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 static 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 static __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 static 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 static 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 static 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 static 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 int
 588 kiblnd_translate_mtu(int value)
 589 {
 590         switch (value) {
 591         default:
 592                 return -1;
 593         case 0:
 594                 return 0;
 595         case 256:
 596                 return IB_MTU_256;
 597         case 512:
 598                 return IB_MTU_512;
 599         case 1024:
 600                 return IB_MTU_1024;
 601         case 2048:
 602                 return IB_MTU_2048;
 603         case 4096:
 604                 return IB_MTU_4096;
 605         }
 606 }
 607
 608 static void
 609 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 610 {
 611         int        mtu;
 612
 613         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 614         if (cmid->route.path_rec == NULL)
 615                 return;
 616
 617         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 618         LASSERT (mtu >= 0);
 619         if (mtu != 0)
 620                 cmid->route.path_rec->mtu = mtu;
 621 }
 622
 623 static int
 624 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 625 {
 626         cpumask_t       *mask;
 627         int             vectors;
 628         int             off;
 629         int             i;
 630         lnet_nid_t      nid = conn->ibc_peer->ibp_nid;
 631
 632         vectors = conn->ibc_cmid->device->num_comp_vectors;
 633         if (vectors <= 1)
 634                 return 0;
 635
 636         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 637         if (mask == NULL)
 638                 return 0;
 639
 640         /* hash NID to CPU id in this partition... */
 641         off = do_div(nid, cpus_weight(*mask));
 642         for_each_cpu_mask(i, *mask) {
 643                 if (off-- == 0)
 644                         return i % vectors;
 645         }
 646
 647         LBUG();
 648         return 1;
 649 }
 650
 651 kib_conn_t *
 652 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 653                    int state, int version)
 654 {
 655         /* CAVEAT EMPTOR:
 656          * If the new conn is created successfully it takes over the caller's
 657          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 658          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 659          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 660          * to destroy 'cmid' here since I'm called from the CM which still has
 661          * its ref on 'cmid'). */
 662         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
 663         kib_net_t             *net = peer->ibp_ni->ni_data;
 664         kib_dev_t             *dev;
 665         struct ib_qp_init_attr *init_qp_attr;
 666         struct kib_sched_info   *sched;
 667         kib_conn_t              *conn;
 668         struct ib_cq            *cq;
 669         unsigned long           flags;
 670         int                     cpt;
 671         int                     rc;
 672         int                     i;
 673
 674         LASSERT(net != NULL);
 675         LASSERT(!in_interrupt());
 676
 677         dev = net->ibn_dev;
 678
 679         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 680         sched = kiblnd_data.kib_scheds[cpt];
 681
 682         LASSERT(sched->ibs_nthreads > 0);
 683
 684         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 685                          sizeof(*init_qp_attr));
 686         if (init_qp_attr == NULL) {
 687                 CERROR("Can't allocate qp_attr for %s\n",
 688                        libcfs_nid2str(peer->ibp_nid));
 689                 goto failed_0;
 690         }
 691
 692         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 693         if (conn == NULL) {
 694                 CERROR("Can't allocate connection for %s\n",
 695                        libcfs_nid2str(peer->ibp_nid));
 696                 goto failed_1;
 697         }
 698
 699         conn->ibc_state = IBLND_CONN_INIT;
 700         conn->ibc_version = version;
 701         conn->ibc_peer = peer;            /* I take the caller's ref */
 702         cmid->context = conn;              /* for future CM callbacks */
 703         conn->ibc_cmid = cmid;
 704
 705         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 706         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 707         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 708         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 709         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 710         INIT_LIST_HEAD(&conn->ibc_active_txs);
 711         spin_lock_init(&conn->ibc_lock);
 712
 713         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 714                          sizeof(*conn->ibc_connvars));
 715         if (conn->ibc_connvars == NULL) {
 716                 CERROR("Can't allocate in-progress connection state\n");
 717                 goto failed_2;
 718         }
 719
 720         write_lock_irqsave(glock, flags);
 721         if (dev->ibd_failover) {
 722                 write_unlock_irqrestore(glock, flags);
 723                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 724                 goto failed_2;
 725         }
 726
 727         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 728                 /* wakeup failover thread and teardown connection */
 729                 if (kiblnd_dev_can_failover(dev)) {
 730                         list_add_tail(&dev->ibd_fail_list,
 731                                       &kiblnd_data.kib_failed_devs);
 732                         wake_up(&kiblnd_data.kib_failover_waitq);
 733                 }
 734
 735                 write_unlock_irqrestore(glock, flags);
 736                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 737                        cmid->device->name, dev->ibd_ifname);
 738                 goto failed_2;
 739         }
 740
 741         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 742         conn->ibc_hdev = dev->ibd_hdev;
 743
 744         kiblnd_setup_mtu_locked(cmid);
 745
 746         write_unlock_irqrestore(glock, flags);
 747
 748         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 749                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 750         if (conn->ibc_rxs == NULL) {
 751                 CERROR("Cannot allocate RX buffers\n");
 752                 goto failed_2;
 753         }
 754
 755         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 756                                 IBLND_RX_MSG_PAGES(version));
 757         if (rc != 0)
 758                 goto failed_2;
 759
 760         kiblnd_map_rx_descs(conn);
 761
 762         cq = ib_create_cq(cmid->device,
 763                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 764                           IBLND_CQ_ENTRIES(version),
 765                           kiblnd_get_completion_vector(conn, cpt));
 766         if (IS_ERR(cq)) {
 767                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 768                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 769                 goto failed_2;
 770         }
 771
 772         conn->ibc_cq = cq;
 773
 774         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 775         if (rc != 0) {
 776                 CERROR("Can't request completion notificiation: %d\n", rc);
 777                 goto failed_2;
 778         }
 779
 780         init_qp_attr->event_handler = kiblnd_qp_event;
 781         init_qp_attr->qp_context = conn;
 782         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 783         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 784         init_qp_attr->cap.max_send_sge = 1;
 785         init_qp_attr->cap.max_recv_sge = 1;
 786         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 787         init_qp_attr->qp_type = IB_QPT_RC;
 788         init_qp_attr->send_cq = cq;
 789         init_qp_attr->recv_cq = cq;
 790
 791         conn->ibc_sched = sched;
 792
 793         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 794         if (rc != 0) {
 795                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 796                        rc, init_qp_attr->cap.max_send_wr,
 797                        init_qp_attr->cap.max_recv_wr);
 798                 goto failed_2;
 799         }
 800
 801         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 802
 803         /* 1 ref for caller and each rxmsg */
 804         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 805         conn->ibc_nrx = IBLND_RX_MSGS(version);
 806
 807         /* post receives */
 808         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 809                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 810                                     IBLND_POSTRX_NO_CREDIT);
 811                 if (rc != 0) {
 812                         CERROR("Can't post rxmsg: %d\n", rc);
 813
 814                         /* Make posted receives complete */
 815                         kiblnd_abort_receives(conn);
 816
 817                         /* correct # of posted buffers
 818                          * NB locking needed now I'm racing with completion */
 819                         spin_lock_irqsave(&sched->ibs_lock, flags);
 820                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 821                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 822
 823                         /* cmid will be destroyed by CM(ofed) after cm_callback
 824                          * returned, so we can't refer it anymore
 825                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 826                         rdma_destroy_qp(conn->ibc_cmid);
 827                         conn->ibc_cmid = NULL;
 828
 829                         /* Drop my own and unused rxbuffer refcounts */
 830                         while (i++ <= IBLND_RX_MSGS(version))
 831                                 kiblnd_conn_decref(conn);
 832
 833                         return NULL;
 834                 }
 835         }
 836
 837         /* Init successful! */
 838         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 839                  state == IBLND_CONN_PASSIVE_WAIT);
 840         conn->ibc_state = state;
 841
 842         /* 1 more conn */
 843         atomic_inc(&net->ibn_nconns);
 844         return conn;
 845
 846  failed_2:
 847         kiblnd_destroy_conn(conn);
 848  failed_1:
 849         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 850  failed_0:
 851         return NULL;
 852 }
 853
 854 void
 855 kiblnd_destroy_conn (kib_conn_t *conn)
 856 {
 857         struct rdma_cm_id *cmid = conn->ibc_cmid;
 858         kib_peer_t      *peer = conn->ibc_peer;
 859         int             rc;
 860
 861         LASSERT (!in_interrupt());
 862         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 863         LASSERT (list_empty(&conn->ibc_early_rxs));
 864         LASSERT (list_empty(&conn->ibc_tx_noops));
 865         LASSERT (list_empty(&conn->ibc_tx_queue));
 866         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 867         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 868         LASSERT (list_empty(&conn->ibc_active_txs));
 869         LASSERT (conn->ibc_noops_posted == 0);
 870         LASSERT (conn->ibc_nsends_posted == 0);
 871
 872         switch (conn->ibc_state) {
 873         default:
 874                 /* conn must be completely disengaged from the network */
 875                 LBUG();
 876
 877         case IBLND_CONN_DISCONNECTED:
 878                 /* connvars should have been freed already */
 879                 LASSERT (conn->ibc_connvars == NULL);
 880                 break;
 881
 882         case IBLND_CONN_INIT:
 883                 break;
