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 kiblnd_cksum(void *ptr, int nob)
  57 {
  58         char *c = ptr;
  59         __u32 sum = 0;
  60
  61         while (nob-- > 0)
  62                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  63
  64         /* ensure I don't return 0 (== no checksum) */
  65         return (sum == 0) ? 1 : sum;
  66 }
  67
  68 static char *kiblnd_msgtype2str(int type)
  69 {
  70         switch (type) {
  71         case IBLND_MSG_CONNREQ:
  72                 return "CONNREQ";
  73
  74         case IBLND_MSG_CONNACK:
  75                 return "CONNACK";
  76
  77         case IBLND_MSG_NOOP:
  78                 return "NOOP";
  79
  80         case IBLND_MSG_IMMEDIATE:
  81                 return "IMMEDIATE";
  82
  83         case IBLND_MSG_PUT_REQ:
  84                 return "PUT_REQ";
  85
  86         case IBLND_MSG_PUT_NAK:
  87                 return "PUT_NAK";
  88
  89         case IBLND_MSG_PUT_ACK:
  90                 return "PUT_ACK";
  91
  92         case IBLND_MSG_PUT_DONE:
  93                 return "PUT_DONE";
  94
  95         case IBLND_MSG_GET_REQ:
  96                 return "GET_REQ";
  97
  98         case IBLND_MSG_GET_DONE:
  99                 return "GET_DONE";
 100
 101         default:
 102                 return "???";
 103         }
 104 }
 105
 106 static int kiblnd_msgtype2size(int type)
 107 {
 108         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 109
 110         switch (type) {
 111         case IBLND_MSG_CONNREQ:
 112         case IBLND_MSG_CONNACK:
 113                 return hdr_size + sizeof(kib_connparams_t);
 114
 115         case IBLND_MSG_NOOP:
 116                 return hdr_size;
 117
 118         case IBLND_MSG_IMMEDIATE:
 119                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 120
 121         case IBLND_MSG_PUT_REQ:
 122                 return hdr_size + sizeof(kib_putreq_msg_t);
 123
 124         case IBLND_MSG_PUT_ACK:
 125                 return hdr_size + sizeof(kib_putack_msg_t);
 126
 127         case IBLND_MSG_GET_REQ:
 128                 return hdr_size + sizeof(kib_get_msg_t);
 129
 130         case IBLND_MSG_PUT_NAK:
 131         case IBLND_MSG_PUT_DONE:
 132         case IBLND_MSG_GET_DONE:
 133                 return hdr_size + sizeof(kib_completion_msg_t);
 134         default:
 135                 return -1;
 136         }
 137 }
 138
 139 static int kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 140 {
 141         kib_rdma_desc_t *rd;
 142         int nob;
 143         int n;
 144         int i;
 145
 146         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 147                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 148
 149         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 150                               &msg->ibm_u.get.ibgm_rd :
 151                               &msg->ibm_u.putack.ibpam_rd;
 152
 153         if (flip) {
 154                 __swab32s(&rd->rd_key);
 155                 __swab32s(&rd->rd_nfrags);
 156         }
 157
 158         n = rd->rd_nfrags;
 159
 160         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 161                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 162                        n, IBLND_MAX_RDMA_FRAGS);
 163                 return 1;
 164         }
 165
 166         nob = offsetof(kib_msg_t, ibm_u) +
 167               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 168
 169         if (msg->ibm_nob < nob) {
 170                 CERROR("Short %s: %d(%d)\n",
 171                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 172                 return 1;
 173         }
 174
 175         if (!flip)
 176                 return 0;
 177
 178         for (i = 0; i < n; i++) {
 179                 __swab32s(&rd->rd_frags[i].rf_nob);
 180                 __swab64s(&rd->rd_frags[i].rf_addr);
 181         }
 182
 183         return 0;
 184 }
 185
 186 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
 187                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 188 {
 189         kib_net_t *net = ni->ni_data;
 190
 191         /* CAVEAT EMPTOR! all message fields not set here should have been
 192          * initialised previously. */
 193         msg->ibm_magic    = IBLND_MSG_MAGIC;
 194         msg->ibm_version  = version;
 195         /*   ibm_type */
 196         msg->ibm_credits  = credits;
 197         /*   ibm_nob */
 198         msg->ibm_cksum    = 0;
 199         msg->ibm_srcnid   = ni->ni_nid;
 200         msg->ibm_srcstamp = net->ibn_incarnation;
 201         msg->ibm_dstnid   = dstnid;
 202         msg->ibm_dststamp = dststamp;
 203
 204         if (*kiblnd_tunables.kib_cksum) {
 205                 /* NB ibm_cksum zero while computing cksum */
 206                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 207         }
 208 }
 209
 210 int kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 211 {
 212         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 213         __u32 msg_cksum;
 214         __u16 version;
 215         int msg_nob;
 216         int flip;
 217
 218         /* 6 bytes are enough to have received magic + version */
 219         if (nob < 6) {
 220                 CERROR("Short message: %d\n", nob);
 221                 return -EPROTO;
 222         }
 223
 224         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 225                 flip = 0;
 226         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 227                 flip = 1;
 228         } else {
 229                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 230                 return -EPROTO;
 231         }
 232
 233         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 234         if (version != IBLND_MSG_VERSION &&
 235             version != IBLND_MSG_VERSION_1) {
 236                 CERROR("Bad version: %x\n", version);
 237                 return -EPROTO;
 238         }
 239
 240         if (nob < hdr_size) {
 241                 CERROR("Short message: %d\n", nob);
 242                 return -EPROTO;
 243         }
 244
 245         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 246         if (msg_nob > nob) {
 247                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 248                 return -EPROTO;
 249         }
 250
 251         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 252          * gets flipped */
 253         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 254         msg->ibm_cksum = 0;
 255         if (msg_cksum != 0 &&
 256             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 257                 CERROR("Bad checksum\n");
 258                 return -EPROTO;
 259         }
 260
 261         msg->ibm_cksum = msg_cksum;
 262
 263         if (flip) {
 264                 /* leave magic unflipped as a clue to peer endianness */
 265                 msg->ibm_version = version;
 266                 CLASSERT(sizeof(msg->ibm_type) == 1);
 267                 CLASSERT(sizeof(msg->ibm_credits) == 1);
 268                 msg->ibm_nob     = msg_nob;
 269                 __swab64s(&msg->ibm_srcnid);
 270                 __swab64s(&msg->ibm_srcstamp);
 271                 __swab64s(&msg->ibm_dstnid);
 272                 __swab64s(&msg->ibm_dststamp);
 273         }
 274
 275         if (msg->ibm_srcnid == LNET_NID_ANY) {
 276                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 277                 return -EPROTO;
 278         }
 279
 280         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 281                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 282                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 283                 return -EPROTO;
 284         }
 285
 286         switch (msg->ibm_type) {
 287         default:
 288                 CERROR("Unknown message type %x\n", msg->ibm_type);
 289                 return -EPROTO;
 290
 291         case IBLND_MSG_NOOP:
 292         case IBLND_MSG_IMMEDIATE:
 293         case IBLND_MSG_PUT_REQ:
 294                 break;
 295
 296         case IBLND_MSG_PUT_ACK:
 297         case IBLND_MSG_GET_REQ:
 298                 if (kiblnd_unpack_rd(msg, flip))
 299                         return -EPROTO;
 300                 break;
 301
 302         case IBLND_MSG_PUT_NAK:
 303         case IBLND_MSG_PUT_DONE:
 304         case IBLND_MSG_GET_DONE:
 305                 if (flip)
 306                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 307                 break;
 308
 309         case IBLND_MSG_CONNREQ:
 310         case IBLND_MSG_CONNACK:
 311                 if (flip) {
 312                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 313                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 314                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 315                 }
 316                 break;
 317         }
 318         return 0;
 319 }
 320
 321 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 322 {
 323         kib_peer_t *peer;
 324         kib_net_t *net = ni->ni_data;
 325         int cpt = lnet_cpt_of_nid(nid);
 326         unsigned long flags;
 327
 328         LASSERT(net != NULL);
 329         LASSERT(nid != LNET_NID_ANY);
 330
 331         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 332         if (peer == NULL) {
 333                 CERROR("Cannot allocate peer\n");
 334                 return -ENOMEM;
 335         }
 336
 337         memset(peer, 0, sizeof(*peer));  /* zero flags etc */
 338
 339         peer->ibp_ni = ni;
 340         peer->ibp_nid = nid;
 341         peer->ibp_error = 0;
 342         peer->ibp_last_alive = 0;
 343         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 344
 345         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 346         INIT_LIST_HEAD(&peer->ibp_conns);
 347         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 348
 349         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 350
 351         /* always called with a ref on ni, which prevents ni being shutdown */
 352         LASSERT(net->ibn_shutdown == 0);
 353
 354         /* npeers only grows with the global lock held */
 355         atomic_inc(&net->ibn_npeers);
 356
 357         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 358
 359         *peerp = peer;
