         }
2014
2015         pool = &ppo->ppo_pool;
2016         kiblnd_init_pool(ps, pool, size);
2017
2018         for (i = 0; i < size; i++) {
2019                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
2020                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
2021                 if (pmr == NULL)
2022                         break;
2023
2024                 pmr->pmr_pool = ppo;
2025                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
2026                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
2027                 if (pmr->pmr_ipb == NULL)
2028                         break;
2029
2030                 list_add(&pmr->pmr_list, &pool->po_free_list);
2031         }
2032
2033         if (i < size) {
2034                 ps->ps_pool_destroy(pool);
2035                 return -ENOMEM;
2036         }
2037
2038         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
2039         *pp_po = pool;
2040         return 0;
2041 }
2042
2043 static void
2044 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2045 {
2046         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2047         int          i;
2048
2049         LASSERT (pool->po_allocated == 0);
2050
2051         if (tpo->tpo_tx_pages != NULL) {
2052                 kiblnd_unmap_tx_pool(tpo);
2053                 kiblnd_free_pages(tpo->tpo_tx_pages);
2054         }
2055
2056         if (tpo->tpo_tx_descs == NULL)
2057                 goto out;
2058
2059         for (i = 0; i < pool->po_size; i++) {
2060                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2061
2062                 list_del(&tx->tx_list);
2063                 if (tx->tx_pages != NULL)
2064                         LIBCFS_FREE(tx->tx_pages,
2065                                     LNET_MAX_IOV *
2066                                     sizeof(*tx->tx_pages));
2067                 if (tx->tx_frags != NULL)
2068                         LIBCFS_FREE(tx->tx_frags,
2069                                     IBLND_MAX_RDMA_FRAGS *
2070                                             sizeof(*tx->tx_frags));
2071                 if (tx->tx_wrq != NULL)
2072                         LIBCFS_FREE(tx->tx_wrq,
2073                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2074                                     sizeof(*tx->tx_wrq));
2075                 if (tx->tx_sge != NULL)
2076                         LIBCFS_FREE(tx->tx_sge,
2077                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2078                                     sizeof(*tx->tx_sge));
2079                 if (tx->tx_rd != NULL)
2080                         LIBCFS_FREE(tx->tx_rd,
2081                                     offsetof(kib_rdma_desc_t,
2082                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2083         }
2084
2085         LIBCFS_FREE(tpo->tpo_tx_descs,
2086                     pool->po_size * sizeof(kib_tx_t));
2087 out:
2088         kiblnd_fini_pool(pool);
2089         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2090 }
2091
2092 static int kiblnd_tx_pool_size(int ncpts)
2093 {
2094         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2095
2096         return max(IBLND_TX_POOL, ntx);
2097 }
2098
2099 static int
2100 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2101 {
2102         int         i;
2103         int         npg;
2104         kib_pool_t    *pool;
2105         kib_tx_pool_t *tpo;
2106
2107         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2108         if (tpo == NULL) {
2109                 CERROR("Failed to allocate TX pool\n");
2110                 return -ENOMEM;
2111         }
2112
2113         pool = &tpo->tpo_pool;
2114         kiblnd_init_pool(ps, pool, size);
2115         tpo->tpo_tx_descs = NULL;
2116         tpo->tpo_tx_pages = NULL;
2117
2118         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2119         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2120                 CERROR("Can't allocate tx pages: %d\n", npg);
2121                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2122                 return -ENOMEM;
2123         }
2124
2125         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2126                          size * sizeof(kib_tx_t));
2127         if (tpo->tpo_tx_descs == NULL) {
2128                 CERROR("Can't allocate %d tx descriptors\n", size);
2129                 ps->ps_pool_destroy(pool);
2130                 return -ENOMEM;
2131         }
2132
2133         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2134
2135         for (i = 0; i < size; i++) {
2136                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2137
2138                 tx->tx_pool = tpo;
2139                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2140                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2141                                          lnet_cpt_table(), ps->ps_cpt,
2142                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2143                         if (tx->tx_pages == NULL)
2144                                 break;
2145                 }
2146
2147                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2148                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2149                 if (tx->tx_frags == NULL)
2150                         break;
2151
2152                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2153
2154                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2155                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2156                                  sizeof(*tx->tx_wrq));
2157                 if (tx->tx_wrq == NULL)
2158                         break;
2159