 884         }
 885
 886         /* conn->ibc_cmid might be destroyed by CM already */
 887         if (cmid != NULL && cmid->qp != NULL)
 888                 rdma_destroy_qp(cmid);
 889
 890         if (conn->ibc_cq != NULL) {
 891                 rc = ib_destroy_cq(conn->ibc_cq);
 892                 if (rc != 0)
 893                         CWARN("Error destroying CQ: %d\n", rc);
 894         }
 895
 896         if (conn->ibc_rx_pages != NULL)
 897                 kiblnd_unmap_rx_descs(conn);
 898
 899         if (conn->ibc_rxs != NULL) {
 900                 LIBCFS_FREE(conn->ibc_rxs,
 901                             IBLND_RX_MSGS(conn->ibc_version) * sizeof(kib_rx_t));
 902         }
 903
 904         if (conn->ibc_connvars != NULL)
 905                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 906
 907         if (conn->ibc_hdev != NULL)
 908                 kiblnd_hdev_decref(conn->ibc_hdev);
 909
 910         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 911         if (conn->ibc_state != IBLND_CONN_INIT) {
 912                 kib_net_t *net = peer->ibp_ni->ni_data;
 913
 914                 kiblnd_peer_decref(peer);
 915                 rdma_destroy_id(cmid);
 916                 atomic_dec(&net->ibn_nconns);
 917         }
 918
 919         LIBCFS_FREE(conn, sizeof(*conn));
 920 }
 921
 922 int
 923 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 924 {
 925         kib_conn_t           *conn;
 926         struct list_head             *ctmp;
 927         struct list_head             *cnxt;
 928         int                  count = 0;
 929
 930         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 931                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 932
 933                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 934                        libcfs_nid2str(peer->ibp_nid),
 935                        conn->ibc_version, why);
 936
 937                 kiblnd_close_conn_locked(conn, why);
 938                 count++;
 939         }
 940
 941         return count;
 942 }
 943
 944 int
 945 kiblnd_close_stale_conns_locked (kib_peer_t *peer,
 946                                  int version, __u64 incarnation)
 947 {
 948         kib_conn_t           *conn;
 949         struct list_head             *ctmp;
 950         struct list_head             *cnxt;
 951         int                  count = 0;
 952
 953         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 954                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 955
 956                 if (conn->ibc_version     == version &&
 957                     conn->ibc_incarnation == incarnation)
 958                         continue;
 959
 960                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 961                        libcfs_nid2str(peer->ibp_nid),
 962                        conn->ibc_version, conn->ibc_incarnation,
 963                        version, incarnation);
 964
 965                 kiblnd_close_conn_locked(conn, -ESTALE);
 966                 count++;
 967         }
 968
 969         return count;
 970 }
 971
 972 static int
 973 kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 974 {
 975         kib_peer_t           *peer;
 976         struct list_head             *ptmp;
 977         struct list_head             *pnxt;
 978         int                  lo;
 979         int                  hi;
 980         int                  i;
 981         unsigned long      flags;
 982         int                  count = 0;
 983
 984         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 985
 986         if (nid != LNET_NID_ANY)
 987                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 988         else {
 989                 lo = 0;
 990                 hi = kiblnd_data.kib_peer_hash_size - 1;
 991         }
 992
 993         for (i = lo; i <= hi; i++) {
 994                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 995
 996                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 997                         LASSERT (peer->ibp_connecting > 0 ||
 998                                  peer->ibp_accepting > 0 ||
 999                                  !list_empty(&peer->ibp_conns));
1000
1001                         if (peer->ibp_ni != ni)
1002                                 continue;
1003
1004                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1005                                 continue;
1006
1007                         count += kiblnd_close_peer_conns_locked(peer, 0);
1008                 }
1009         }
1010
1011         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1012
1013         /* wildcards always succeed */
1014         if (nid == LNET_NID_ANY)
1015                 return 0;
1016
1017         return (count == 0) ? -ENOENT : 0;
1018 }
1019
1020 int
1021 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1022 {
1023         struct libcfs_ioctl_data *data = arg;
1024         int                    rc = -EINVAL;
1025
1026         switch (cmd) {
1027         case IOC_LIBCFS_GET_PEER: {
1028                 lnet_nid_t   nid = 0;
1029                 int       count = 0;
1030
1031                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1032                                           &nid, &count);
1033                 data->ioc_nid    = nid;
1034                 data->ioc_count  = count;
1035                 break;
1036         }
1037
1038         case IOC_LIBCFS_DEL_PEER: {
1039                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1040                 break;
1041         }
1042         case IOC_LIBCFS_GET_CONN: {
1043                 kib_conn_t *conn;
1044
1045                 rc = 0;
1046                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1047                 if (conn == NULL) {
1048                         rc = -ENOENT;
1049                         break;
1050                 }
1051
1052                 LASSERT (conn->ibc_cmid != NULL);
1053                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1054                 if (conn->ibc_cmid->route.path_rec == NULL)
1055                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1056                 else
1057                         data->ioc_u32[0] =
1058                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1059                 kiblnd_conn_decref(conn);
1060                 break;
1061         }
1062         case IOC_LIBCFS_CLOSE_CONNECTION: {
1063                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1064                 break;
1065         }
1066
1067         default:
1068                 break;
1069         }
1070
1071         return rc;
1072 }
1073
1074 void
1075 kiblnd_query (lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1076 {
1077         unsigned long   last_alive = 0;
1078         unsigned long   now = cfs_time_current();
1079         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1080         kib_peer_t      *peer;
1081         unsigned long   flags;
1082
1083         read_lock_irqsave(glock, flags);
1084
1085         peer = kiblnd_find_peer_locked(nid);
1086         if (peer != NULL) {
1087                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
1088                          peer->ibp_accepting > 0 ||
1089                          !list_empty(&peer->ibp_conns));  /* active conn */
1090                 last_alive = peer->ibp_last_alive;
1091         }
1092
1093         read_unlock_irqrestore(glock, flags);
1094
1095         if (last_alive != 0)
1096                 *when = last_alive;
1097
1098         /* peer is not persistent in hash, trigger peer creation
1099          * and connection establishment with a NULL tx */
1100         if (peer == NULL)
1101                 kiblnd_launch_tx(ni, NULL, nid);
1102
1103         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1104                libcfs_nid2str(nid), peer,
1105                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1106         return;
1107 }
1108
1109 void
1110 kiblnd_free_pages(kib_pages_t *p)
1111 {
1112         int     npages = p->ibp_npages;
1113         int     i;
1114
1115         for (i = 0; i < npages; i++) {
1116                 if (p->ibp_pages[i] != NULL)
1117                         __free_page(p->ibp_pages[i]);
1118         }
1119
1120         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1121 }
1122
1123 int
1124 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1125 {
1126         kib_pages_t     *p;
1127         int             i;
1128
1129         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1130                          offsetof(kib_pages_t, ibp_pages[npages]));
1131         if (p == NULL) {
1132                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1133                 return -ENOMEM;
1134         }
1135
1136         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1137         p->ibp_npages = npages;
1138
1139         for (i = 0; i < npages; i++) {
1140                 p->ibp_pages[i] = alloc_pages_node(
1141                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1142                                     GFP_NOFS, 0);
1143                 if (p->ibp_pages[i] == NULL) {
1144                         CERROR("Can't allocate page %d of %d\n", i, npages);
1145                         kiblnd_free_pages(p);
1146                         return -ENOMEM;
1147                 }
1148         }
1149
1150         *pp = p;
1151         return 0;
1152 }
1153
1154 void
1155 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1156 {
1157         kib_rx_t *rx;
1158         int       i;
1159
1160         LASSERT (conn->ibc_rxs != NULL);
1161         LASSERT (conn->ibc_hdev != NULL);
1162
1163         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1164                 rx = &conn->ibc_rxs[i];
1165
1166                 LASSERT (rx->rx_nob >= 0); /* not posted */
1167
1168                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1169                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1170                                                           rx->rx_msgaddr),
1171                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1172         }
1173
1174         kiblnd_free_pages(conn->ibc_rx_pages);
1175
1176         conn->ibc_rx_pages = NULL;
1177 }
1178
1179 void
1180 kiblnd_map_rx_descs(kib_conn_t *conn)
1181 {
1182         kib_rx_t       *rx;
1183         struct page    *pg;
1184         int          pg_off;
1185         int          ipg;
1186         int          i;
1187
1188         for (pg_off = ipg = i = 0;
1189              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1190                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1191                 rx = &conn->ibc_rxs[i];
1192
1193                 rx->rx_conn = conn;
1194                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1195
1196                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1197                                                        rx->rx_msg, IBLND_MSG_SIZE,
1198                                                        DMA_FROM_DEVICE);
1199                 LASSERT (!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1200                                                    rx->rx_msgaddr));
1201                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1202
1203                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1204                        i, rx->rx_msg, rx->rx_msgaddr,
1205                        lnet_page2phys(pg) + pg_off);
1206
1207                 pg_off += IBLND_MSG_SIZE;
1208                 LASSERT (pg_off <= PAGE_SIZE);
1209
1210                 if (pg_off == PAGE_SIZE) {
1211                         pg_off = 0;
1212                         ipg++;
1213                         LASSERT (ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1214                 }
1215         }
1216 }
1217
1218 static void
1219 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1220 {