 360         return 0;
 361 }
 362
 363 void kiblnd_destroy_peer(kib_peer_t *peer)
 364 {
 365         kib_net_t *net = peer->ibp_ni->ni_data;
 366
 367         LASSERT(net != NULL);
 368         LASSERT(atomic_read(&peer->ibp_refcount) == 0);
 369         LASSERT(!kiblnd_peer_active(peer));
 370         LASSERT(peer->ibp_connecting == 0);
 371         LASSERT(peer->ibp_accepting == 0);
 372         LASSERT(list_empty(&peer->ibp_conns));
 373         LASSERT(list_empty(&peer->ibp_tx_queue));
 374
 375         LIBCFS_FREE(peer, sizeof(*peer));
 376
 377         /* NB a peer's connections keep a reference on their peer until
 378          * they are destroyed, so we can be assured that _all_ state to do
 379          * with this peer has been cleaned up when its refcount drops to
 380          * zero. */
 381         atomic_dec(&net->ibn_npeers);
 382 }
 383
 384 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
 385 {
 386         /* the caller is responsible for accounting the additional reference
 387          * that this creates */
 388         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 389         struct list_head *tmp;
 390         kib_peer_t *peer;
 391
 392         list_for_each(tmp, peer_list) {
 393
 394                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 395
 396                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
 397                          peer->ibp_accepting > 0 ||
 398                          !list_empty(&peer->ibp_conns));  /* active conn */
 399
 400                 if (peer->ibp_nid != nid)
 401                         continue;
 402
 403                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 404                        peer, libcfs_nid2str(nid),
 405                        atomic_read(&peer->ibp_refcount),
 406                        peer->ibp_version);
 407                 return peer;
 408         }
 409         return NULL;
 410 }
 411
 412 void kiblnd_unlink_peer_locked(kib_peer_t *peer)
 413 {
 414         LASSERT(list_empty(&peer->ibp_conns));
 415
 416         LASSERT(kiblnd_peer_active(peer));
 417         list_del_init(&peer->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer);
 420 }
 421
 422 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 423                                 lnet_nid_t *nidp, int *count)
 424 {
 425         kib_peer_t *peer;
 426         struct list_head *ptmp;
 427         int i;
 428         unsigned long flags;
 429
 430         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 431
 432         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 433
 434                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 435
 436                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 437                         LASSERT(peer->ibp_connecting > 0 ||
 438                                  peer->ibp_accepting > 0 ||
 439                                  !list_empty(&peer->ibp_conns));
 440
 441                         if (peer->ibp_ni != ni)
 442                                 continue;
 443
 444                         if (index-- > 0)
 445                                 continue;
 446
 447                         *nidp = peer->ibp_nid;
 448                         *count = atomic_read(&peer->ibp_refcount);
 449
 450                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 451                                                flags);
 452                         return 0;
 453                 }
 454         }
 455
 456         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 457         return -ENOENT;
 458 }
 459
 460 static void kiblnd_del_peer_locked(kib_peer_t *peer)
 461 {
 462         struct list_head *ctmp;
 463         struct list_head *cnxt;
 464         kib_conn_t *conn;
 465
 466         if (list_empty(&peer->ibp_conns)) {
 467                 kiblnd_unlink_peer_locked(peer);
 468         } else {
 469                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 470                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 471
 472                         kiblnd_close_conn_locked(conn, 0);
 473                 }
 474                 /* NB closing peer's last conn unlinked it. */
 475         }
 476         /* NB peer now unlinked; might even be freed if the peer table had the
 477          * last ref on it. */
 478 }
 479
 480 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 481 {
 482         LIST_HEAD(zombies);
 483         struct list_head *ptmp;
 484         struct list_head *pnxt;
 485         kib_peer_t *peer;
 486         int lo;
 487         int hi;
 488         int i;
 489         unsigned long flags;
 490         int rc = -ENOENT;
 491
 492         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 493
 494         if (nid != LNET_NID_ANY) {
 495                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 496         } else {
 497                 lo = 0;
 498                 hi = kiblnd_data.kib_peer_hash_size - 1;
 499         }
 500
 501         for (i = lo; i <= hi; i++) {
 502                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 503                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 504                         LASSERT(peer->ibp_connecting > 0 ||
 505                                  peer->ibp_accepting > 0 ||
 506                                  !list_empty(&peer->ibp_conns));
 507
 508                         if (peer->ibp_ni != ni)
 509                                 continue;
 510
 511                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 512                                 continue;
 513
 514                         if (!list_empty(&peer->ibp_tx_queue)) {
 515                                 LASSERT(list_empty(&peer->ibp_conns));
 516
 517                                 list_splice_init(&peer->ibp_tx_queue,
 518                                                      &zombies);
 519                         }
 520
 521                         kiblnd_del_peer_locked(peer);
 522                         rc = 0;  /* matched something */
 523                 }
 524         }
 525
 526         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 527
 528         kiblnd_txlist_done(ni, &zombies, -EIO);
 529
 530         return rc;
 531 }
 532
 533 static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 534 {
 535         kib_peer_t *peer;
 536         struct list_head *ptmp;
 537         kib_conn_t *conn;
 538         struct list_head *ctmp;
 539         int i;
 540         unsigned long flags;
 541
 542         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 543
 544         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 545                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 546
 547                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 548                         LASSERT(peer->ibp_connecting > 0 ||
 549                                  peer->ibp_accepting > 0 ||
 550                                  !list_empty(&peer->ibp_conns));
 551
 552                         if (peer->ibp_ni != ni)
 553                                 continue;
 554
 555                         list_for_each(ctmp, &peer->ibp_conns) {
 556                                 if (index-- > 0)
 557                                         continue;
 558
 559                                 conn = list_entry(ctmp, kib_conn_t,
 560                                                       ibc_list);
 561                                 kiblnd_conn_addref(conn);
 562                                 read_unlock_irqrestore(
 563                                         &kiblnd_data.kib_global_lock,
 564                                         flags);
 565                                 return conn;
 566                         }
 567                 }
 568         }
 569
 570         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 571         return NULL;
 572 }
 573
 574 int kiblnd_translate_mtu(int value)
 575 {
 576         switch (value) {
 577         default:
 578                 return -1;
 579         case 0:
 580                 return 0;
 581         case 256:
 582                 return IB_MTU_256;
 583         case 512:
 584                 return IB_MTU_512;
 585         case 1024:
 586                 return IB_MTU_1024;
 587         case 2048:
 588                 return IB_MTU_2048;
 589         case 4096:
 590                 return IB_MTU_4096;
 591         }
 592 }
 593
 594 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 595 {
 596         int mtu;
 597
 598         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 599         if (cmid->route.path_rec == NULL)
 600                 return;
 601
 602         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 603         LASSERT(mtu >= 0);
 604         if (mtu != 0)
 605                 cmid->route.path_rec->mtu = mtu;
 606 }
 607
 608 static int kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 609 {
 610         cpumask_t *mask;
 611         int vectors;
 612         int off;
 613         int i;
 614         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 615
 616         vectors = conn->ibc_cmid->device->num_comp_vectors;
 617         if (vectors <= 1)
 618                 return 0;
 619
 620         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 621         if (mask == NULL)
 622                 return 0;
 623
 624         /* hash NID to CPU id in this partition... */
 625         off = do_div(nid, cpumask_weight(mask));
 626         for_each_cpu(i, mask) {
 627                 if (off-- == 0)
 628                         return i % vectors;
 629         }
 630
 631         LBUG();
 632         return 1;
 633 }
 634
 635 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 636                                 int state, int version)
 637 {
 638         /* CAVEAT EMPTOR:
 639          * If the new conn is created successfully it takes over the caller's
 640          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 641          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 642          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 643          * to destroy 'cmid' here since I'm called from the CM which still has
 644          * its ref on 'cmid'). */
 645         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 646         kib_net_t *net = peer->ibp_ni->ni_data;