2160                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2161                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2162                                  sizeof(*tx->tx_sge));
2163                 if (tx->tx_sge == NULL)
2164                         break;
2165
2166                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2167                                  offsetof(kib_rdma_desc_t,
2168                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2169                 if (tx->tx_rd == NULL)
2170                         break;
2171         }
2172
2173         if (i == size) {
2174                 kiblnd_map_tx_pool(tpo);
2175                 *pp_po = pool;
2176                 return 0;
2177         }
2178
2179         ps->ps_pool_destroy(pool);
2180         return -ENOMEM;
2181 }
2182
2183 static void
2184 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2185 {
2186         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2187                                              tps_poolset);
2188         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2189
2190         tx->tx_cookie = tps->tps_next_tx_cookie ++;
2191 }
2192
2193 void
2194 kiblnd_net_fini_pools(kib_net_t *net)
2195 {
2196         int     i;
2197
2198         cfs_cpt_for_each(i, lnet_cpt_table()) {
2199                 kib_tx_poolset_t        *tps;
2200                 kib_fmr_poolset_t       *fps;
2201                 kib_pmr_poolset_t       *pps;
2202
2203                 if (net->ibn_tx_ps != NULL) {
2204                         tps = net->ibn_tx_ps[i];
2205                         kiblnd_fini_poolset(&tps->tps_poolset);
2206                 }
2207
2208                 if (net->ibn_fmr_ps != NULL) {
2209                         fps = net->ibn_fmr_ps[i];
2210                         kiblnd_fini_fmr_poolset(fps);
2211                 }
2212
2213                 if (net->ibn_pmr_ps != NULL) {
2214                         pps = net->ibn_pmr_ps[i];
2215                         kiblnd_fini_poolset(&pps->pps_poolset);
2216                 }
2217         }
2218
2219         if (net->ibn_tx_ps != NULL) {
2220                 cfs_percpt_free(net->ibn_tx_ps);
2221                 net->ibn_tx_ps = NULL;
2222         }
2223
2224         if (net->ibn_fmr_ps != NULL) {
2225                 cfs_percpt_free(net->ibn_fmr_ps);
2226                 net->ibn_fmr_ps = NULL;
2227         }
2228
2229         if (net->ibn_pmr_ps != NULL) {
2230                 cfs_percpt_free(net->ibn_pmr_ps);
2231                 net->ibn_pmr_ps = NULL;
2232         }
2233 }
2234
2235 int
2236 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2237 {
2238         unsigned long   flags;
2239         int             cpt;
2240         int             rc;
2241         int             i;
2242
2243         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2244         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2245             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2246                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2247                                            flags);
2248                 goto create_tx_pool;
2249         }
2250
2251         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2252
2253         if (*kiblnd_tunables.kib_fmr_pool_size <
2254             *kiblnd_tunables.kib_ntx / 4) {
2255                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2256                        *kiblnd_tunables.kib_fmr_pool_size,
2257                        *kiblnd_tunables.kib_ntx / 4);
2258                 rc = -EINVAL;
2259                 goto failed;
2260         }
2261
2262         /* TX pool must be created later than FMR/PMR, see LU-2268
2263          * for details */
2264         LASSERT(net->ibn_tx_ps == NULL);
2265
2266         /* premapping can fail if ibd_nmr > 1, so we always create
2267          * FMR/PMR pool and map-on-demand if premapping failed */
2268
2269         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2270                                            sizeof(kib_fmr_poolset_t));
2271         if (net->ibn_fmr_ps == NULL) {
2272                 CERROR("Failed to allocate FMR pool array\n");
2273                 rc = -ENOMEM;
2274                 goto failed;
2275         }
2276
2277         for (i = 0; i < ncpts; i++) {
2278                 cpt = (cpts == NULL) ? i : cpts[i];
2279                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2280                                              kiblnd_fmr_pool_size(ncpts),
2281                                              kiblnd_fmr_flush_trigger(ncpts));
2282                 if (rc == -ENOSYS && i == 0) /* no FMR */
2283                         break; /* create PMR pool */
2284
2285                 if (rc != 0) { /* a real error */
2286                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2287                                cpt, rc);
2288                         goto failed;
2289                 }
2290         }
2291
2292         if (i > 0) {
2293                 LASSERT(i == ncpts);
2294                 goto create_tx_pool;
2295         }
2296
2297         cfs_percpt_free(net->ibn_fmr_ps);
2298         net->ibn_fmr_ps = NULL;
2299
2300         CWARN("Device does not support FMR, failing back to PMR\n");
2301
2302         if (*kiblnd_tunables.kib_pmr_pool_size <
2303             *kiblnd_tunables.kib_ntx / 4) {
2304                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2305                        *kiblnd_tunables.kib_pmr_pool_size,
2306                        *kiblnd_tunables.kib_ntx / 4);
2307                 rc = -EINVAL;
2308                 goto failed;
2309         }
2310
2311         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2312                                            sizeof(kib_pmr_poolset_t));