1221         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1222         kib_tx_t       *tx;
1223         int          i;
1224
1225         LASSERT (tpo->tpo_pool.po_allocated == 0);
1226
1227         if (hdev == NULL)
1228                 return;
1229
1230         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1231                 tx = &tpo->tpo_tx_descs[i];
1232                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1233                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1234                                                           tx->tx_msgaddr),
1235                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1236         }
1237
1238         kiblnd_hdev_decref(hdev);
1239         tpo->tpo_hdev = NULL;
1240 }
1241
1242 static kib_hca_dev_t *
1243 kiblnd_current_hdev(kib_dev_t *dev)
1244 {
1245         kib_hca_dev_t *hdev;
1246         unsigned long  flags;
1247         int         i = 0;
1248
1249         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1250         while (dev->ibd_failover) {
1251                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1252                 if (i++ % 50 == 0)
1253                         CDEBUG(D_NET, "%s: Wait for failover\n",
1254                                dev->ibd_ifname);
1255                 schedule_timeout(cfs_time_seconds(1) / 100);
1256
1257                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1258         }
1259
1260         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1261         hdev = dev->ibd_hdev;
1262
1263         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1264
1265         return hdev;
1266 }
1267
1268 static void
1269 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1270 {
1271         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1272         kib_pool_t     *pool  = &tpo->tpo_pool;
1273         kib_net_t      *net   = pool->po_owner->ps_net;
1274         kib_dev_t      *dev;
1275         struct page    *page;
1276         kib_tx_t       *tx;
1277         int          page_offset;
1278         int          ipage;
1279         int          i;
1280
1281         LASSERT (net != NULL);
1282
1283         dev = net->ibn_dev;
1284
1285         /* pre-mapped messages are not bigger than 1 page */
1286         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1287
1288         /* No fancy arithmetic when we do the buffer calculations */
1289         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1290
1291         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1292
1293         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1294                 page = txpgs->ibp_pages[ipage];
1295                 tx = &tpo->tpo_tx_descs[i];
1296
1297                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1298                                            page_offset);
1299
1300                 tx->tx_msgaddr = kiblnd_dma_map_single(
1301                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1302                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1303                 LASSERT (!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1304                                                    tx->tx_msgaddr));
1305                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1306
1307                 list_add(&tx->tx_list, &pool->po_free_list);
1308
1309                 page_offset += IBLND_MSG_SIZE;
1310                 LASSERT (page_offset <= PAGE_SIZE);
1311
1312                 if (page_offset == PAGE_SIZE) {
1313                         page_offset = 0;
1314                         ipage++;
1315                         LASSERT (ipage <= txpgs->ibp_npages);
1316                 }
1317         }
1318 }
1319
1320 struct ib_mr *
1321 kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1322 {
1323         __u64   index;
1324
1325         LASSERT (hdev->ibh_mrs[0] != NULL);
1326
1327         if (hdev->ibh_nmrs == 1)
1328                 return hdev->ibh_mrs[0];
1329
1330         index = addr >> hdev->ibh_mr_shift;
1331
1332         if (index <  hdev->ibh_nmrs &&
1333             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1334                 return hdev->ibh_mrs[index];
1335
1336         return NULL;
1337 }
1338
1339 struct ib_mr *
1340 kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1341 {
1342         struct ib_mr *prev_mr;
1343         struct ib_mr *mr;
1344         int        i;
1345
1346         LASSERT (hdev->ibh_mrs[0] != NULL);
1347
1348         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1349             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1350                 return NULL;
1351
1352         if (hdev->ibh_nmrs == 1)
1353                 return hdev->ibh_mrs[0];
1354
1355         for (i = 0, mr = prev_mr = NULL;
1356              i < rd->rd_nfrags; i++) {
1357                 mr = kiblnd_find_dma_mr(hdev,
1358                                         rd->rd_frags[i].rf_addr,
1359                                         rd->rd_frags[i].rf_nob);
1360                 if (prev_mr == NULL)
1361                         prev_mr = mr;
1362
1363                 if (mr == NULL || prev_mr != mr) {
1364                         /* Can't covered by one single MR */
1365                         mr = NULL;
1366                         break;
1367                 }
1368         }
1369
1370         return mr;
1371 }
1372
1373 static void
1374 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1375 {
1376         LASSERT (pool->fpo_map_count == 0);
1377
1378         if (pool->fpo_fmr_pool != NULL)
1379                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1380
1381         if (pool->fpo_hdev != NULL)
1382                 kiblnd_hdev_decref(pool->fpo_hdev);
1383
1384         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1385 }
1386
1387 static void
1388 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1389 {
1390         kib_fmr_pool_t *pool;
1391
1392         while (!list_empty(head)) {
1393                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1394                 list_del(&pool->fpo_list);
1395                 kiblnd_destroy_fmr_pool(pool);
1396         }
1397 }
1398
1399 static int kiblnd_fmr_pool_size(int ncpts)
1400 {
1401         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1402
1403         return max(IBLND_FMR_POOL, size);
1404 }
1405
1406 static int kiblnd_fmr_flush_trigger(int ncpts)
1407 {
1408         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1409
1410         return max(IBLND_FMR_POOL_FLUSH, size);
1411 }
1412
1413 static int
1414 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1415 {
1416         /* FMR pool for RDMA */
1417         kib_dev_t              *dev = fps->fps_net->ibn_dev;
1418         kib_fmr_pool_t    *fpo;
1419         struct ib_fmr_pool_param param = {
1420                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1421                 .page_shift     = PAGE_SHIFT,
1422                 .access     = (IB_ACCESS_LOCAL_WRITE |
1423                                       IB_ACCESS_REMOTE_WRITE),
1424                 .pool_size         = fps->fps_pool_size,
1425                 .dirty_watermark   = fps->fps_flush_trigger,
1426                 .flush_function    = NULL,
1427                 .flush_arg       = NULL,
1428                 .cache       = !!*kiblnd_tunables.kib_fmr_cache};
1429         int rc;
1430
1431         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1432         if (fpo == NULL)
1433                 return -ENOMEM;
1434
1435         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1436
1437         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1438         if (IS_ERR(fpo->fpo_fmr_pool)) {
1439                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1440                 CERROR("Failed to create FMR pool: %d\n", rc);
1441
1442                 kiblnd_hdev_decref(fpo->fpo_hdev);
1443                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1444                 return rc;
1445         }
1446
1447         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1448         fpo->fpo_owner    = fps;
1449         *pp_fpo = fpo;
1450
1451         return 0;
1452 }
1453
1454 static void
1455 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1456 {
1457         if (fps->fps_net == NULL) /* intialized? */
1458                 return;
1459
1460         spin_lock(&fps->fps_lock);
1461
1462         while (!list_empty(&fps->fps_pool_list)) {
1463                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1464                                                  kib_fmr_pool_t, fpo_list);
1465                 fpo->fpo_failed = 1;
1466                 list_del(&fpo->fpo_list);
1467                 if (fpo->fpo_map_count == 0)
1468                         list_add(&fpo->fpo_list, zombies);
1469                 else
1470                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1471         }
1472
1473         spin_unlock(&fps->fps_lock);
1474 }
1475
1476 static void
1477 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1478 {
1479         if (fps->fps_net != NULL) { /* initialized? */
1480                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1481                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1482         }
1483 }
1484
1485 static int
1486 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, kib_net_t *net,
1487                         int pool_size, int flush_trigger)
1488 {
1489         kib_fmr_pool_t *fpo;
1490         int          rc;
1491
1492         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1493
1494         fps->fps_net = net;
1495         fps->fps_cpt = cpt;
1496         fps->fps_pool_size = pool_size;
1497         fps->fps_flush_trigger = flush_trigger;
1498         spin_lock_init(&fps->fps_lock);
1499         INIT_LIST_HEAD(&fps->fps_pool_list);
1500         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1501
1502         rc = kiblnd_create_fmr_pool(fps, &fpo);
1503         if (rc == 0)
1504                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1505
1506         return rc;
1507 }
1508
1509 static int
1510 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1511 {
1512         if (fpo->fpo_map_count != 0) /* still in use */
1513                 return 0;
1514         if (fpo->fpo_failed)
1515                 return 1;
1516         return cfs_time_aftereq(now, fpo->fpo_deadline);
1517 }
1518
1519 void
1520 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1521 {
1522         LIST_HEAD     (zombies);
1523         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1524         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1525         unsigned long    now = cfs_time_current();
1526         kib_fmr_pool_t    *tmp;
1527         int             rc;
1528
1529         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1530         LASSERT (rc == 0);
1531
1532         if (status != 0) {
1533                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1534                 LASSERT (rc == 0);
1535         }
1536
1537         fmr->fmr_pool = NULL;
1538         fmr->fmr_pfmr = NULL;
1539
1540         spin_lock(&fps->fps_lock);
1541         fpo->fpo_map_count--;  /* decref the pool */
1542
1543         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1544                 /* the first pool is persistent */
1545                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1546                         continue;
1547
1548                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1549                         list_move(&fpo->fpo_list, &zombies);
1550                         fps->fps_version++;
1551                 }
1552         }
1553         spin_unlock(&fps->fps_lock);
1554
1555         if (!list_empty(&zombies))
1556                 kiblnd_destroy_fmr_pool_list(&zombies);
1557 }
1558
1559 int
1560 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1561                     __u64 iov, kib_fmr_t *fmr)
1562 {
1563         struct ib_pool_fmr *pfmr;
1564         kib_fmr_pool_t     *fpo;