 647         kib_dev_t *dev;
 648         struct ib_qp_init_attr *init_qp_attr;
 649         struct kib_sched_info *sched;
 650         kib_conn_t *conn;
 651         struct ib_cq *cq;
 652         unsigned long flags;
 653         int cpt;
 654         int rc;
 655         int i;
 656
 657         LASSERT(net != NULL);
 658         LASSERT(!in_interrupt());
 659
 660         dev = net->ibn_dev;
 661
 662         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 663         sched = kiblnd_data.kib_scheds[cpt];
 664
 665         LASSERT(sched->ibs_nthreads > 0);
 666
 667         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 668                          sizeof(*init_qp_attr));
 669         if (init_qp_attr == NULL) {
 670                 CERROR("Can't allocate qp_attr for %s\n",
 671                        libcfs_nid2str(peer->ibp_nid));
 672                 goto failed_0;
 673         }
 674
 675         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 676         if (conn == NULL) {
 677                 CERROR("Can't allocate connection for %s\n",
 678                        libcfs_nid2str(peer->ibp_nid));
 679                 goto failed_1;
 680         }
 681
 682         conn->ibc_state = IBLND_CONN_INIT;
 683         conn->ibc_version = version;
 684         conn->ibc_peer = peer;            /* I take the caller's ref */
 685         cmid->context = conn;              /* for future CM callbacks */
 686         conn->ibc_cmid = cmid;
 687
 688         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 689         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 690         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 691         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 692         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 693         INIT_LIST_HEAD(&conn->ibc_active_txs);
 694         spin_lock_init(&conn->ibc_lock);
 695
 696         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 697                          sizeof(*conn->ibc_connvars));
 698         if (conn->ibc_connvars == NULL) {
 699                 CERROR("Can't allocate in-progress connection state\n");
 700                 goto failed_2;
 701         }
 702
 703         write_lock_irqsave(glock, flags);
 704         if (dev->ibd_failover) {
 705                 write_unlock_irqrestore(glock, flags);
 706                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 707                 goto failed_2;
 708         }
 709
 710         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 711                 /* wakeup failover thread and teardown connection */
 712                 if (kiblnd_dev_can_failover(dev)) {
 713                         list_add_tail(&dev->ibd_fail_list,
 714                                       &kiblnd_data.kib_failed_devs);
 715                         wake_up(&kiblnd_data.kib_failover_waitq);
 716                 }
 717
 718                 write_unlock_irqrestore(glock, flags);
 719                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 720                        cmid->device->name, dev->ibd_ifname);
 721                 goto failed_2;
 722         }
 723
 724         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 725         conn->ibc_hdev = dev->ibd_hdev;
 726
 727         kiblnd_setup_mtu_locked(cmid);
 728
 729         write_unlock_irqrestore(glock, flags);
 730
 731         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 732                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 733         if (conn->ibc_rxs == NULL) {
 734                 CERROR("Cannot allocate RX buffers\n");
 735                 goto failed_2;
 736         }
 737
 738         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 739                                 IBLND_RX_MSG_PAGES(version));
 740         if (rc != 0)
 741                 goto failed_2;
 742
 743         kiblnd_map_rx_descs(conn);
 744
 745         cq = ib_create_cq(cmid->device,
 746                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 747                           IBLND_CQ_ENTRIES(version),
 748                           kiblnd_get_completion_vector(conn, cpt));
 749         if (IS_ERR(cq)) {
 750                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 751                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 752                 goto failed_2;
 753         }
 754
 755         conn->ibc_cq = cq;
 756
 757         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 758         if (rc != 0) {
 759                 CERROR("Can't request completion notificiation: %d\n", rc);
 760                 goto failed_2;
 761         }
 762
 763         init_qp_attr->event_handler = kiblnd_qp_event;
 764         init_qp_attr->qp_context = conn;
 765         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 766         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 767         init_qp_attr->cap.max_send_sge = 1;
 768         init_qp_attr->cap.max_recv_sge = 1;
 769         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 770         init_qp_attr->qp_type = IB_QPT_RC;
 771         init_qp_attr->send_cq = cq;
 772         init_qp_attr->recv_cq = cq;
 773
 774         conn->ibc_sched = sched;
 775
 776         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 777         if (rc != 0) {
 778                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 779                        rc, init_qp_attr->cap.max_send_wr,
 780                        init_qp_attr->cap.max_recv_wr);
 781                 goto failed_2;
 782         }
 783
 784         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 785
 786         /* 1 ref for caller and each rxmsg */
 787         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 788         conn->ibc_nrx = IBLND_RX_MSGS(version);
 789
 790         /* post receives */
 791         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 792                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 793                                     IBLND_POSTRX_NO_CREDIT);
 794                 if (rc != 0) {
 795                         CERROR("Can't post rxmsg: %d\n", rc);
 796
 797                         /* Make posted receives complete */
 798                         kiblnd_abort_receives(conn);
 799
 800                         /* correct # of posted buffers
 801                          * NB locking needed now I'm racing with completion */
 802                         spin_lock_irqsave(&sched->ibs_lock, flags);
 803                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 804                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 805
 806                         /* cmid will be destroyed by CM(ofed) after cm_callback
 807                          * returned, so we can't refer it anymore
 808                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 809                         rdma_destroy_qp(conn->ibc_cmid);
 810                         conn->ibc_cmid = NULL;
 811
 812                         /* Drop my own and unused rxbuffer refcounts */
 813                         while (i++ <= IBLND_RX_MSGS(version))
 814                                 kiblnd_conn_decref(conn);
 815
 816                         return NULL;
 817                 }
 818         }
 819
 820         /* Init successful! */
 821         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 822                  state == IBLND_CONN_PASSIVE_WAIT);
 823         conn->ibc_state = state;
 824
 825         /* 1 more conn */
 826         atomic_inc(&net->ibn_nconns);
 827         return conn;
 828
 829  failed_2:
 830         kiblnd_destroy_conn(conn);
 831  failed_1:
 832         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 833  failed_0:
 834         return NULL;
 835 }
 836
 837 void kiblnd_destroy_conn(kib_conn_t *conn)
 838 {
 839         struct rdma_cm_id *cmid = conn->ibc_cmid;
 840         kib_peer_t *peer = conn->ibc_peer;
 841         int rc;
 842
 843         LASSERT(!in_interrupt());
 844         LASSERT(atomic_read(&conn->ibc_refcount) == 0);
 845         LASSERT(list_empty(&conn->ibc_early_rxs));
 846         LASSERT(list_empty(&conn->ibc_tx_noops));
 847         LASSERT(list_empty(&conn->ibc_tx_queue));
 848         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 849         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 850         LASSERT(list_empty(&conn->ibc_active_txs));
 851         LASSERT(conn->ibc_noops_posted == 0);
 852         LASSERT(conn->ibc_nsends_posted == 0);
 853
 854         switch (conn->ibc_state) {
 855         default:
 856                 /* conn must be completely disengaged from the network */
 857                 LBUG();
 858
 859         case IBLND_CONN_DISCONNECTED:
 860                 /* connvars should have been freed already */
 861                 LASSERT(conn->ibc_connvars == NULL);
 862                 break;
 863
 864         case IBLND_CONN_INIT:
 865                 break;
 866         }
 867
 868         /* conn->ibc_cmid might be destroyed by CM already */
 869         if (cmid != NULL && cmid->qp != NULL)
 870                 rdma_destroy_qp(cmid);
 871
 872         if (conn->ibc_cq != NULL) {
 873                 rc = ib_destroy_cq(conn->ibc_cq);
 874                 if (rc != 0)
 875                         CWARN("Error destroying CQ: %d\n", rc);
 876         }
 877
 878         if (conn->ibc_rx_pages != NULL)
 879                 kiblnd_unmap_rx_descs(conn);
 880
 881         if (conn->ibc_rxs != NULL) {
 882                 LIBCFS_FREE(conn->ibc_rxs,
 883                             IBLND_RX_MSGS(conn->ibc_version)
 884                               * sizeof(kib_rx_t));
 885         }
 886
 887         if (conn->ibc_connvars != NULL)
 888                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 889
 890         if (conn->ibc_hdev != NULL)
 891                 kiblnd_hdev_decref(conn->ibc_hdev);
 892
 893         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 894         if (conn->ibc_state != IBLND_CONN_INIT) {
 895                 kib_net_t *net = peer->ibp_ni->ni_data;
 896
 897                 kiblnd_peer_decref(peer);
 898                 rdma_destroy_id(cmid);
 899                 atomic_dec(&net->ibn_nconns);
 900         }
 901