2313         if (net->ibn_pmr_ps == NULL) {
2314                 CERROR("Failed to allocate PMR pool array\n");
2315                 rc = -ENOMEM;
2316                 goto failed;
2317         }
2318
2319         for (i = 0; i < ncpts; i++) {
2320                 cpt = (cpts == NULL) ? i : cpts[i];
2321                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2322                                          cpt, net, "PMR",
2323                                          kiblnd_pmr_pool_size(ncpts),
2324                                          kiblnd_create_pmr_pool,
2325                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2326                 if (rc != 0) {
2327                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2328                                cpt, rc);
2329                         goto failed;
2330                 }
2331         }
2332
2333  create_tx_pool:
2334         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2335                                           sizeof(kib_tx_poolset_t));
2336         if (net->ibn_tx_ps == NULL) {
2337                 CERROR("Failed to allocate tx pool array\n");
2338                 rc = -ENOMEM;
2339                 goto failed;
2340         }
2341
2342         for (i = 0; i < ncpts; i++) {
2343                 cpt = (cpts == NULL) ? i : cpts[i];
2344                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2345                                          cpt, net, "TX",
2346                                          kiblnd_tx_pool_size(ncpts),
2347                                          kiblnd_create_tx_pool,
2348                                          kiblnd_destroy_tx_pool,
2349                                          kiblnd_tx_init, NULL);
2350                 if (rc != 0) {
2351                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2352                                cpt, rc);
2353                         goto failed;
2354                 }
2355         }
2356
2357         return 0;
2358  failed:
2359         kiblnd_net_fini_pools(net);
2360         LASSERT(rc != 0);
2361         return rc;
2362 }
2363
2364 static int
2365 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2366 {
2367         struct ib_device_attr *attr;
2368         int                 rc;
2369
2370         /* It's safe to assume a HCA can handle a page size
2371          * matching that of the native system */
2372         hdev->ibh_page_shift = PAGE_SHIFT;
2373         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2374         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2375
2376         LIBCFS_ALLOC(attr, sizeof(*attr));
2377         if (attr == NULL) {
2378                 CERROR("Out of memory\n");
2379                 return -ENOMEM;
2380         }
2381
2382         rc = ib_query_device(hdev->ibh_ibdev, attr);
2383         if (rc == 0)
2384                 hdev->ibh_mr_size = attr->max_mr_size;
2385
2386         LIBCFS_FREE(attr, sizeof(*attr));
2387
2388         if (rc != 0) {
2389                 CERROR("Failed to query IB device: %d\n", rc);
2390                 return rc;
2391         }
2392
2393         if (hdev->ibh_mr_size == ~0ULL) {
2394                 hdev->ibh_mr_shift = 64;
2395                 return 0;
2396         }
2397
2398         for (hdev->ibh_mr_shift = 0;
2399              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift ++) {
2400                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2401                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2402                         return 0;
2403         }
2404
2405         CERROR("Invalid mr size: "LPX64"\n", hdev->ibh_mr_size);
2406         return -EINVAL;
2407 }
2408
2409 void
2410 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2411 {
2412         int     i;
2413
2414         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2415                 return;
2416
2417         for (i = 0; i < hdev->ibh_nmrs; i++) {
2418                 if (hdev->ibh_mrs[i] == NULL)
2419                         break;
2420
2421                 ib_dereg_mr(hdev->ibh_mrs[i]);
2422         }
2423
2424         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2425         hdev->ibh_mrs  = NULL;
2426         hdev->ibh_nmrs = 0;
2427 }
2428
2429 void
2430 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2431 {
2432         kiblnd_hdev_cleanup_mrs(hdev);
2433
2434         if (hdev->ibh_pd != NULL)
2435                 ib_dealloc_pd(hdev->ibh_pd);
2436
2437         if (hdev->ibh_cmid != NULL)
2438                 rdma_destroy_id(hdev->ibh_cmid);
2439
2440         LIBCFS_FREE(hdev, sizeof(*hdev));
2441 }
2442
2443 int
2444 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2445 {
2446         struct ib_mr *mr;
2447         int        i;
2448         int        rc;
2449         __u64    mm_size;
2450         __u64    mr_size;
2451         int        acflags = IB_ACCESS_LOCAL_WRITE |
2452                                 IB_ACCESS_REMOTE_WRITE;
2453
2454         rc = kiblnd_hdev_get_attr(hdev);
2455         if (rc != 0)
2456                 return rc;
2457
2458         if (hdev->ibh_mr_shift == 64) {
2459                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2460                 if (hdev->ibh_mrs == NULL) {
2461                         CERROR("Failed to allocate MRs table\n");
2462                         return -ENOMEM;
2463                 }
2464
2465                 hdev->ibh_mrs[0] = NULL;
2466                 hdev->ibh_nmrs   = 1;
2467
2468                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2469                 if (IS_ERR(mr)) {