1565         __u64          version;
1566         int              rc;
1567
1568  again:
1569         spin_lock(&fps->fps_lock);
1570         version = fps->fps_version;
1571         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1572                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1573                 fpo->fpo_map_count++;
1574                 spin_unlock(&fps->fps_lock);
1575
1576                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1577                                             pages, npages, iov);
1578                 if (likely(!IS_ERR(pfmr))) {
1579                         fmr->fmr_pool = fpo;
1580                         fmr->fmr_pfmr = pfmr;
1581                         return 0;
1582                 }
1583
1584                 spin_lock(&fps->fps_lock);
1585                 fpo->fpo_map_count--;
1586                 if (PTR_ERR(pfmr) != -EAGAIN) {
1587                         spin_unlock(&fps->fps_lock);
1588                         return PTR_ERR(pfmr);
1589                 }
1590
1591                 /* EAGAIN and ... */
1592                 if (version != fps->fps_version) {
1593                         spin_unlock(&fps->fps_lock);
1594                         goto again;
1595                 }
1596         }
1597
1598         if (fps->fps_increasing) {
1599                 spin_unlock(&fps->fps_lock);
1600                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1601                 schedule();
1602                 goto again;
1603
1604         }
1605
1606         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1607                 /* someone failed recently */
1608                 spin_unlock(&fps->fps_lock);
1609                 return -EAGAIN;
1610         }
1611
1612         fps->fps_increasing = 1;
1613         spin_unlock(&fps->fps_lock);
1614
1615         CDEBUG(D_NET, "Allocate new FMR pool\n");
1616         rc = kiblnd_create_fmr_pool(fps, &fpo);
1617         spin_lock(&fps->fps_lock);
1618         fps->fps_increasing = 0;
1619         if (rc == 0) {
1620                 fps->fps_version++;
1621                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1622         } else {
1623                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1624         }
1625         spin_unlock(&fps->fps_lock);
1626
1627         goto again;
1628 }
1629
1630 static void
1631 kiblnd_fini_pool(kib_pool_t *pool)
1632 {
1633         LASSERT (list_empty(&pool->po_free_list));
1634         LASSERT (pool->po_allocated == 0);
1635
1636         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1637 }
1638
1639 static void
1640 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1641 {
1642         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1643
1644         memset(pool, 0, sizeof(kib_pool_t));
1645         INIT_LIST_HEAD(&pool->po_free_list);
1646         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1647         pool->po_owner    = ps;
1648         pool->po_size     = size;
1649 }
1650
1651 static void
1652 kiblnd_destroy_pool_list(struct list_head *head)
1653 {
1654         kib_pool_t *pool;
1655
1656         while (!list_empty(head)) {
1657                 pool = list_entry(head->next, kib_pool_t, po_list);
1658                 list_del(&pool->po_list);
1659
1660                 LASSERT (pool->po_owner != NULL);
1661                 pool->po_owner->ps_pool_destroy(pool);
1662         }
1663 }
1664
1665 static void
1666 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1667 {
1668         if (ps->ps_net == NULL) /* intialized? */
1669                 return;
1670
1671         spin_lock(&ps->ps_lock);
1672         while (!list_empty(&ps->ps_pool_list)) {
1673                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1674                                             kib_pool_t, po_list);
1675                 po->po_failed = 1;
1676                 list_del(&po->po_list);
1677                 if (po->po_allocated == 0)
1678                         list_add(&po->po_list, zombies);
1679                 else
1680                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1681         }
1682         spin_unlock(&ps->ps_lock);
1683 }
1684
1685 static void
1686 kiblnd_fini_poolset(kib_poolset_t *ps)
1687 {
1688         if (ps->ps_net != NULL) { /* initialized? */
1689                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1690                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1691         }
1692 }
1693
1694 static int
1695 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1696                     kib_net_t *net, char *name, int size,
1697                     kib_ps_pool_create_t po_create,
1698                     kib_ps_pool_destroy_t po_destroy,
1699                     kib_ps_node_init_t nd_init,
1700                     kib_ps_node_fini_t nd_fini)
1701 {
1702         kib_pool_t      *pool;
1703         int             rc;
1704
1705         memset(ps, 0, sizeof(kib_poolset_t));
1706
1707         ps->ps_cpt          = cpt;
1708         ps->ps_net        = net;
1709         ps->ps_pool_create  = po_create;
1710         ps->ps_pool_destroy = po_destroy;
1711         ps->ps_node_init    = nd_init;
1712         ps->ps_node_fini    = nd_fini;
1713         ps->ps_pool_size    = size;
1714         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1715             >= sizeof(ps->ps_name))
1716                 return -E2BIG;
1717         spin_lock_init(&ps->ps_lock);
1718         INIT_LIST_HEAD(&ps->ps_pool_list);
1719         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1720
1721         rc = ps->ps_pool_create(ps, size, &pool);
1722         if (rc == 0)
1723                 list_add(&pool->po_list, &ps->ps_pool_list);
1724         else
1725                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1726
1727         return rc;
1728 }
1729
1730 static int
1731 kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1732 {
1733         if (pool->po_allocated != 0) /* still in use */
1734                 return 0;
1735         if (pool->po_failed)
1736                 return 1;
1737         return cfs_time_aftereq(now, pool->po_deadline);
1738 }
1739
1740 void
1741 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1742 {
1743         LIST_HEAD  (zombies);
1744         kib_poolset_t  *ps = pool->po_owner;
1745         kib_pool_t     *tmp;
1746         unsigned long      now = cfs_time_current();
1747
1748         spin_lock(&ps->ps_lock);
1749
1750         if (ps->ps_node_fini != NULL)
1751                 ps->ps_node_fini(pool, node);
1752
1753         LASSERT (pool->po_allocated > 0);
1754         list_add(node, &pool->po_free_list);
1755         pool->po_allocated--;
1756
1757         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1758                 /* the first pool is persistent */
1759                 if (ps->ps_pool_list.next == &pool->po_list)
1760                         continue;
1761
1762                 if (kiblnd_pool_is_idle(pool, now))
1763                         list_move(&pool->po_list, &zombies);
1764         }
1765         spin_unlock(&ps->ps_lock);
1766
1767         if (!list_empty(&zombies))
1768                 kiblnd_destroy_pool_list(&zombies);
1769 }
1770
1771 struct list_head *
1772 kiblnd_pool_alloc_node(kib_poolset_t *ps)
1773 {
1774         struct list_head            *node;
1775         kib_pool_t          *pool;
1776         int                 rc;
1777
1778  again:
1779         spin_lock(&ps->ps_lock);
1780         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1781                 if (list_empty(&pool->po_free_list))
1782                         continue;
1783
1784                 pool->po_allocated++;
1785                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1786                 node = pool->po_free_list.next;
1787                 list_del(node);
1788
1789                 if (ps->ps_node_init != NULL) {
1790                         /* still hold the lock */
1791                         ps->ps_node_init(pool, node);
1792                 }
1793                 spin_unlock(&ps->ps_lock);
1794                 return node;
1795         }
1796
1797         /* no available tx pool and ... */
1798         if (ps->ps_increasing) {
1799                 /* another thread is allocating a new pool */
1800                 spin_unlock(&ps->ps_lock);
1801                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1802                        ps->ps_name);
1803                 schedule();
1804                 goto again;
1805         }
1806
1807         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1808                 /* someone failed recently */
1809                 spin_unlock(&ps->ps_lock);
1810                 return NULL;
1811         }
1812
1813         ps->ps_increasing = 1;
1814         spin_unlock(&ps->ps_lock);
1815
1816         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1817
1818         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1819
1820         spin_lock(&ps->ps_lock);
1821         ps->ps_increasing = 0;
1822         if (rc == 0) {
1823                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1824         } else {
1825                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1826                 CERROR("Can't allocate new %s pool because out of memory\n",
1827                        ps->ps_name);
1828         }
1829         spin_unlock(&ps->ps_lock);
1830
1831         goto again;
1832 }
1833
1834 void
1835 kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1836 {
1837         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1838         struct ib_mr    *mr  = pmr->pmr_mr;
1839
1840         pmr->pmr_mr = NULL;
1841         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1842         if (mr != NULL)
1843                 ib_dereg_mr(mr);
1844 }
1845
1846 int
1847 kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1848                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1849 {
1850         kib_phys_mr_t *pmr;
1851         struct list_head    *node;
1852         int         rc;
1853         int         i;
1854
1855         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1856         if (node == NULL) {
1857                 CERROR("Failed to allocate PMR descriptor\n");
1858                 return -ENOMEM;
1859         }
1860
1861         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1862         if (pmr->pmr_pool->ppo_hdev != hdev) {
1863                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1864                 return -EAGAIN;
1865         }
1866
1867         for (i = 0; i < rd->rd_nfrags; i++) {
1868                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1869                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1870         }
1871
1872         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1873                                      pmr->pmr_ipb, rd->rd_nfrags,
1874                                      IB_ACCESS_LOCAL_WRITE |
1875                                      IB_ACCESS_REMOTE_WRITE,
1876                                      iova);
1877         if (!IS_ERR(pmr->pmr_mr)) {
1878                 pmr->pmr_iova = *iova;
1879                 *pp_pmr = pmr;
1880                 return 0;
1881         }
1882
1883         rc = PTR_ERR(pmr->pmr_mr);
1884         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1885
1886         pmr->pmr_mr = NULL;
1887         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1888
1889         return rc;
1890 }
1891
1892 static void
1893 kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1894 {
1895         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1896         kib_phys_mr_t  *pmr;
1897
1898         LASSERT (pool->po_allocated == 0);
1899
1900         while (!list_empty(&pool->po_free_list)) {
1901                 pmr = list_entry(pool->po_free_list.next,
1902                                      kib_phys_mr_t, pmr_list);
1903
1904                 LASSERT (pmr->pmr_mr == NULL);
1905                 list_del(&pmr->pmr_list);