 902         LIBCFS_FREE(conn, sizeof(*conn));
 903 }
 904
 905 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 906 {
 907         kib_conn_t *conn;
 908         struct list_head *ctmp;
 909         struct list_head *cnxt;
 910         int count = 0;
 911
 912         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 913                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 914
 915                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 916                        libcfs_nid2str(peer->ibp_nid),
 917                        conn->ibc_version, why);
 918
 919                 kiblnd_close_conn_locked(conn, why);
 920                 count++;
 921         }
 922
 923         return count;
 924 }
 925
 926 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
 927                                      int version, __u64 incarnation)
 928 {
 929         kib_conn_t *conn;
 930         struct list_head *ctmp;
 931         struct list_head *cnxt;
 932         int count = 0;
 933
 934         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 935                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 936
 937                 if (conn->ibc_version     == version &&
 938                     conn->ibc_incarnation == incarnation)
 939                         continue;
 940
 941                 CDEBUG(D_NET,
 942                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 943                        libcfs_nid2str(peer->ibp_nid),
 944                        conn->ibc_version, conn->ibc_incarnation,
 945                        version, incarnation);
 946
 947                 kiblnd_close_conn_locked(conn, -ESTALE);
 948                 count++;
 949         }
 950
 951         return count;
 952 }
 953
 954 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 955 {
 956         kib_peer_t *peer;
 957         struct list_head *ptmp;
 958         struct list_head *pnxt;
 959         int lo;
 960         int hi;
 961         int i;
 962         unsigned long flags;
 963         int count = 0;
 964
 965         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 966
 967         if (nid != LNET_NID_ANY)
 968                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 969         else {
 970                 lo = 0;
 971                 hi = kiblnd_data.kib_peer_hash_size - 1;
 972         }
 973
 974         for (i = lo; i <= hi; i++) {
 975                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 976
 977                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 978                         LASSERT(peer->ibp_connecting > 0 ||
 979                                  peer->ibp_accepting > 0 ||
 980                                  !list_empty(&peer->ibp_conns));
 981
 982                         if (peer->ibp_ni != ni)
 983                                 continue;
 984
 985                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 986                                 continue;
 987
 988                         count += kiblnd_close_peer_conns_locked(peer, 0);
 989                 }
 990         }
 991
 992         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 993
 994         /* wildcards always succeed */
 995         if (nid == LNET_NID_ANY)
 996                 return 0;
 997
 998         return (count == 0) ? -ENOENT : 0;
 999 }
1000
1001 int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1002 {
1003         struct libcfs_ioctl_data *data = arg;
1004         int rc = -EINVAL;
1005
1006         switch (cmd) {
1007         case IOC_LIBCFS_GET_PEER: {
1008                 lnet_nid_t nid = 0;
1009                 int count = 0;
1010
1011                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1012                                           &nid, &count);
1013                 data->ioc_nid   = nid;
1014                 data->ioc_count = count;
1015                 break;
1016         }
1017
1018         case IOC_LIBCFS_DEL_PEER: {
1019                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1020                 break;
1021         }
1022         case IOC_LIBCFS_GET_CONN: {
1023                 kib_conn_t *conn;
1024
1025                 rc = 0;
1026                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1027                 if (conn == NULL) {
1028                         rc = -ENOENT;
1029                         break;
1030                 }
1031
1032                 LASSERT(conn->ibc_cmid != NULL);
1033                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1034                 if (conn->ibc_cmid->route.path_rec == NULL)
1035                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1036                 else
1037                         data->ioc_u32[0] =
1038                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1039                 kiblnd_conn_decref(conn);
1040                 break;
1041         }
1042         case IOC_LIBCFS_CLOSE_CONNECTION: {
1043                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1044                 break;
1045         }
1046
1047         default:
1048                 break;
1049         }
1050
1051         return rc;
1052 }
1053
1054 void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1055 {
1056         unsigned long last_alive = 0;
1057         unsigned long now = cfs_time_current();
1058         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1059         kib_peer_t *peer;
1060         unsigned long flags;
1061
1062         read_lock_irqsave(glock, flags);
1063
1064         peer = kiblnd_find_peer_locked(nid);
1065         if (peer != NULL) {
1066                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1067                          peer->ibp_accepting > 0 ||
1068                          !list_empty(&peer->ibp_conns));  /* active conn */
1069                 last_alive = peer->ibp_last_alive;
1070         }
1071
1072         read_unlock_irqrestore(glock, flags);
1073
1074         if (last_alive != 0)
1075                 *when = last_alive;
1076
1077         /* peer is not persistent in hash, trigger peer creation
1078          * and connection establishment with a NULL tx */
1079         if (peer == NULL)
1080                 kiblnd_launch_tx(ni, NULL, nid);
1081
1082         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1083                libcfs_nid2str(nid), peer,
1084                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1085 }
1086
1087 void kiblnd_free_pages(kib_pages_t *p)
1088 {
1089         int npages = p->ibp_npages;
1090         int i;
1091
1092         for (i = 0; i < npages; i++) {
1093                 if (p->ibp_pages[i] != NULL)
1094                         __free_page(p->ibp_pages[i]);
1095         }
1096
1097         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1098 }
1099
1100 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1101 {
1102         kib_pages_t *p;
1103         int i;
1104
1105         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1106                          offsetof(kib_pages_t, ibp_pages[npages]));
1107         if (p == NULL) {
1108                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1109                 return -ENOMEM;
1110         }
1111
1112         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1113         p->ibp_npages = npages;
1114
1115         for (i = 0; i < npages; i++) {
1116                 p->ibp_pages[i] = alloc_pages_node(
1117                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1118                                     GFP_NOFS, 0);
1119                 if (p->ibp_pages[i] == NULL) {
1120                         CERROR("Can't allocate page %d of %d\n", i, npages);
1121                         kiblnd_free_pages(p);
1122                         return -ENOMEM;
1123                 }
1124         }
1125
1126         *pp = p;
1127         return 0;
1128 }
1129
1130 void kiblnd_unmap_rx_descs(kib_conn_t *conn)
1131 {
1132         kib_rx_t *rx;
1133         int i;
1134
1135         LASSERT(conn->ibc_rxs != NULL);
1136         LASSERT(conn->ibc_hdev != NULL);
1137
1138         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1139                 rx = &conn->ibc_rxs[i];
1140
1141                 LASSERT(rx->rx_nob >= 0); /* not posted */
1142
1143                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1144                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1145                                                           rx->rx_msgaddr),
1146                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1147         }
1148
1149         kiblnd_free_pages(conn->ibc_rx_pages);
1150
1151         conn->ibc_rx_pages = NULL;
1152 }
1153
1154 void kiblnd_map_rx_descs(kib_conn_t *conn)
1155 {
1156         kib_rx_t *rx;
1157         struct page *pg;
1158         int pg_off;
1159         int ipg;
1160         int i;
1161
1162         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1163                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1164                 rx = &conn->ibc_rxs[i];
1165
1166                 rx->rx_conn = conn;
1167                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1168
1169                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1170                                                        rx->rx_msg,
1171                                                        IBLND_MSG_SIZE,
1172                                                        DMA_FROM_DEVICE);
1173                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1174                                                    rx->rx_msgaddr));
1175                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1176
1177                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1178                        i, rx->rx_msg, rx->rx_msgaddr,
1179                        lnet_page2phys(pg) + pg_off);
1180
1181                 pg_off += IBLND_MSG_SIZE;
1182                 LASSERT(pg_off <= PAGE_SIZE);
1183
1184                 if (pg_off == PAGE_SIZE) {
1185                         pg_off = 0;
1186                         ipg++;
1187                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1188                 }
1189         }
1190 }
1191
1192 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1193 {
1194         kib_hca_dev_t *hdev = tpo->tpo_hdev;