2470                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2471                         kiblnd_hdev_cleanup_mrs(hdev);
2472                         return PTR_ERR(mr);
2473                 }
2474
2475                 hdev->ibh_mrs[0] = mr;
2476
2477                 goto out;
2478         }
2479
2480         mr_size = (1ULL << hdev->ibh_mr_shift);
2481         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2482
2483         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2484
2485         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2486                 /* it's 4T..., assume we will re-code at that time */
2487                 CERROR("Can't support memory size: x"LPX64
2488                        " with MR size: x"LPX64"\n", mm_size, mr_size);
2489                 return -EINVAL;
2490         }
2491
2492         /* create an array of MRs to cover all memory */
2493         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2494         if (hdev->ibh_mrs == NULL) {
2495                 CERROR("Failed to allocate MRs' table\n");
2496                 return -ENOMEM;
2497         }
2498
2499         memset(hdev->ibh_mrs, 0, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2500
2501         for (i = 0; i < hdev->ibh_nmrs; i++) {
2502                 struct ib_phys_buf ipb;
2503                 __u64         iova;
2504
2505                 ipb.size = hdev->ibh_mr_size;
2506                 ipb.addr = i * mr_size;
2507                 iova     = ipb.addr;
2508
2509                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2510                 if (IS_ERR(mr)) {
2511                         CERROR("Failed ib_reg_phys_mr addr "LPX64
2512                                " size "LPX64" : %ld\n",
2513                                ipb.addr, ipb.size, PTR_ERR(mr));
2514                         kiblnd_hdev_cleanup_mrs(hdev);
2515                         return PTR_ERR(mr);
2516                 }
2517
2518                 LASSERT (iova == ipb.addr);
2519
2520                 hdev->ibh_mrs[i] = mr;
2521         }
2522
2523 out:
2524         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2525                 LCONSOLE_INFO("Register global MR array, MR size: "
2526                               LPX64", array size: %d\n",
2527                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2528         return 0;
2529 }
2530
2531 static int
2532 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2533 {       /* DUMMY */
2534         return 0;
2535 }
2536
2537 static int
2538 kiblnd_dev_need_failover(kib_dev_t *dev)
2539 {
2540         struct rdma_cm_id  *cmid;
2541         struct sockaddr_in  srcaddr;
2542         struct sockaddr_in  dstaddr;
2543         int              rc;
2544
2545         if (dev->ibd_hdev == NULL || /* initializing */
2546             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2547             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2548                 return 1;
2549
2550         /* XXX: it's UGLY, but I don't have better way to find
2551          * ib-bonding HCA failover because:
2552          *
2553          * a. no reliable CM event for HCA failover...
2554          * b. no OFED API to get ib_device for current net_device...
2555          *
2556          * We have only two choices at this point:
2557          *
2558          * a. rdma_bind_addr(), it will conflict with listener cmid
2559          * b. rdma_resolve_addr() to zero addr */
2560         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2561                                      IB_QPT_RC);
2562         if (IS_ERR(cmid)) {
2563                 rc = PTR_ERR(cmid);
2564                 CERROR("Failed to create cmid for failover: %d\n", rc);
2565                 return rc;
2566         }
2567
2568         memset(&srcaddr, 0, sizeof(srcaddr));
2569         srcaddr.sin_family      = AF_INET;
2570         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2571
2572         memset(&dstaddr, 0, sizeof(dstaddr));
2573         dstaddr.sin_family = AF_INET;
2574         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2575                                (struct sockaddr *)&dstaddr, 1);
2576         if (rc != 0 || cmid->device == NULL) {
2577                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2578                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2579                        cmid->device, rc);
2580                 rdma_destroy_id(cmid);
2581                 return rc;
2582         }
2583
2584         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2585                 /* don't need device failover */
2586                 rdma_destroy_id(cmid);
2587                 return 0;
2588         }
2589
2590         return 1;
2591 }
2592
2593 int
2594 kiblnd_dev_failover(kib_dev_t *dev)
2595 {
2596         LIST_HEAD      (zombie_tpo);
2597         LIST_HEAD      (zombie_ppo);
2598         LIST_HEAD      (zombie_fpo);
2599         struct rdma_cm_id  *cmid  = NULL;
2600         kib_hca_dev_t      *hdev  = NULL;
2601         kib_hca_dev_t      *old;
2602         struct ib_pd       *pd;
2603         kib_net_t         *net;
2604         struct sockaddr_in  addr;
2605         unsigned long       flags;
2606         int              rc = 0;
2607         int                 i;
2608
2609         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2610                  dev->ibd_can_failover ||
2611                  dev->ibd_hdev == NULL);
2612
2613         rc = kiblnd_dev_need_failover(dev);
2614         if (rc <= 0)
2615                 goto out;
2616
2617         if (dev->ibd_hdev != NULL &&
2618             dev->ibd_hdev->ibh_cmid != NULL) {
2619                 /* XXX it's not good to close old listener at here,
2620                  * because we can fail to create new listener.