1906
1907                 if (pmr->pmr_ipb != NULL) {
1908                         LIBCFS_FREE(pmr->pmr_ipb,
1909                                     IBLND_MAX_RDMA_FRAGS *
1910                                     sizeof(struct ib_phys_buf));
1911                 }
1912
1913                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1914         }
1915
1916         kiblnd_fini_pool(pool);
1917         if (ppo->ppo_hdev != NULL)
1918                 kiblnd_hdev_decref(ppo->ppo_hdev);
1919
1920         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1921 }
1922
1923 static inline int kiblnd_pmr_pool_size(int ncpts)
1924 {
1925         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1926
1927         return max(IBLND_PMR_POOL, size);
1928 }
1929
1930 static int
1931 kiblnd_create_pmr_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
1932 {
1933         struct kib_pmr_pool     *ppo;
1934         struct kib_pool         *pool;
1935         kib_phys_mr_t           *pmr;
1936         int                     i;
1937
1938         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
1939                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
1940         if (ppo == NULL) {
1941                 CERROR("Failed to allocate PMR pool\n");
1942                 return -ENOMEM;
1943         }
1944
1945         pool = &ppo->ppo_pool;
1946         kiblnd_init_pool(ps, pool, size);
1947
1948         for (i = 0; i < size; i++) {
1949                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
1950                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
1951                 if (pmr == NULL)
1952                         break;
1953
1954                 pmr->pmr_pool = ppo;
1955                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
1956                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
1957                 if (pmr->pmr_ipb == NULL)
1958                         break;
1959
1960                 list_add(&pmr->pmr_list, &pool->po_free_list);
1961         }
1962
1963         if (i < size) {
1964                 ps->ps_pool_destroy(pool);
1965                 return -ENOMEM;
1966         }
1967
1968         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1969         *pp_po = pool;
1970         return 0;
1971 }
1972
1973 static void
1974 kiblnd_destroy_tx_pool(kib_pool_t *pool)
1975 {
1976         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1977         int          i;
1978
1979         LASSERT (pool->po_allocated == 0);
1980
1981         if (tpo->tpo_tx_pages != NULL) {
1982                 kiblnd_unmap_tx_pool(tpo);
1983                 kiblnd_free_pages(tpo->tpo_tx_pages);
1984         }
1985
1986         if (tpo->tpo_tx_descs == NULL)
1987                 goto out;
1988
1989         for (i = 0; i < pool->po_size; i++) {
1990                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1991
1992                 list_del(&tx->tx_list);
1993                 if (tx->tx_pages != NULL)
1994                         LIBCFS_FREE(tx->tx_pages,
1995                                     LNET_MAX_IOV *
1996                                     sizeof(*tx->tx_pages));
1997                 if (tx->tx_frags != NULL)
1998                         LIBCFS_FREE(tx->tx_frags,
1999                                     IBLND_MAX_RDMA_FRAGS *
2000                                             sizeof(*tx->tx_frags));
2001                 if (tx->tx_wrq != NULL)
2002                         LIBCFS_FREE(tx->tx_wrq,
2003                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2004                                     sizeof(*tx->tx_wrq));
2005                 if (tx->tx_sge != NULL)
2006                         LIBCFS_FREE(tx->tx_sge,
2007                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2008                                     sizeof(*tx->tx_sge));
2009                 if (tx->tx_rd != NULL)
2010                         LIBCFS_FREE(tx->tx_rd,
2011                                     offsetof(kib_rdma_desc_t,
2012                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2013         }
2014
2015         LIBCFS_FREE(tpo->tpo_tx_descs,
2016                     pool->po_size * sizeof(kib_tx_t));
2017 out:
2018         kiblnd_fini_pool(pool);
2019         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2020 }
2021
2022 static int kiblnd_tx_pool_size(int ncpts)
2023 {
2024         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2025
2026         return max(IBLND_TX_POOL, ntx);
2027 }
2028
2029 static int
2030 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2031 {
2032         int         i;
2033         int         npg;
2034         kib_pool_t    *pool;
2035         kib_tx_pool_t *tpo;
2036
2037         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2038         if (tpo == NULL) {
2039                 CERROR("Failed to allocate TX pool\n");
2040                 return -ENOMEM;
2041         }
2042
2043         pool = &tpo->tpo_pool;
2044         kiblnd_init_pool(ps, pool, size);
2045         tpo->tpo_tx_descs = NULL;
2046         tpo->tpo_tx_pages = NULL;
2047
2048         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2049         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2050                 CERROR("Can't allocate tx pages: %d\n", npg);
2051                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2052                 return -ENOMEM;
2053         }
2054
2055         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2056                          size * sizeof(kib_tx_t));
2057         if (tpo->tpo_tx_descs == NULL) {
2058                 CERROR("Can't allocate %d tx descriptors\n", size);
2059                 ps->ps_pool_destroy(pool);
2060                 return -ENOMEM;
2061         }
2062
2063         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2064
2065         for (i = 0; i < size; i++) {
2066                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2067
2068                 tx->tx_pool = tpo;
2069                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2070                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2071                                          lnet_cpt_table(), ps->ps_cpt,
2072                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2073                         if (tx->tx_pages == NULL)
2074                                 break;
2075                 }
2076
2077                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2078                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2079                 if (tx->tx_frags == NULL)
2080                         break;
2081
2082                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2083
2084                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2085                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2086                                  sizeof(*tx->tx_wrq));
2087                 if (tx->tx_wrq == NULL)
2088                         break;
2089
2090                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2091                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2092                                  sizeof(*tx->tx_sge));
2093                 if (tx->tx_sge == NULL)
2094                         break;
2095
2096                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2097                                  offsetof(kib_rdma_desc_t,
2098                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2099                 if (tx->tx_rd == NULL)
2100                         break;
2101         }
2102
2103         if (i == size) {
2104                 kiblnd_map_tx_pool(tpo);
2105                 *pp_po = pool;
2106                 return 0;
2107         }
2108
2109         ps->ps_pool_destroy(pool);
2110         return -ENOMEM;
2111 }
2112
2113 static void
2114 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2115 {
2116         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2117                                              tps_poolset);
2118         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2119
2120         tx->tx_cookie = tps->tps_next_tx_cookie++;
2121 }
2122
2123 static void
2124 kiblnd_net_fini_pools(kib_net_t *net)
2125 {
2126         int     i;
2127
2128         cfs_cpt_for_each(i, lnet_cpt_table()) {
2129                 kib_tx_poolset_t        *tps;
2130                 kib_fmr_poolset_t       *fps;
2131                 kib_pmr_poolset_t       *pps;
2132
2133                 if (net->ibn_tx_ps != NULL) {
2134                         tps = net->ibn_tx_ps[i];
2135                         kiblnd_fini_poolset(&tps->tps_poolset);
2136                 }
2137
2138                 if (net->ibn_fmr_ps != NULL) {
2139                         fps = net->ibn_fmr_ps[i];
2140                         kiblnd_fini_fmr_poolset(fps);
2141                 }
2142
2143                 if (net->ibn_pmr_ps != NULL) {
2144                         pps = net->ibn_pmr_ps[i];
2145                         kiblnd_fini_poolset(&pps->pps_poolset);
2146                 }
2147         }
2148
2149         if (net->ibn_tx_ps != NULL) {
2150                 cfs_percpt_free(net->ibn_tx_ps);
2151                 net->ibn_tx_ps = NULL;
2152         }
2153
2154         if (net->ibn_fmr_ps != NULL) {
2155                 cfs_percpt_free(net->ibn_fmr_ps);
2156                 net->ibn_fmr_ps = NULL;
2157         }
2158
2159         if (net->ibn_pmr_ps != NULL) {
2160                 cfs_percpt_free(net->ibn_pmr_ps);
2161                 net->ibn_pmr_ps = NULL;
2162         }
2163 }
2164
2165 static int
2166 kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2167 {
2168         unsigned long   flags;
2169         int             cpt;
2170         int             rc;
2171         int             i;
2172
2173         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2174         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2175             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2176                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2177                                            flags);
2178                 goto create_tx_pool;
2179         }
2180
2181         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2182
2183         if (*kiblnd_tunables.kib_fmr_pool_size <
2184             *kiblnd_tunables.kib_ntx / 4) {
2185                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2186                        *kiblnd_tunables.kib_fmr_pool_size,
2187                        *kiblnd_tunables.kib_ntx / 4);
2188                 rc = -EINVAL;
2189                 goto failed;
2190         }
2191
2192         /* TX pool must be created later than FMR/PMR, see LU-2268
2193          * for details */
2194         LASSERT(net->ibn_tx_ps == NULL);
2195
2196         /* premapping can fail if ibd_nmr > 1, so we always create
2197          * FMR/PMR pool and map-on-demand if premapping failed */
2198
2199         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2200                                            sizeof(kib_fmr_poolset_t));
2201         if (net->ibn_fmr_ps == NULL) {
2202                 CERROR("Failed to allocate FMR pool array\n");
2203                 rc = -ENOMEM;
2204                 goto failed;
2205         }
2206
2207         for (i = 0; i < ncpts; i++) {
2208                 cpt = (cpts == NULL) ? i : cpts[i];
2209                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2210                                              kiblnd_fmr_pool_size(ncpts),
2211                                              kiblnd_fmr_flush_trigger(ncpts));
2212                 if (rc == -ENOSYS && i == 0) /* no FMR */
2213                         break; /* create PMR pool */
2214
2215                 if (rc != 0) { /* a real error */
2216                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2217                                cpt, rc);
2218                         goto failed;
2219                 }
2220         }