1195         kib_tx_t *tx;
1196         int i;
1197
1198         LASSERT(tpo->tpo_pool.po_allocated == 0);
1199
1200         if (hdev == NULL)
1201                 return;
1202
1203         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1204                 tx = &tpo->tpo_tx_descs[i];
1205                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1206                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1207                                                           tx->tx_msgaddr),
1208                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1209         }
1210
1211         kiblnd_hdev_decref(hdev);
1212         tpo->tpo_hdev = NULL;
1213 }
1214
1215 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1216 {
1217         kib_hca_dev_t *hdev;
1218         unsigned long flags;
1219         int i = 0;
1220
1221         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1222         while (dev->ibd_failover) {
1223                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1224                 if (i++ % 50 == 0)
1225                         CDEBUG(D_NET, "%s: Wait for failover\n",
1226                                dev->ibd_ifname);
1227                 schedule_timeout(cfs_time_seconds(1) / 100);
1228
1229                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1230         }
1231
1232         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1233         hdev = dev->ibd_hdev;
1234
1235         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1236
1237         return hdev;
1238 }
1239
1240 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1241 {
1242         kib_pages_t *txpgs = tpo->tpo_tx_pages;
1243         kib_pool_t *pool = &tpo->tpo_pool;
1244         kib_net_t *net = pool->po_owner->ps_net;
1245         kib_dev_t *dev;
1246         struct page *page;
1247         kib_tx_t *tx;
1248         int page_offset;
1249         int ipage;
1250         int i;
1251
1252         LASSERT(net != NULL);
1253
1254         dev = net->ibn_dev;
1255
1256         /* pre-mapped messages are not bigger than 1 page */
1257         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1258
1259         /* No fancy arithmetic when we do the buffer calculations */
1260         CLASSERT(PAGE_SIZE % IBLND_MSG_SIZE == 0);
1261
1262         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1263
1264         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1265                 page = txpgs->ibp_pages[ipage];
1266                 tx = &tpo->tpo_tx_descs[i];
1267
1268                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1269                                            page_offset);
1270
1271                 tx->tx_msgaddr = kiblnd_dma_map_single(
1272                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1273                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1274                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1275                                                    tx->tx_msgaddr));
1276                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1277
1278                 list_add(&tx->tx_list, &pool->po_free_list);
1279
1280                 page_offset += IBLND_MSG_SIZE;
1281                 LASSERT(page_offset <= PAGE_SIZE);
1282
1283                 if (page_offset == PAGE_SIZE) {
1284                         page_offset = 0;
1285                         ipage++;
1286                         LASSERT(ipage <= txpgs->ibp_npages);
1287                 }
1288         }
1289 }
1290
1291 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1292 {
1293         __u64 index;
1294
1295         LASSERT(hdev->ibh_mrs[0] != NULL);
1296
1297         if (hdev->ibh_nmrs == 1)
1298                 return hdev->ibh_mrs[0];
1299
1300         index = addr >> hdev->ibh_mr_shift;
1301
1302         if (index <  hdev->ibh_nmrs &&
1303             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1304                 return hdev->ibh_mrs[index];
1305
1306         return NULL;
1307 }
1308
1309 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1310 {
1311         struct ib_mr *prev_mr;
1312         struct ib_mr *mr;
1313         int i;
1314
1315         LASSERT(hdev->ibh_mrs[0] != NULL);
1316
1317         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1318             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1319                 return NULL;
1320
1321         if (hdev->ibh_nmrs == 1)
1322                 return hdev->ibh_mrs[0];
1323
1324         for (i = 0, mr = prev_mr = NULL;
1325              i < rd->rd_nfrags; i++) {
1326                 mr = kiblnd_find_dma_mr(hdev,
1327                                         rd->rd_frags[i].rf_addr,
1328                                         rd->rd_frags[i].rf_nob);
1329                 if (prev_mr == NULL)
1330                         prev_mr = mr;
1331
1332                 if (mr == NULL || prev_mr != mr) {
1333                         /* Can't covered by one single MR */
1334                         mr = NULL;
1335                         break;
1336                 }
1337         }
1338
1339         return mr;
1340 }
1341
1342 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1343 {
1344         LASSERT(pool->fpo_map_count == 0);
1345
1346         if (pool->fpo_fmr_pool != NULL)
1347                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1348
1349         if (pool->fpo_hdev != NULL)
1350                 kiblnd_hdev_decref(pool->fpo_hdev);
1351
1352         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1353 }
1354
1355 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1356 {
1357         kib_fmr_pool_t *pool;
1358
1359         while (!list_empty(head)) {
1360                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1361                 list_del(&pool->fpo_list);
1362                 kiblnd_destroy_fmr_pool(pool);
1363         }
1364 }
1365
1366 static int kiblnd_fmr_pool_size(int ncpts)
1367 {
1368         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1369
1370         return max(IBLND_FMR_POOL, size);
1371 }
1372
1373 static int kiblnd_fmr_flush_trigger(int ncpts)
1374 {
1375         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1376
1377         return max(IBLND_FMR_POOL_FLUSH, size);
1378 }
1379
1380 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1381                                   kib_fmr_pool_t **pp_fpo)
1382 {
1383         /* FMR pool for RDMA */
1384         kib_dev_t *dev = fps->fps_net->ibn_dev;
1385         kib_fmr_pool_t *fpo;
1386         struct ib_fmr_pool_param param = {
1387                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1388                 .page_shift        = PAGE_SHIFT,
1389                 .access            = (IB_ACCESS_LOCAL_WRITE |
1390                                       IB_ACCESS_REMOTE_WRITE),
1391                 .pool_size         = fps->fps_pool_size,
1392                 .dirty_watermark   = fps->fps_flush_trigger,
1393                 .flush_function    = NULL,
1394                 .flush_arg         = NULL,
1395                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1396         int rc;
1397
1398         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1399         if (fpo == NULL)
1400                 return -ENOMEM;
1401
1402         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1403
1404         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1405         if (IS_ERR(fpo->fpo_fmr_pool)) {
1406                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1407                 CERROR("Failed to create FMR pool: %d\n", rc);
1408
1409                 kiblnd_hdev_decref(fpo->fpo_hdev);
1410                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1411                 return rc;
1412         }
1413
1414         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1415         fpo->fpo_owner    = fps;
1416         *pp_fpo = fpo;
1417
1418         return 0;
1419 }
1420
1421 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1422                                     struct list_head *zombies)
1423 {
1424         if (fps->fps_net == NULL) /* intialized? */
1425                 return;
1426
1427         spin_lock(&fps->fps_lock);
1428
1429         while (!list_empty(&fps->fps_pool_list)) {
1430                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1431                                                  kib_fmr_pool_t, fpo_list);
1432                 fpo->fpo_failed = 1;
1433                 list_del(&fpo->fpo_list);
1434                 if (fpo->fpo_map_count == 0)
1435                         list_add(&fpo->fpo_list, zombies);
1436                 else
1437                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1438         }
1439
1440         spin_unlock(&fps->fps_lock);
1441 }
1442
1443 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1444 {
1445         if (fps->fps_net != NULL) { /* initialized? */
1446                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1447                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1448         }
1449 }
1450
1451 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1452                                    kib_net_t *net, int pool_size,
1453                                    int flush_trigger)
1454 {
1455         kib_fmr_pool_t *fpo;
1456         int rc;
1457
1458         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1459
1460         fps->fps_net = net;
1461         fps->fps_cpt = cpt;
1462         fps->fps_pool_size = pool_size;
1463         fps->fps_flush_trigger = flush_trigger;
1464         spin_lock_init(&fps->fps_lock);
1465         INIT_LIST_HEAD(&fps->fps_pool_list);
1466         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1467
1468         rc = kiblnd_create_fmr_pool(fps, &fpo);
1469         if (rc == 0)
1470                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1471
1472         return rc;
1473 }
1474
1475 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1476 {
1477         if (fpo->fpo_map_count != 0) /* still in use */
1478                 return 0;
1479         if (fpo->fpo_failed)
1480                 return 1;
1481         return cfs_time_aftereq(now, fpo->fpo_deadline);
1482 }
1483
1484 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1485 {