2621                  * But we have to close it now, otherwise rdma_bind_addr
2622                  * will return EADDRINUSE... How crap! */
2623                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2624
2625                 cmid = dev->ibd_hdev->ibh_cmid;
2626                 /* make next schedule of kiblnd_dev_need_failover()
2627                  * return 1 for me */
2628                 dev->ibd_hdev->ibh_cmid  = NULL;
2629                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2630
2631                 rdma_destroy_id(cmid);
2632         }
2633
2634         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2635                                      IB_QPT_RC);
2636         if (IS_ERR(cmid)) {
2637                 rc = PTR_ERR(cmid);
2638                 CERROR("Failed to create cmid for failover: %d\n", rc);
2639                 goto out;
2640         }
2641
2642         memset(&addr, 0, sizeof(addr));
2643         addr.sin_family      = AF_INET;
2644         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2645         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2646
2647         /* Bind to failover device or port */
2648         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2649         if (rc != 0 || cmid->device == NULL) {
2650                 CERROR("Failed to bind %s:%u.%u.%u.%u to device(%p): %d\n",
2651                        dev->ibd_ifname, HIPQUAD(dev->ibd_ifip),
2652                        cmid->device, rc);
2653                 rdma_destroy_id(cmid);
2654                 goto out;
2655         }
2656
2657         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2658         if (hdev == NULL) {
2659                 CERROR("Failed to allocate kib_hca_dev\n");
2660                 rdma_destroy_id(cmid);
2661                 rc = -ENOMEM;
2662                 goto out;
2663         }
2664
2665         atomic_set(&hdev->ibh_ref, 1);
2666         hdev->ibh_dev   = dev;
2667         hdev->ibh_cmid  = cmid;
2668         hdev->ibh_ibdev = cmid->device;
2669
2670         pd = ib_alloc_pd(cmid->device);
2671         if (IS_ERR(pd)) {
2672                 rc = PTR_ERR(pd);
2673                 CERROR("Can't allocate PD: %d\n", rc);
2674                 goto out;
2675         }
2676
2677         hdev->ibh_pd = pd;
2678
2679         rc = rdma_listen(cmid, 0);
2680         if (rc != 0) {
2681                 CERROR("Can't start new listener: %d\n", rc);
2682                 goto out;
2683         }
2684
2685         rc = kiblnd_hdev_setup_mrs(hdev);
2686         if (rc != 0) {
2687                 CERROR("Can't setup device: %d\n", rc);
2688                 goto out;
2689         }
2690
2691         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2692
2693         old = dev->ibd_hdev;
2694         dev->ibd_hdev = hdev; /* take over the refcount */
2695         hdev = old;
2696
2697         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2698                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2699                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2700                                             &zombie_tpo);
2701
2702                         if (net->ibn_fmr_ps != NULL) {
2703                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2704                                                         &zombie_fpo);
2705
2706                         } else if (net->ibn_pmr_ps != NULL) {
2707                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2708                                                     pps_poolset, &zombie_ppo);
2709                         }
2710                 }
2711         }
2712
2713         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2714  out:
2715         if (!list_empty(&zombie_tpo))
2716                 kiblnd_destroy_pool_list(&zombie_tpo);
2717         if (!list_empty(&zombie_ppo))
2718                 kiblnd_destroy_pool_list(&zombie_ppo);
2719         if (!list_empty(&zombie_fpo))
2720                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2721         if (hdev != NULL)
2722                 kiblnd_hdev_decref(hdev);
2723
2724         if (rc != 0)
2725                 dev->ibd_failed_failover++;
2726         else
2727                 dev->ibd_failed_failover = 0;
2728
2729         return rc;
2730 }
2731
2732 void
2733 kiblnd_destroy_dev (kib_dev_t *dev)
2734 {
2735         LASSERT (dev->ibd_nnets == 0);
2736         LASSERT (list_empty(&dev->ibd_nets));
2737
2738         list_del(&dev->ibd_fail_list);
2739         list_del(&dev->ibd_list);
2740
2741         if (dev->ibd_hdev != NULL)
2742                 kiblnd_hdev_decref(dev->ibd_hdev);
2743
2744         LIBCFS_FREE(dev, sizeof(*dev));
2745 }
2746
2747 kib_dev_t *
2748 kiblnd_create_dev(char *ifname)
2749 {
2750         struct net_device *netdev;
2751         kib_dev_t        *dev;
2752         __u32         netmask;
2753         __u32         ip;
2754         int             up;
2755         int             rc;
2756
2757         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2758         if (rc != 0) {
2759                 CERROR("Can't query IPoIB interface %s: %d\n",
2760                        ifname, rc);
2761                 return NULL;
2762         }
2763
2764         if (!up) {
2765                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2766                 return NULL;
2767         }
2768
2769         LIBCFS_ALLOC(dev, sizeof(*dev));
2770         if (dev == NULL)
2771                 return NULL;
2772
2773         memset(dev, 0, sizeof(*dev));
2774         netdev = dev_get_by_name(&init_net, ifname);
2775         if (netdev == NULL) {
2776                 dev->ibd_can_failover = 0;
2777         } else {
2778                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2779                 dev_put(netdev);