2221
2222         if (i > 0) {
2223                 LASSERT(i == ncpts);
2224                 goto create_tx_pool;
2225         }
2226
2227         cfs_percpt_free(net->ibn_fmr_ps);
2228         net->ibn_fmr_ps = NULL;
2229
2230         CWARN("Device does not support FMR, failing back to PMR\n");
2231
2232         if (*kiblnd_tunables.kib_pmr_pool_size <
2233             *kiblnd_tunables.kib_ntx / 4) {
2234                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2235                        *kiblnd_tunables.kib_pmr_pool_size,
2236                        *kiblnd_tunables.kib_ntx / 4);
2237                 rc = -EINVAL;
2238                 goto failed;
2239         }
2240
2241         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2242                                            sizeof(kib_pmr_poolset_t));
2243         if (net->ibn_pmr_ps == NULL) {
2244                 CERROR("Failed to allocate PMR pool array\n");
2245                 rc = -ENOMEM;
2246                 goto failed;
2247         }
2248
2249         for (i = 0; i < ncpts; i++) {
2250                 cpt = (cpts == NULL) ? i : cpts[i];
2251                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2252                                          cpt, net, "PMR",
2253                                          kiblnd_pmr_pool_size(ncpts),
2254                                          kiblnd_create_pmr_pool,
2255                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2256                 if (rc != 0) {
2257                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2258                                cpt, rc);
2259                         goto failed;
2260                 }
2261         }
2262
2263  create_tx_pool:
2264         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2265                                           sizeof(kib_tx_poolset_t));
2266         if (net->ibn_tx_ps == NULL) {
2267                 CERROR("Failed to allocate tx pool array\n");
2268                 rc = -ENOMEM;
2269                 goto failed;
2270         }
2271
2272         for (i = 0; i < ncpts; i++) {
2273                 cpt = (cpts == NULL) ? i : cpts[i];
2274                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2275                                          cpt, net, "TX",
2276                                          kiblnd_tx_pool_size(ncpts),
2277                                          kiblnd_create_tx_pool,
2278                                          kiblnd_destroy_tx_pool,
2279                                          kiblnd_tx_init, NULL);
2280                 if (rc != 0) {
2281                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2282                                cpt, rc);
2283                         goto failed;
2284                 }
2285         }
2286
2287         return 0;
2288  failed:
2289         kiblnd_net_fini_pools(net);
2290         LASSERT(rc != 0);
2291         return rc;
2292 }
2293
2294 static int
2295 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2296 {
2297         struct ib_device_attr *attr;
2298         int                 rc;
2299
2300         /* It's safe to assume a HCA can handle a page size
2301          * matching that of the native system */
2302         hdev->ibh_page_shift = PAGE_SHIFT;
2303         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2304         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2305
2306         LIBCFS_ALLOC(attr, sizeof(*attr));
2307         if (attr == NULL) {
2308                 CERROR("Out of memory\n");
2309                 return -ENOMEM;
2310         }
2311
2312         rc = ib_query_device(hdev->ibh_ibdev, attr);
2313         if (rc == 0)
2314                 hdev->ibh_mr_size = attr->max_mr_size;
2315
2316         LIBCFS_FREE(attr, sizeof(*attr));
2317
2318         if (rc != 0) {
2319                 CERROR("Failed to query IB device: %d\n", rc);
2320                 return rc;
2321         }
2322
2323         if (hdev->ibh_mr_size == ~0ULL) {
2324                 hdev->ibh_mr_shift = 64;
2325                 return 0;
2326         }
2327
2328         for (hdev->ibh_mr_shift = 0;
2329              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2330                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2331                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2332                         return 0;
2333         }
2334
2335         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2336         return -EINVAL;
2337 }
2338
2339 static void
2340 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2341 {
2342         int     i;
2343
2344         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2345                 return;
2346
2347         for (i = 0; i < hdev->ibh_nmrs; i++) {
2348                 if (hdev->ibh_mrs[i] == NULL)
2349                         break;
2350
2351                 ib_dereg_mr(hdev->ibh_mrs[i]);
2352         }
2353
2354         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2355         hdev->ibh_mrs  = NULL;
2356         hdev->ibh_nmrs = 0;
2357 }
2358
2359 void
2360 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2361 {
2362         kiblnd_hdev_cleanup_mrs(hdev);
2363
2364         if (hdev->ibh_pd != NULL)
2365                 ib_dealloc_pd(hdev->ibh_pd);
2366
2367         if (hdev->ibh_cmid != NULL)
2368                 rdma_destroy_id(hdev->ibh_cmid);
2369
2370         LIBCFS_FREE(hdev, sizeof(*hdev));
2371 }
2372
2373 static int
2374 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2375 {
2376         struct ib_mr *mr;
2377         int        i;
2378         int        rc;
2379         __u64    mm_size;
2380         __u64    mr_size;
2381         int        acflags = IB_ACCESS_LOCAL_WRITE |
2382                                 IB_ACCESS_REMOTE_WRITE;
2383
2384         rc = kiblnd_hdev_get_attr(hdev);
2385         if (rc != 0)
2386                 return rc;
2387
2388         if (hdev->ibh_mr_shift == 64) {
2389                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2390                 if (hdev->ibh_mrs == NULL) {
2391                         CERROR("Failed to allocate MRs table\n");
2392                         return -ENOMEM;
2393                 }
2394
2395                 hdev->ibh_mrs[0] = NULL;
2396                 hdev->ibh_nmrs   = 1;
2397
2398                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2399                 if (IS_ERR(mr)) {
2400                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2401                         kiblnd_hdev_cleanup_mrs(hdev);
2402                         return PTR_ERR(mr);
2403                 }
2404
2405                 hdev->ibh_mrs[0] = mr;
2406
2407                 goto out;
2408         }
2409
2410         mr_size = 1ULL << hdev->ibh_mr_shift;
2411         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2412
2413         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2414
2415         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2416                 /* it's 4T..., assume we will re-code at that time */
2417                 CERROR("Can't support memory size: x%#llx with MR size: x%#llx\n",
2418                        mm_size, mr_size);
2419                 return -EINVAL;
2420         }
2421
2422         /* create an array of MRs to cover all memory */
2423         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2424         if (hdev->ibh_mrs == NULL) {
2425                 CERROR("Failed to allocate MRs' table\n");
2426                 return -ENOMEM;
2427         }
2428
2429         for (i = 0; i < hdev->ibh_nmrs; i++) {
2430                 struct ib_phys_buf ipb;
2431                 __u64         iova;
2432
2433                 ipb.size = hdev->ibh_mr_size;
2434                 ipb.addr = i * mr_size;
2435                 iova     = ipb.addr;
2436
2437                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2438                 if (IS_ERR(mr)) {
2439                         CERROR("Failed ib_reg_phys_mr addr %#llx size %#llx : %ld\n",
2440                                ipb.addr, ipb.size, PTR_ERR(mr));
2441                         kiblnd_hdev_cleanup_mrs(hdev);
2442                         return PTR_ERR(mr);
2443                 }
2444
2445                 LASSERT (iova == ipb.addr);
2446
2447                 hdev->ibh_mrs[i] = mr;
2448         }
2449
2450 out:
2451         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2452                 LCONSOLE_INFO("Register global MR array, MR size: %#llx, array size: %d\n",
2453                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2454         return 0;
2455 }
2456
2457 static int
2458 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2459 {       /* DUMMY */
2460         return 0;
2461 }
2462
2463 static int
2464 kiblnd_dev_need_failover(kib_dev_t *dev)
2465 {
2466         struct rdma_cm_id  *cmid;
2467         struct sockaddr_in  srcaddr;
2468         struct sockaddr_in  dstaddr;
2469         int              rc;
2470
2471         if (dev->ibd_hdev == NULL || /* initializing */
2472             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2473             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2474                 return 1;
2475
2476         /* XXX: it's UGLY, but I don't have better way to find
2477          * ib-bonding HCA failover because:
2478          *
2479          * a. no reliable CM event for HCA failover...
2480          * b. no OFED API to get ib_device for current net_device...
2481          *
2482          * We have only two choices at this point:
2483          *
2484          * a. rdma_bind_addr(), it will conflict with listener cmid
2485          * b. rdma_resolve_addr() to zero addr */
2486         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2487                                      IB_QPT_RC);
2488         if (IS_ERR(cmid)) {
2489                 rc = PTR_ERR(cmid);
2490                 CERROR("Failed to create cmid for failover: %d\n", rc);
2491                 return rc;
2492         }
2493
2494         memset(&srcaddr, 0, sizeof(srcaddr));
2495         srcaddr.sin_family      = AF_INET;
2496         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2497
2498         memset(&dstaddr, 0, sizeof(dstaddr));
2499         dstaddr.sin_family = AF_INET;
2500         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2501                                (struct sockaddr *)&dstaddr, 1);
2502         if (rc != 0 || cmid->device == NULL) {
2503                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2504                        dev->ibd_ifname, &dev->ibd_ifip,
2505                        cmid->device, rc);
2506                 rdma_destroy_id(cmid);
2507                 return rc;
2508         }
2509
2510         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2511                 /* don't need device failover */
2512                 rdma_destroy_id(cmid);
2513                 return 0;
2514         }
2515
2516         return 1;
2517 }
2518
2519 int
2520 kiblnd_dev_failover(kib_dev_t *dev)
2521 {
2522         LIST_HEAD      (zombie_tpo);
2523         LIST_HEAD      (zombie_ppo);
2524         LIST_HEAD      (zombie_fpo);
2525         struct rdma_cm_id  *cmid  = NULL;
2526         kib_hca_dev_t      *hdev  = NULL;
2527         kib_hca_dev_t      *old;
2528         struct ib_pd       *pd;
2529         kib_net_t         *net;
2530         struct sockaddr_in  addr;
2531         unsigned long       flags;
2532         int              rc = 0;
2533         int                 i;
2534
2535         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2536                  dev->ibd_can_failover ||
2537                  dev->ibd_hdev == NULL);
2538
2539         rc = kiblnd_dev_need_failover(dev);
2540         if (rc <= 0)
2541                 goto out;
2542
2543         if (dev->ibd_hdev != NULL &&
2544             dev->ibd_hdev->ibh_cmid != NULL) {
2545                 /* XXX it's not good to close old listener at here,
2546                  * because we can fail to create new listener.