1486         LIST_HEAD(zombies);
1487         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1488         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1489         unsigned long now = cfs_time_current();
1490         kib_fmr_pool_t *tmp;
1491         int rc;
1492
1493         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1494         LASSERT(rc == 0);
1495
1496         if (status != 0) {
1497                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1498                 LASSERT(rc == 0);
1499         }
1500
1501         fmr->fmr_pool = NULL;
1502         fmr->fmr_pfmr = NULL;
1503
1504         spin_lock(&fps->fps_lock);
1505         fpo->fpo_map_count--;  /* decref the pool */
1506
1507         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1508                 /* the first pool is persistent */
1509                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1510                         continue;
1511
1512                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1513                         list_move(&fpo->fpo_list, &zombies);
1514                         fps->fps_version++;
1515                 }
1516         }
1517         spin_unlock(&fps->fps_lock);
1518
1519         if (!list_empty(&zombies))
1520                 kiblnd_destroy_fmr_pool_list(&zombies);
1521 }
1522
1523 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1524                         __u64 iov, kib_fmr_t *fmr)
1525 {
1526         struct ib_pool_fmr *pfmr;
1527         kib_fmr_pool_t *fpo;
1528         __u64 version;
1529         int rc;
1530
1531  again:
1532         spin_lock(&fps->fps_lock);
1533         version = fps->fps_version;
1534         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1535                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1536                 fpo->fpo_map_count++;
1537                 spin_unlock(&fps->fps_lock);
1538
1539                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1540                                             pages, npages, iov);
1541                 if (likely(!IS_ERR(pfmr))) {
1542                         fmr->fmr_pool = fpo;
1543                         fmr->fmr_pfmr = pfmr;
1544                         return 0;
1545                 }
1546
1547                 spin_lock(&fps->fps_lock);
1548                 fpo->fpo_map_count--;
1549                 if (PTR_ERR(pfmr) != -EAGAIN) {
1550                         spin_unlock(&fps->fps_lock);
1551                         return PTR_ERR(pfmr);
1552                 }
1553
1554                 /* EAGAIN and ... */
1555                 if (version != fps->fps_version) {
1556                         spin_unlock(&fps->fps_lock);
1557                         goto again;
1558                 }
1559         }
1560
1561         if (fps->fps_increasing) {
1562                 spin_unlock(&fps->fps_lock);
1563                 CDEBUG(D_NET,
1564                         "Another thread is allocating new FMR pool, waiting for her to complete\n");
1565                 schedule();
1566                 goto again;
1567
1568         }
1569
1570         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1571                 /* someone failed recently */
1572                 spin_unlock(&fps->fps_lock);
1573                 return -EAGAIN;
1574         }
1575
1576         fps->fps_increasing = 1;
1577         spin_unlock(&fps->fps_lock);
1578
1579         CDEBUG(D_NET, "Allocate new FMR pool\n");
1580         rc = kiblnd_create_fmr_pool(fps, &fpo);
1581         spin_lock(&fps->fps_lock);
1582         fps->fps_increasing = 0;
1583         if (rc == 0) {
1584                 fps->fps_version++;
1585                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1586         } else {
1587                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1588         }
1589         spin_unlock(&fps->fps_lock);
1590
1591         goto again;
1592 }
1593
1594 static void kiblnd_fini_pool(kib_pool_t *pool)
1595 {
1596         LASSERT(list_empty(&pool->po_free_list));
1597         LASSERT(pool->po_allocated == 0);
1598
1599         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1600 }
1601
1602 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1603 {
1604         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1605
1606         memset(pool, 0, sizeof(kib_pool_t));
1607         INIT_LIST_HEAD(&pool->po_free_list);
1608         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1609         pool->po_owner    = ps;
1610         pool->po_size     = size;
1611 }
1612
1613 static void kiblnd_destroy_pool_list(struct list_head *head)
1614 {
1615         kib_pool_t *pool;
1616
1617         while (!list_empty(head)) {
1618                 pool = list_entry(head->next, kib_pool_t, po_list);
1619                 list_del(&pool->po_list);
1620
1621                 LASSERT(pool->po_owner != NULL);
1622                 pool->po_owner->ps_pool_destroy(pool);
1623         }
1624 }
1625
1626 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1627 {
1628         if (ps->ps_net == NULL) /* intialized? */
1629                 return;
1630
1631         spin_lock(&ps->ps_lock);
1632         while (!list_empty(&ps->ps_pool_list)) {
1633                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1634                                             kib_pool_t, po_list);
1635                 po->po_failed = 1;
1636                 list_del(&po->po_list);
1637                 if (po->po_allocated == 0)
1638                         list_add(&po->po_list, zombies);
1639                 else
1640                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1641         }
1642         spin_unlock(&ps->ps_lock);
1643 }
1644
1645 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1646 {
1647         if (ps->ps_net != NULL) { /* initialized? */
1648                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1649                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1650         }
1651 }
1652
1653 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1654                                kib_net_t *net, char *name, int size,
1655                                kib_ps_pool_create_t po_create,
1656                                kib_ps_pool_destroy_t po_destroy,
1657                                kib_ps_node_init_t nd_init,
1658                                kib_ps_node_fini_t nd_fini)
1659 {
1660         kib_pool_t *pool;
1661         int rc;
1662
1663         memset(ps, 0, sizeof(kib_poolset_t));
1664
1665         ps->ps_cpt          = cpt;
1666         ps->ps_net          = net;
1667         ps->ps_pool_create  = po_create;
1668         ps->ps_pool_destroy = po_destroy;
1669         ps->ps_node_init    = nd_init;
1670         ps->ps_node_fini    = nd_fini;
1671         ps->ps_pool_size    = size;
1672         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1673             >= sizeof(ps->ps_name))
1674                 return -E2BIG;
1675         spin_lock_init(&ps->ps_lock);
1676         INIT_LIST_HEAD(&ps->ps_pool_list);
1677         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1678
1679         rc = ps->ps_pool_create(ps, size, &pool);
1680         if (rc == 0)
1681                 list_add(&pool->po_list, &ps->ps_pool_list);
1682         else
1683                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1684
1685         return rc;
1686 }
1687
1688 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1689 {
1690         if (pool->po_allocated != 0) /* still in use */
1691                 return 0;
1692         if (pool->po_failed)
1693                 return 1;
1694         return cfs_time_aftereq(now, pool->po_deadline);
1695 }
1696
1697 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1698 {
1699         LIST_HEAD(zombies);
1700         kib_poolset_t *ps = pool->po_owner;
1701         kib_pool_t *tmp;
1702         unsigned long now = cfs_time_current();
1703
1704         spin_lock(&ps->ps_lock);
1705
1706         if (ps->ps_node_fini != NULL)
1707                 ps->ps_node_fini(pool, node);
1708
1709         LASSERT(pool->po_allocated > 0);
1710         list_add(node, &pool->po_free_list);
1711         pool->po_allocated--;
1712
1713         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1714                 /* the first pool is persistent */
1715                 if (ps->ps_pool_list.next == &pool->po_list)
1716                         continue;
1717
1718                 if (kiblnd_pool_is_idle(pool, now))
1719                         list_move(&pool->po_list, &zombies);
1720         }
1721         spin_unlock(&ps->ps_lock);
1722
1723         if (!list_empty(&zombies))
1724                 kiblnd_destroy_pool_list(&zombies);
1725 }
1726
1727 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1728 {
1729         struct list_head *node;
1730         kib_pool_t *pool;
1731         int rc;
1732
1733  again:
1734         spin_lock(&ps->ps_lock);
1735         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1736                 if (list_empty(&pool->po_free_list))
1737                         continue;
1738
1739                 pool->po_allocated++;
1740                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1741                 node = pool->po_free_list.next;
1742                 list_del(node);
1743
1744                 if (ps->ps_node_init != NULL) {
1745                         /* still hold the lock */
1746                         ps->ps_node_init(pool, node);
1747                 }
1748                 spin_unlock(&ps->ps_lock);
1749                 return node;
1750         }
1751
1752         /* no available tx pool and ... */
1753         if (ps->ps_increasing) {
1754                 /* another thread is allocating a new pool */
1755                 spin_unlock(&ps->ps_lock);
1756                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1757                        ps->ps_name);
1758                 schedule();
1759                 goto again;
1760         }
1761
1762         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1763                 /* someone failed recently */
1764                 spin_unlock(&ps->ps_lock);
1765                 return NULL;
1766         }
1767
1768         ps->ps_increasing = 1;
1769         spin_unlock(&ps->ps_lock);
1770