2780         }
2781
2782         INIT_LIST_HEAD(&dev->ibd_nets);
2783         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2784         INIT_LIST_HEAD(&dev->ibd_fail_list);
2785         dev->ibd_ifip = ip;
2786         strcpy(&dev->ibd_ifname[0], ifname);
2787
2788         /* initialize the device */
2789         rc = kiblnd_dev_failover(dev);
2790         if (rc != 0) {
2791                 CERROR("Can't initialize device: %d\n", rc);
2792                 LIBCFS_FREE(dev, sizeof(*dev));
2793                 return NULL;
2794         }
2795
2796         list_add_tail(&dev->ibd_list,
2797                           &kiblnd_data.kib_devs);
2798         return dev;
2799 }
2800
2801 void
2802 kiblnd_base_shutdown(void)
2803 {
2804         struct kib_sched_info   *sched;
2805         int                     i;
2806
2807         LASSERT (list_empty(&kiblnd_data.kib_devs));
2808
2809         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2810                atomic_read(&libcfs_kmemory));
2811
2812         switch (kiblnd_data.kib_init) {
2813         default:
2814                 LBUG();
2815
2816         case IBLND_INIT_ALL:
2817         case IBLND_INIT_DATA:
2818                 LASSERT (kiblnd_data.kib_peers != NULL);
2819                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2820                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2821                 }
2822                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2823                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2824
2825                 /* flag threads to terminate; wake and wait for them to die */
2826                 kiblnd_data.kib_shutdown = 1;
2827
2828                 /* NB: we really want to stop scheduler threads net by net
2829                  * instead of the whole module, this should be improved
2830                  * with dynamic configuration LNet */
2831                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2832                         wake_up_all(&sched->ibs_waitq);
2833
2834                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2835                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2836
2837                 i = 2;
2838                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2839                         i++;
2840                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2841                                "Waiting for %d threads to terminate\n",
2842                                atomic_read(&kiblnd_data.kib_nthreads));
2843                         cfs_pause(cfs_time_seconds(1));
2844                 }
2845
2846                 /* fall through */
2847
2848         case IBLND_INIT_NOTHING:
2849                 break;
2850         }
2851
2852         if (kiblnd_data.kib_peers != NULL) {
2853                 LIBCFS_FREE(kiblnd_data.kib_peers,
2854                             sizeof(struct list_head) *
2855                             kiblnd_data.kib_peer_hash_size);
2856         }
2857
2858         if (kiblnd_data.kib_scheds != NULL)
2859                 cfs_percpt_free(kiblnd_data.kib_scheds);
2860
2861         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2862                atomic_read(&libcfs_kmemory));
2863
2864         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2865         module_put(THIS_MODULE);
2866 }
2867
2868 void
2869 kiblnd_shutdown (lnet_ni_t *ni)
2870 {
2871         kib_net_t       *net = ni->ni_data;
2872         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2873         int            i;
2874         unsigned long     flags;
2875
2876         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2877
2878         if (net == NULL)
2879                 goto out;
2880
2881         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2882                atomic_read(&libcfs_kmemory));
2883
2884         write_lock_irqsave(g_lock, flags);
2885         net->ibn_shutdown = 1;
2886         write_unlock_irqrestore(g_lock, flags);
2887
2888         switch (net->ibn_init) {
2889         default:
2890                 LBUG();
2891
2892         case IBLND_INIT_ALL:
2893                 /* nuke all existing peers within this net */
2894                 kiblnd_del_peer(ni, LNET_NID_ANY);
2895
2896                 /* Wait for all peer state to clean up */
2897                 i = 2;
2898                 while (atomic_read(&net->ibn_npeers) != 0) {
2899                         i++;
2900                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2901                                "%s: waiting for %d peers to disconnect\n",
2902                                libcfs_nid2str(ni->ni_nid),
2903                                atomic_read(&net->ibn_npeers));
2904                         cfs_pause(cfs_time_seconds(1));
2905                 }
2906
2907                 kiblnd_net_fini_pools(net);
2908
2909                 write_lock_irqsave(g_lock, flags);
2910                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2911                 net->ibn_dev->ibd_nnets--;
2912                 list_del(&net->ibn_list);
2913                 write_unlock_irqrestore(g_lock, flags);
2914
2915                 /* fall through */
2916
2917         case IBLND_INIT_NOTHING:
2918                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2919
2920                 if (net->ibn_dev != NULL &&
2921                     net->ibn_dev->ibd_nnets == 0)
2922                         kiblnd_destroy_dev(net->ibn_dev);
2923
2924                 break;
2925         }
2926
2927         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2928                atomic_read(&libcfs_kmemory));
2929
2930         net->ibn_init = IBLND_INIT_NOTHING;
2931         ni->ni_data = NULL;
2932
2933         LIBCFS_FREE(net, sizeof(*net));