2547                  * But we have to close it now, otherwise rdma_bind_addr
2548                  * will return EADDRINUSE... How crap! */
2549                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2550
2551                 cmid = dev->ibd_hdev->ibh_cmid;
2552                 /* make next schedule of kiblnd_dev_need_failover()
2553                  * return 1 for me */
2554                 dev->ibd_hdev->ibh_cmid  = NULL;
2555                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2556
2557                 rdma_destroy_id(cmid);
2558         }
2559
2560         cmid = kiblnd_rdma_create_id(kiblnd_cm_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                 goto out;
2566         }
2567
2568         memset(&addr, 0, sizeof(addr));
2569         addr.sin_family      = AF_INET;
2570         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2571         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2572
2573         /* Bind to failover device or port */
2574         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2575         if (rc != 0 || cmid->device == NULL) {
2576                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2577                        dev->ibd_ifname, &dev->ibd_ifip,
2578                        cmid->device, rc);
2579                 rdma_destroy_id(cmid);
2580                 goto out;
2581         }
2582
2583         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2584         if (hdev == NULL) {
2585                 CERROR("Failed to allocate kib_hca_dev\n");
2586                 rdma_destroy_id(cmid);
2587                 rc = -ENOMEM;
2588                 goto out;
2589         }
2590
2591         atomic_set(&hdev->ibh_ref, 1);
2592         hdev->ibh_dev   = dev;
2593         hdev->ibh_cmid  = cmid;
2594         hdev->ibh_ibdev = cmid->device;
2595
2596         pd = ib_alloc_pd(cmid->device);
2597         if (IS_ERR(pd)) {
2598                 rc = PTR_ERR(pd);
2599                 CERROR("Can't allocate PD: %d\n", rc);
2600                 goto out;
2601         }
2602
2603         hdev->ibh_pd = pd;
2604
2605         rc = rdma_listen(cmid, 0);
2606         if (rc != 0) {
2607                 CERROR("Can't start new listener: %d\n", rc);
2608                 goto out;
2609         }
2610
2611         rc = kiblnd_hdev_setup_mrs(hdev);
2612         if (rc != 0) {
2613                 CERROR("Can't setup device: %d\n", rc);
2614                 goto out;
2615         }
2616
2617         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2618
2619         old = dev->ibd_hdev;
2620         dev->ibd_hdev = hdev; /* take over the refcount */
2621         hdev = old;
2622
2623         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2624                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2625                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2626                                             &zombie_tpo);
2627
2628                         if (net->ibn_fmr_ps != NULL) {
2629                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2630                                                         &zombie_fpo);
2631
2632                         } else if (net->ibn_pmr_ps != NULL) {
2633                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2634                                                     pps_poolset, &zombie_ppo);
2635                         }
2636                 }
2637         }
2638
2639         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2640  out:
2641         if (!list_empty(&zombie_tpo))
2642                 kiblnd_destroy_pool_list(&zombie_tpo);
2643         if (!list_empty(&zombie_ppo))
2644                 kiblnd_destroy_pool_list(&zombie_ppo);
2645         if (!list_empty(&zombie_fpo))
2646                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2647         if (hdev != NULL)
2648                 kiblnd_hdev_decref(hdev);
2649
2650         if (rc != 0)
2651                 dev->ibd_failed_failover++;
2652         else
2653                 dev->ibd_failed_failover = 0;
2654
2655         return rc;
2656 }
2657
2658 void
2659 kiblnd_destroy_dev (kib_dev_t *dev)
2660 {
2661         LASSERT (dev->ibd_nnets == 0);
2662         LASSERT (list_empty(&dev->ibd_nets));
2663
2664         list_del(&dev->ibd_fail_list);
2665         list_del(&dev->ibd_list);
2666
2667         if (dev->ibd_hdev != NULL)
2668                 kiblnd_hdev_decref(dev->ibd_hdev);
2669
2670         LIBCFS_FREE(dev, sizeof(*dev));
2671 }
2672
2673 static kib_dev_t *
2674 kiblnd_create_dev(char *ifname)
2675 {
2676         struct net_device *netdev;
2677         kib_dev_t        *dev;
2678         __u32         netmask;
2679         __u32         ip;
2680         int             up;
2681         int             rc;
2682
2683         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2684         if (rc != 0) {
2685                 CERROR("Can't query IPoIB interface %s: %d\n",
2686                        ifname, rc);
2687                 return NULL;
2688         }
2689
2690         if (!up) {
2691                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2692                 return NULL;
2693         }
2694
2695         LIBCFS_ALLOC(dev, sizeof(*dev));
2696         if (dev == NULL)
2697                 return NULL;
2698
2699         netdev = dev_get_by_name(&init_net, ifname);
2700         if (netdev == NULL) {
2701                 dev->ibd_can_failover = 0;
2702         } else {
2703                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2704                 dev_put(netdev);
2705         }
2706
2707         INIT_LIST_HEAD(&dev->ibd_nets);
2708         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2709         INIT_LIST_HEAD(&dev->ibd_fail_list);
2710         dev->ibd_ifip = ip;
2711         strcpy(&dev->ibd_ifname[0], ifname);
2712
2713         /* initialize the device */
2714         rc = kiblnd_dev_failover(dev);
2715         if (rc != 0) {
2716                 CERROR("Can't initialize device: %d\n", rc);
2717                 LIBCFS_FREE(dev, sizeof(*dev));
2718                 return NULL;
2719         }
2720
2721         list_add_tail(&dev->ibd_list,
2722                           &kiblnd_data.kib_devs);
2723         return dev;
2724 }
2725
2726 static void
2727 kiblnd_base_shutdown(void)
2728 {
2729         struct kib_sched_info   *sched;
2730         int                     i;
2731
2732         LASSERT (list_empty(&kiblnd_data.kib_devs));
2733
2734         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2735                atomic_read(&libcfs_kmemory));
2736
2737         switch (kiblnd_data.kib_init) {
2738         default:
2739                 LBUG();
2740
2741         case IBLND_INIT_ALL:
2742         case IBLND_INIT_DATA:
2743                 LASSERT (kiblnd_data.kib_peers != NULL);
2744                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2745                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
2746                 }
2747                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
2748                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
2749
2750                 /* flag threads to terminate; wake and wait for them to die */
2751                 kiblnd_data.kib_shutdown = 1;
2752
2753                 /* NB: we really want to stop scheduler threads net by net
2754                  * instead of the whole module, this should be improved
2755                  * with dynamic configuration LNet */
2756                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2757                         wake_up_all(&sched->ibs_waitq);
2758
2759                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2760                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2761
2762                 i = 2;
2763                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2764                         i++;
2765                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2766                                "Waiting for %d threads to terminate\n",
2767                                atomic_read(&kiblnd_data.kib_nthreads));
2768                         set_current_state(TASK_UNINTERRUPTIBLE);
2769                         schedule_timeout(cfs_time_seconds(1));
2770                 }
2771
2772                 /* fall through */
2773
2774         case IBLND_INIT_NOTHING:
2775                 break;
2776         }
2777
2778         if (kiblnd_data.kib_peers != NULL) {
2779                 LIBCFS_FREE(kiblnd_data.kib_peers,
2780                             sizeof(struct list_head) *
2781                             kiblnd_data.kib_peer_hash_size);
2782         }
2783
2784         if (kiblnd_data.kib_scheds != NULL)
2785                 cfs_percpt_free(kiblnd_data.kib_scheds);
2786
2787         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2788                atomic_read(&libcfs_kmemory));
2789
2790         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2791         module_put(THIS_MODULE);
2792 }
2793
2794 void
2795 kiblnd_shutdown (lnet_ni_t *ni)
2796 {
2797         kib_net_t       *net = ni->ni_data;
2798         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2799         int            i;
2800         unsigned long     flags;
2801
2802         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2803
2804         if (net == NULL)
2805                 goto out;
2806
2807         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2808                atomic_read(&libcfs_kmemory));
2809
2810         write_lock_irqsave(g_lock, flags);
2811         net->ibn_shutdown = 1;
2812         write_unlock_irqrestore(g_lock, flags);
2813
2814         switch (net->ibn_init) {
2815         default:
2816                 LBUG();
2817
2818         case IBLND_INIT_ALL:
2819                 /* nuke all existing peers within this net */
2820                 kiblnd_del_peer(ni, LNET_NID_ANY);
2821
2822                 /* Wait for all peer state to clean up */
2823                 i = 2;
2824                 while (atomic_read(&net->ibn_npeers) != 0) {
2825                         i++;
2826                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2827                                "%s: waiting for %d peers to disconnect\n",
2828                                libcfs_nid2str(ni->ni_nid),
2829                                atomic_read(&net->ibn_npeers));
2830                         set_current_state(TASK_UNINTERRUPTIBLE);
2831                         schedule_timeout(cfs_time_seconds(1));
2832                 }
2833
2834                 kiblnd_net_fini_pools(net);
2835
2836                 write_lock_irqsave(g_lock, flags);
2837                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2838                 net->ibn_dev->ibd_nnets--;
2839                 list_del(&net->ibn_list);
2840                 write_unlock_irqrestore(g_lock, flags);
2841
2842                 /* fall through */
2843
2844         case IBLND_INIT_NOTHING:
2845                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2846
2847                 if (net->ibn_dev != NULL &&
2848                     net->ibn_dev->ibd_nnets == 0)
2849                         kiblnd_destroy_dev(net->ibn_dev);
2850
2851                 break;
2852         }
2853
2854         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2855                atomic_read(&libcfs_kmemory));
2856
2857         net->ibn_init = IBLND_INIT_NOTHING;
2858         ni->ni_data = NULL;
2859
2860         LIBCFS_FREE(net, sizeof(*net));
2861