1771         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1772
1773         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1774
1775         spin_lock(&ps->ps_lock);
1776         ps->ps_increasing = 0;
1777         if (rc == 0) {
1778                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1779         } else {
1780                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1781                 CERROR("Can't allocate new %s pool because out of memory\n",
1782                        ps->ps_name);
1783         }
1784         spin_unlock(&ps->ps_lock);
1785
1786         goto again;
1787 }
1788
1789 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1790 {
1791         kib_pmr_pool_t *ppo = pmr->pmr_pool;
1792         struct ib_mr *mr = pmr->pmr_mr;
1793
1794         pmr->pmr_mr = NULL;
1795         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1796         if (mr != NULL)
1797                 ib_dereg_mr(mr);
1798 }
1799
1800 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1801                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1802 {
1803         kib_phys_mr_t *pmr;
1804         struct list_head *node;
1805         int rc;
1806         int i;
1807
1808         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1809         if (node == NULL) {
1810                 CERROR("Failed to allocate PMR descriptor\n");
1811                 return -ENOMEM;
1812         }
1813
1814         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1815         if (pmr->pmr_pool->ppo_hdev != hdev) {
1816                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1817                 return -EAGAIN;
1818         }
1819
1820         for (i = 0; i < rd->rd_nfrags; i++) {
1821                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1822                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1823         }
1824
1825         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1826                                      pmr->pmr_ipb, rd->rd_nfrags,
1827                                      IB_ACCESS_LOCAL_WRITE |
1828                                      IB_ACCESS_REMOTE_WRITE,
1829                                      iova);
1830         if (!IS_ERR(pmr->pmr_mr)) {
1831                 pmr->pmr_iova = *iova;
1832                 *pp_pmr = pmr;
1833                 return 0;
1834         }
1835
1836         rc = PTR_ERR(pmr->pmr_mr);
1837         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1838
1839         pmr->pmr_mr = NULL;
1840         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1841
1842         return rc;
1843 }
1844
1845 static void kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1846 {
1847         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1848         kib_phys_mr_t *pmr;
1849         kib_phys_mr_t *tmp;
1850
1851         LASSERT(pool->po_allocated == 0);
1852
1853         list_for_each_entry_safe(pmr, tmp, &pool->po_free_list, pmr_list) {
1854                 LASSERT(pmr->pmr_mr == NULL);
1855                 list_del(&pmr->pmr_list);
1856
1857                 if (pmr->pmr_ipb != NULL) {
1858                         LIBCFS_FREE(pmr->pmr_ipb,
1859                                     IBLND_MAX_RDMA_FRAGS *
1860                                     sizeof(struct ib_phys_buf));
1861                 }
1862
1863                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1864         }
1865
1866         kiblnd_fini_pool(pool);
1867         if (ppo->ppo_hdev != NULL)
1868                 kiblnd_hdev_decref(ppo->ppo_hdev);
1869
1870         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1871 }
1872
1873 static inline int kiblnd_pmr_pool_size(int ncpts)
1874 {
1875         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1876
1877         return max(IBLND_PMR_POOL, size);
1878 }
1879
1880 static int kiblnd_create_pmr_pool(kib_poolset_t *ps, int size,
1881                                   kib_pool_t **pp_po)
1882 {
1883         struct kib_pmr_pool *ppo;
1884         struct kib_pool *pool;
1885         kib_phys_mr_t *pmr;
1886         int i;
1887
1888         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
1889                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
1890         if (ppo == NULL) {
1891                 CERROR("Failed to allocate PMR pool\n");
1892                 return -ENOMEM;
1893         }
1894
1895         pool = &ppo->ppo_pool;
1896         kiblnd_init_pool(ps, pool, size);
1897
1898         for (i = 0; i < size; i++) {
1899                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
1900                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
1901                 if (pmr == NULL)
1902                         break;
1903
1904                 pmr->pmr_pool = ppo;
1905                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
1906                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
1907                 if (pmr->pmr_ipb == NULL)
1908                         break;
1909
1910                 list_add(&pmr->pmr_list, &pool->po_free_list);
1911         }
1912
1913         if (i < size) {
1914                 ps->ps_pool_destroy(pool);
1915                 return -ENOMEM;
1916         }
1917
1918         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1919         *pp_po = pool;
1920         return 0;
1921 }
1922
1923 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1924 {
1925         kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1926         int i;
1927
1928         LASSERT(pool->po_allocated == 0);
1929
1930         if (tpo->tpo_tx_pages != NULL) {
1931                 kiblnd_unmap_tx_pool(tpo);
1932                 kiblnd_free_pages(tpo->tpo_tx_pages);
1933         }
1934
1935         if (tpo->tpo_tx_descs == NULL)
1936                 goto out;
1937
1938         for (i = 0; i < pool->po_size; i++) {
1939                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1940
1941                 list_del(&tx->tx_list);
1942                 if (tx->tx_pages != NULL)
1943                         LIBCFS_FREE(tx->tx_pages,
1944                                     LNET_MAX_IOV *
1945                                     sizeof(*tx->tx_pages));
1946                 if (tx->tx_frags != NULL)
1947                         LIBCFS_FREE(tx->tx_frags,
1948                                     IBLND_MAX_RDMA_FRAGS *
1949                                             sizeof(*tx->tx_frags));
1950                 if (tx->tx_wrq != NULL)
1951                         LIBCFS_FREE(tx->tx_wrq,
1952                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1953                                     sizeof(*tx->tx_wrq));
1954                 if (tx->tx_sge != NULL)
1955                         LIBCFS_FREE(tx->tx_sge,
1956                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1957                                     sizeof(*tx->tx_sge));
1958                 if (tx->tx_rd != NULL)
1959                         LIBCFS_FREE(tx->tx_rd,
1960                                     offsetof(kib_rdma_desc_t,
1961                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1962         }
1963
1964         LIBCFS_FREE(tpo->tpo_tx_descs,
1965                     pool->po_size * sizeof(kib_tx_t));
1966 out:
1967         kiblnd_fini_pool(pool);
1968         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1969 }
1970
1971 static int kiblnd_tx_pool_size(int ncpts)
1972 {
1973         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1974
1975         return max(IBLND_TX_POOL, ntx);
1976 }
1977
1978 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1979                                  kib_pool_t **pp_po)
1980 {
1981         int i;
1982         int npg;
1983         kib_pool_t *pool;
1984         kib_tx_pool_t *tpo;
1985
1986         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1987         if (tpo == NULL) {
1988                 CERROR("Failed to allocate TX pool\n");
1989                 return -ENOMEM;
1990         }
1991
1992         pool = &tpo->tpo_pool;
1993         kiblnd_init_pool(ps, pool, size);
1994         tpo->tpo_tx_descs = NULL;
1995         tpo->tpo_tx_pages = NULL;
1996
1997         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1998         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1999                 CERROR("Can't allocate tx pages: %d\n", npg);
2000                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2001                 return -ENOMEM;
2002         }
2003
2004         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2005                          size * sizeof(kib_tx_t));
2006         if (tpo->tpo_tx_descs == NULL) {
2007                 CERROR("Can't allocate %d tx descriptors\n", size);
2008                 ps->ps_pool_destroy(pool);
2009                 return -ENOMEM;
2010         }
2011
2012         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2013
2014         for (i = 0; i < size; i++) {
2015                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2016
2017                 tx->tx_pool = tpo;
2018                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2019                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2020                                          lnet_cpt_table(), ps->ps_cpt,
2021                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2022                         if (tx->tx_pages == NULL)
2023                                 break;
2024                 }
2025
2026                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2027                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2028                 if (tx->tx_frags == NULL)
2029                         break;
2030
2031                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2032
2033                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2034                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2035                                  sizeof(*tx->tx_wrq));
2036                 if (tx->tx_wrq == NULL)
2037                         break;
2038
2039                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2040                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2041                                  sizeof(*tx->tx_sge));
2042                 if (tx->tx_sge == NULL)