2934
2935 out:
2936         if (list_empty(&kiblnd_data.kib_devs))
2937                 kiblnd_base_shutdown();
2938         return;
2939 }
2940
2941 int
2942 kiblnd_base_startup(void)
2943 {
2944         struct kib_sched_info   *sched;
2945         int                     rc;
2946         int                     i;
2947
2948         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2949
2950         try_module_get(THIS_MODULE);
2951         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2952
2953         rwlock_init(&kiblnd_data.kib_global_lock);
2954
2955         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2956         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2957
2958         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2959         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2960                      sizeof(struct list_head) *
2961                             kiblnd_data.kib_peer_hash_size);
2962         if (kiblnd_data.kib_peers == NULL) {
2963                 goto failed;
2964         }
2965         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2966                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2967
2968         spin_lock_init(&kiblnd_data.kib_connd_lock);
2969         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2970         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2971         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2972         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2973
2974         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2975                                                   sizeof(*sched));
2976         if (kiblnd_data.kib_scheds == NULL)
2977                 goto failed;
2978
2979         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2980                 int     nthrs;
2981
2982                 spin_lock_init(&sched->ibs_lock);
2983                 INIT_LIST_HEAD(&sched->ibs_conns);
2984                 init_waitqueue_head(&sched->ibs_waitq);
2985
2986                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2987                 if (*kiblnd_tunables.kib_nscheds > 0) {
2988                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2989                 } else {
2990                         /* max to half of CPUs, another half is reserved for
2991                          * upper layer modules */
2992                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2993                 }
2994
2995                 sched->ibs_nthreads_max = nthrs;
2996                 sched->ibs_cpt = i;
2997         }
2998
2999         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3000
3001         /* lists/ptrs/locks initialised */
3002         kiblnd_data.kib_init = IBLND_INIT_DATA;
3003         /*****************************************************/
3004
3005         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3006         if (rc != 0) {
3007                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3008                 goto failed;
3009         }
3010
3011         if (*kiblnd_tunables.kib_dev_failover != 0)
3012                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3013                                          "kiblnd_failover");
3014
3015         if (rc != 0) {
3016                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3017                 goto failed;
3018         }
3019
3020         /* flag everything initialised */
3021         kiblnd_data.kib_init = IBLND_INIT_ALL;
3022         /*****************************************************/
3023
3024         return 0;
3025
3026  failed:
3027         kiblnd_base_shutdown();
3028         return -ENETDOWN;
3029 }
3030
3031 int
3032 kiblnd_start_schedulers(struct kib_sched_info *sched)
3033 {
3034         int     rc = 0;
3035         int     nthrs;
3036         int     i;
3037
3038         if (sched->ibs_nthreads == 0) {
3039                 if (*kiblnd_tunables.kib_nscheds > 0) {
3040                         nthrs = sched->ibs_nthreads_max;
3041                 } else {
3042                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3043                                                sched->ibs_cpt);
3044                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3045                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3046                 }
3047         } else {
3048                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3049                 /* increase one thread if there is new interface */
3050                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3051         }
3052
3053         for (i = 0; i < nthrs; i++) {
3054                 long    id;
3055                 char    name[20];
3056                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3057                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3058                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3059                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3060                 if (rc == 0)
3061                         continue;
3062
3063                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3064                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3065                 break;
3066         }
3067
3068         sched->ibs_nthreads += i;
3069         return rc;
3070 }
3071
3072 int
3073 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3074 {
3075         int     cpt;
3076         int     rc;
3077         int     i;
3078
3079         for (i = 0; i < ncpts; i++) {
3080                 struct kib_sched_info *sched;
3081
3082                 cpt = (cpts == NULL) ? i : cpts[i];
3083                 sched = kiblnd_data.kib_scheds[cpt];
3084
3085                 if (!newdev && sched->ibs_nthreads > 0)