2862 out:
2863         if (list_empty(&kiblnd_data.kib_devs))
2864                 kiblnd_base_shutdown();
2865         return;
2866 }
2867
2868 static int
2869 kiblnd_base_startup(void)
2870 {
2871         struct kib_sched_info   *sched;
2872         int                     rc;
2873         int                     i;
2874
2875         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2876
2877         try_module_get(THIS_MODULE);
2878         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
2879
2880         rwlock_init(&kiblnd_data.kib_global_lock);
2881
2882         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2883         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2884
2885         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2886         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2887                      sizeof(struct list_head) *
2888                             kiblnd_data.kib_peer_hash_size);
2889         if (kiblnd_data.kib_peers == NULL) {
2890                 goto failed;
2891         }
2892         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2893                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2894
2895         spin_lock_init(&kiblnd_data.kib_connd_lock);
2896         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2897         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2898         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2899         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2900
2901         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2902                                                   sizeof(*sched));
2903         if (kiblnd_data.kib_scheds == NULL)
2904                 goto failed;
2905
2906         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2907                 int     nthrs;
2908
2909                 spin_lock_init(&sched->ibs_lock);
2910                 INIT_LIST_HEAD(&sched->ibs_conns);
2911                 init_waitqueue_head(&sched->ibs_waitq);
2912
2913                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2914                 if (*kiblnd_tunables.kib_nscheds > 0) {
2915                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2916                 } else {
2917                         /* max to half of CPUs, another half is reserved for
2918                          * upper layer modules */
2919                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2920                 }
2921
2922                 sched->ibs_nthreads_max = nthrs;
2923                 sched->ibs_cpt = i;
2924         }
2925
2926         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2927
2928         /* lists/ptrs/locks initialised */
2929         kiblnd_data.kib_init = IBLND_INIT_DATA;
2930         /*****************************************************/
2931
2932         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2933         if (rc != 0) {
2934                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2935                 goto failed;
2936         }
2937
2938         if (*kiblnd_tunables.kib_dev_failover != 0)
2939                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2940                                          "kiblnd_failover");
2941
2942         if (rc != 0) {
2943                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2944                 goto failed;
2945         }
2946
2947         /* flag everything initialised */
2948         kiblnd_data.kib_init = IBLND_INIT_ALL;
2949         /*****************************************************/
2950
2951         return 0;
2952
2953  failed:
2954         kiblnd_base_shutdown();
2955         return -ENETDOWN;
2956 }
2957
2958 static int
2959 kiblnd_start_schedulers(struct kib_sched_info *sched)
2960 {
2961         int     rc = 0;
2962         int     nthrs;
2963         int     i;
2964
2965         if (sched->ibs_nthreads == 0) {
2966                 if (*kiblnd_tunables.kib_nscheds > 0) {
2967                         nthrs = sched->ibs_nthreads_max;
2968                 } else {
2969                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2970                                                sched->ibs_cpt);
2971                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2972                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2973                 }
2974         } else {
2975                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2976                 /* increase one thread if there is new interface */
2977                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
2978         }
2979
2980         for (i = 0; i < nthrs; i++) {
2981                 long    id;
2982                 char    name[20];
2983                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2984                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2985                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2986                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2987                 if (rc == 0)
2988                         continue;
2989
2990                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2991                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2992                 break;
2993         }
2994
2995         sched->ibs_nthreads += i;
2996         return rc;
2997 }
2998
2999 static int
3000 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3001 {
3002         int     cpt;
3003         int     rc;
3004         int     i;
3005
3006         for (i = 0; i < ncpts; i++) {
3007                 struct kib_sched_info *sched;
3008
3009                 cpt = (cpts == NULL) ? i : cpts[i];
3010                 sched = kiblnd_data.kib_scheds[cpt];
3011
3012                 if (!newdev && sched->ibs_nthreads > 0)
3013                         continue;
3014
3015                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3016                 if (rc != 0) {
3017                         CERROR("Failed to start scheduler threads for %s\n",
3018                                dev->ibd_ifname);
3019                         return rc;
3020                 }
3021         }
3022         return 0;
3023 }
3024
3025 static kib_dev_t *
3026 kiblnd_dev_search(char *ifname)
3027 {
3028         kib_dev_t       *alias = NULL;
3029         kib_dev_t       *dev;
3030         char            *colon;
3031         char            *colon2;
3032
3033         colon = strchr(ifname, ':');
3034         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3035                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3036                         return dev;
3037
3038                 if (alias != NULL)
3039                         continue;
3040
3041                 colon2 = strchr(dev->ibd_ifname, ':');
3042                 if (colon != NULL)
3043                         *colon = 0;
3044                 if (colon2 != NULL)
3045                         *colon2 = 0;
3046
3047                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3048                         alias = dev;
3049
3050                 if (colon != NULL)
3051                         *colon = ':';
3052                 if (colon2 != NULL)
3053                         *colon2 = ':';
3054         }
3055         return alias;
3056 }
3057
3058 int
3059 kiblnd_startup (lnet_ni_t *ni)
3060 {
3061         char                 *ifname;
3062         kib_dev_t               *ibdev = NULL;
3063         kib_net_t               *net;
3064         struct timeval      tv;
3065         unsigned long        flags;
3066         int                    rc;
3067         int                       newdev;
3068
3069         LASSERT (ni->ni_lnd == &the_o2iblnd);
3070
3071         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3072                 rc = kiblnd_base_startup();
3073                 if (rc != 0)
3074                         return rc;
3075         }
3076
3077         LIBCFS_ALLOC(net, sizeof(*net));
3078         ni->ni_data = net;
3079         if (net == NULL)
3080                 goto net_failed;
3081
3082         do_gettimeofday(&tv);
3083         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3084
3085         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3086         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3087         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3088         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3089
3090         if (ni->ni_interfaces[0] != NULL) {
3091                 /* Use the IPoIB interface specified in 'networks=' */
3092
3093                 CLASSERT (LNET_MAX_INTERFACES > 1);
3094                 if (ni->ni_interfaces[1] != NULL) {
3095                         CERROR("Multiple interfaces not supported\n");
3096                         goto failed;
3097                 }
3098
3099                 ifname = ni->ni_interfaces[0];
3100         } else {
3101                 ifname = *kiblnd_tunables.kib_default_ipif;
3102         }
3103
3104         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3105                 CERROR("IPoIB interface name too long: %s\n", ifname);
3106                 goto failed;
3107         }
3108
3109         ibdev = kiblnd_dev_search(ifname);
3110
3111         newdev = ibdev == NULL;
3112         /* hmm...create kib_dev even for alias */
3113         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3114                 ibdev = kiblnd_create_dev(ifname);
3115
3116         if (ibdev == NULL)
3117                 goto failed;
3118
3119         net->ibn_dev = ibdev;
3120         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3121
3122         rc = kiblnd_dev_start_threads(ibdev, newdev,
3123                                       ni->ni_cpts, ni->ni_ncpts);
3124         if (rc != 0)
3125                 goto failed;
3126
3127         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3128         if (rc != 0) {
3129                 CERROR("Failed to initialize NI pools: %d\n", rc);
3130                 goto failed;
3131         }
3132
3133         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3134         ibdev->ibd_nnets++;
3135         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3136         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3137
3138         net->ibn_init = IBLND_INIT_ALL;
3139
3140         return 0;
3141
3142 failed:
3143         if (net->ibn_dev == NULL && ibdev != NULL)
3144                 kiblnd_destroy_dev(ibdev);
3145
3146 net_failed:
3147         kiblnd_shutdown(ni);
3148
3149         CDEBUG(D_NET, "kiblnd_startup failed\n");
3150         return -ENETDOWN;
3151 }
3152
3153 static void __exit
3154 kiblnd_module_fini (void)
3155 {
3156         lnet_unregister_lnd(&the_o2iblnd);
3157 }
3158
3159 static int __init
3160 kiblnd_module_init (void)
3161 {
3162         int    rc;
3163
3164         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3165         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3166                   <= IBLND_MSG_SIZE);
3167         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3168                   <= IBLND_MSG_SIZE);
3169
3170         rc = kiblnd_tunables_init();
3171         if (rc != 0)
3172                 return rc;
3173
3174         lnet_register_lnd(&the_o2iblnd);
3175
3176         return 0;
3177 }
3178
3179 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3180 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3181 MODULE_LICENSE("GPL");
3182
3183 module_init(kiblnd_module_init);
3184 module_exit(kiblnd_module_fini);