3086                         continue;
3087
3088                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3089                 if (rc != 0) {
3090                         CERROR("Failed to start scheduler threads for %s\n",
3091                                dev->ibd_ifname);
3092                         return rc;
3093                 }
3094         }
3095         return 0;
3096 }
3097
3098 kib_dev_t *
3099 kiblnd_dev_search(char *ifname)
3100 {
3101         kib_dev_t       *alias = NULL;
3102         kib_dev_t       *dev;
3103         char            *colon;
3104         char            *colon2;
3105
3106         colon = strchr(ifname, ':');
3107         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3108                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3109                         return dev;
3110
3111                 if (alias != NULL)
3112                         continue;
3113
3114                 colon2 = strchr(dev->ibd_ifname, ':');
3115                 if (colon != NULL)
3116                         *colon = 0;
3117                 if (colon2 != NULL)
3118                         *colon2 = 0;
3119
3120                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3121                         alias = dev;
3122
3123                 if (colon != NULL)
3124                         *colon = ':';
3125                 if (colon2 != NULL)
3126                         *colon2 = ':';
3127         }
3128         return alias;
3129 }
3130
3131 int
3132 kiblnd_startup (lnet_ni_t *ni)
3133 {
3134         char                 *ifname;
3135         kib_dev_t               *ibdev = NULL;
3136         kib_net_t               *net;
3137         struct timeval      tv;
3138         unsigned long        flags;
3139         int                    rc;
3140         int                       newdev;
3141
3142         LASSERT (ni->ni_lnd == &the_o2iblnd);
3143
3144         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3145                 rc = kiblnd_base_startup();
3146                 if (rc != 0)
3147                         return rc;
3148         }
3149
3150         LIBCFS_ALLOC(net, sizeof(*net));
3151         ni->ni_data = net;
3152         if (net == NULL)
3153                 goto failed;
3154
3155         memset(net, 0, sizeof(*net));
3156
3157         do_gettimeofday(&tv);
3158         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3159
3160         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3161         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3162         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3163         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3164
3165         if (ni->ni_interfaces[0] != NULL) {
3166                 /* Use the IPoIB interface specified in 'networks=' */
3167
3168                 CLASSERT (LNET_MAX_INTERFACES > 1);
3169                 if (ni->ni_interfaces[1] != NULL) {
3170                         CERROR("Multiple interfaces not supported\n");
3171                         goto failed;
3172                 }
3173
3174                 ifname = ni->ni_interfaces[0];
3175         } else {
3176                 ifname = *kiblnd_tunables.kib_default_ipif;
3177         }
3178
3179         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3180                 CERROR("IPoIB interface name too long: %s\n", ifname);
3181                 goto failed;
3182         }
3183
3184         ibdev = kiblnd_dev_search(ifname);
3185
3186         newdev = ibdev == NULL;
3187         /* hmm...create kib_dev even for alias */
3188         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3189                 ibdev = kiblnd_create_dev(ifname);
3190
3191         if (ibdev == NULL)
3192                 goto failed;
3193
3194         net->ibn_dev = ibdev;
3195         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3196
3197         rc = kiblnd_dev_start_threads(ibdev, newdev,
3198                                       ni->ni_cpts, ni->ni_ncpts);
3199         if (rc != 0)
3200                 goto failed;
3201
3202         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3203         if (rc != 0) {
3204                 CERROR("Failed to initialize NI pools: %d\n", rc);
3205                 goto failed;
3206         }
3207
3208         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3209         ibdev->ibd_nnets++;
3210         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3211         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3212
3213         net->ibn_init = IBLND_INIT_ALL;
3214
3215         return 0;
3216
3217 failed:
3218         if (net->ibn_dev == NULL && ibdev != NULL)
3219                 kiblnd_destroy_dev(ibdev);
3220
3221         kiblnd_shutdown(ni);
3222
3223         CDEBUG(D_NET, "kiblnd_startup failed\n");
3224         return -ENETDOWN;
3225 }
3226
3227 void __exit
3228 kiblnd_module_fini (void)
3229 {
3230         lnet_unregister_lnd(&the_o2iblnd);
3231         kiblnd_tunables_fini();
3232 }
3233
3234 int __init
3235 kiblnd_module_init (void)
3236 {
3237         int    rc;
3238
3239         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3240         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3241                   <= IBLND_MSG_SIZE);
3242         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3243                   <= IBLND_MSG_SIZE);
3244
3245         rc = kiblnd_tunables_init();
3246         if (rc != 0)
3247                 return rc;
3248
3249         lnet_register_lnd(&the_o2iblnd);
3250
3251         return 0;
3252 }
3253
3254 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3255 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3256 MODULE_LICENSE("GPL");
3257
3258 module_init(kiblnd_module_init);
3259 module_exit(kiblnd_module_fini);