2043                         break;
2044
2045                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2046                                  offsetof(kib_rdma_desc_t,
2047                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2048                 if (tx->tx_rd == NULL)
2049                         break;
2050         }
2051
2052         if (i == size) {
2053                 kiblnd_map_tx_pool(tpo);
2054                 *pp_po = pool;
2055                 return 0;
2056         }
2057
2058         ps->ps_pool_destroy(pool);
2059         return -ENOMEM;
2060 }
2061
2062 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2063 {
2064         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2065                                              tps_poolset);
2066         kib_tx_t *tx  = list_entry(node, kib_tx_t, tx_list);
2067
2068         tx->tx_cookie = tps->tps_next_tx_cookie++;
2069 }
2070
2071 static void kiblnd_net_fini_pools(kib_net_t *net)
2072 {
2073         int i;
2074
2075         cfs_cpt_for_each(i, lnet_cpt_table()) {
2076                 kib_tx_poolset_t *tps;
2077                 kib_fmr_poolset_t *fps;
2078                 kib_pmr_poolset_t *pps;
2079
2080                 if (net->ibn_tx_ps != NULL) {
2081                         tps = net->ibn_tx_ps[i];
2082                         kiblnd_fini_poolset(&tps->tps_poolset);
2083                 }
2084
2085                 if (net->ibn_fmr_ps != NULL) {
2086                         fps = net->ibn_fmr_ps[i];
2087                         kiblnd_fini_fmr_poolset(fps);
2088                 }
2089
2090                 if (net->ibn_pmr_ps != NULL) {
2091                         pps = net->ibn_pmr_ps[i];
2092                         kiblnd_fini_poolset(&pps->pps_poolset);
2093                 }
2094         }
2095
2096         if (net->ibn_tx_ps != NULL) {
2097                 cfs_percpt_free(net->ibn_tx_ps);
2098                 net->ibn_tx_ps = NULL;
2099         }
2100
2101         if (net->ibn_fmr_ps != NULL) {
2102                 cfs_percpt_free(net->ibn_fmr_ps);
2103                 net->ibn_fmr_ps = NULL;
2104         }
2105
2106         if (net->ibn_pmr_ps != NULL) {
2107                 cfs_percpt_free(net->ibn_pmr_ps);
2108                 net->ibn_pmr_ps = NULL;
2109         }
2110 }
2111
2112 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2113 {
2114         unsigned long flags;
2115         int cpt;
2116         int rc;
2117         int i;
2118
2119         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2120         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2121             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2122                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2123                 goto create_tx_pool;
2124         }
2125
2126         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2127
2128         if (*kiblnd_tunables.kib_fmr_pool_size <
2129             *kiblnd_tunables.kib_ntx / 4) {
2130                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2131                        *kiblnd_tunables.kib_fmr_pool_size,
2132                        *kiblnd_tunables.kib_ntx / 4);
2133                 rc = -EINVAL;
2134                 goto failed;
2135         }
2136
2137         /* TX pool must be created later than FMR/PMR, see LU-2268
2138          * for details */
2139         LASSERT(net->ibn_tx_ps == NULL);
2140
2141         /* premapping can fail if ibd_nmr > 1, so we always create
2142          * FMR/PMR pool and map-on-demand if premapping failed */
2143
2144         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2145                                            sizeof(kib_fmr_poolset_t));
2146         if (net->ibn_fmr_ps == NULL) {
2147                 CERROR("Failed to allocate FMR pool array\n");
2148                 rc = -ENOMEM;
2149                 goto failed;
2150         }
2151
2152         for (i = 0; i < ncpts; i++) {
2153                 cpt = (cpts == NULL) ? i : cpts[i];
2154                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2155                                              kiblnd_fmr_pool_size(ncpts),
2156                                              kiblnd_fmr_flush_trigger(ncpts));
2157                 if (rc == -ENOSYS && i == 0) /* no FMR */
2158                         break; /* create PMR pool */
2159
2160                 if (rc != 0) { /* a real error */
2161                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2162                                cpt, rc);
2163                         goto failed;
2164                 }
2165         }
2166
2167         if (i > 0) {
2168                 LASSERT(i == ncpts);
2169                 goto create_tx_pool;
2170         }
2171
2172         cfs_percpt_free(net->ibn_fmr_ps);
2173         net->ibn_fmr_ps = NULL;
2174
2175         CWARN("Device does not support FMR, failing back to PMR\n");
2176
2177         if (*kiblnd_tunables.kib_pmr_pool_size <
2178             *kiblnd_tunables.kib_ntx / 4) {
2179                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2180                        *kiblnd_tunables.kib_pmr_pool_size,
2181                        *kiblnd_tunables.kib_ntx / 4);
2182                 rc = -EINVAL;
2183                 goto failed;
2184         }
2185
2186         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2187                                            sizeof(kib_pmr_poolset_t));
2188         if (net->ibn_pmr_ps == NULL) {
2189                 CERROR("Failed to allocate PMR pool array\n");
2190                 rc = -ENOMEM;
2191                 goto failed;
2192         }
2193
2194         for (i = 0; i < ncpts; i++) {
2195                 cpt = (cpts == NULL) ? i : cpts[i];
2196                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2197                                          cpt, net, "PMR",
2198                                          kiblnd_pmr_pool_size(ncpts),
2199                                          kiblnd_create_pmr_pool,
2200                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2201                 if (rc != 0) {
2202                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2203                                cpt, rc);
2204                         goto failed;
2205                 }
2206         }
2207
2208  create_tx_pool:
2209         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2210                                           sizeof(kib_tx_poolset_t));
2211         if (net->ibn_tx_ps == NULL) {
2212                 CERROR("Failed to allocate tx pool array\n");
2213                 rc = -ENOMEM;
2214                 goto failed;
2215         }
2216
2217         for (i = 0; i < ncpts; i++) {
2218                 cpt = (cpts == NULL) ? i : cpts[i];
2219                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2220                                          cpt, net, "TX",
2221                                          kiblnd_tx_pool_size(ncpts),
2222                                          kiblnd_create_tx_pool,
2223                                          kiblnd_destroy_tx_pool,
2224                                          kiblnd_tx_init, NULL);
2225                 if (rc != 0) {
2226                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2227                                cpt, rc);
2228                         goto failed;
2229                 }
2230         }
2231
2232         return 0;
2233  failed:
2234         kiblnd_net_fini_pools(net);
2235         LASSERT(rc != 0);
2236         return rc;
2237 }
2238
2239 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2240 {
2241         struct ib_device_attr *attr;
2242         int rc;
2243
2244         /* It's safe to assume a HCA can handle a page size
2245          * matching that of the native system */
2246         hdev->ibh_page_shift = PAGE_SHIFT;
2247         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2248         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2249
2250         LIBCFS_ALLOC(attr, sizeof(*attr));
2251         if (attr == NULL) {
2252                 CERROR("Out of memory\n");
2253                 return -ENOMEM;
2254         }
2255
2256         rc = ib_query_device(hdev->ibh_ibdev, attr);
2257         if (rc == 0)
2258                 hdev->ibh_mr_size = attr->max_mr_size;
2259
2260         LIBCFS_FREE(attr, sizeof(*attr));
2261
2262         if (rc != 0) {
2263                 CERROR("Failed to query IB device: %d\n", rc);
2264                 return rc;
2265         }
2266
2267         if (hdev->ibh_mr_size == ~0ULL) {
2268                 hdev->ibh_mr_shift = 64;
2269                 return 0;
2270         }
2271
2272         for (hdev->ibh_mr_shift = 0;
2273              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2274                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2275                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2276                         return 0;
2277         }
2278
2279         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2280         return -EINVAL;
2281 }
2282
2283 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2284 {
2285         int i;
2286
2287         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2288                 return;
2289
2290         for (i = 0; i < hdev->ibh_nmrs; i++) {
2291                 if (hdev->ibh_mrs[i] == NULL)
2292                         break;
2293
2294                 ib_dereg_mr(hdev->ibh_mrs[i]);
2295         }
2296
2297         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2298         hdev->ibh_mrs  = NULL;
2299         hdev->ibh_nmrs = 0;
2300 }
2301
2302 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2303 {
2304         kiblnd_hdev_cleanup_mrs(hdev);
2305
2306         if (hdev->ibh_pd != NULL)
2307                 ib_dealloc_pd(hdev->ibh_pd);
2308
2309         if (hdev->ibh_cmid != NULL)
2310                 rdma_destroy_id(hdev->ibh_cmid);
2311
2312         LIBCFS_FREE(hdev, sizeof(*hdev));
2313 }
2314
2315 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2316 {
2317         struct ib_mr *mr;
2318         int i;
2319         int rc;
2320         __u64 mm_size;
2321         __u64 mr_size;
2322         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2323
2324         rc = kiblnd_hdev_get_attr(hdev);
2325         if (rc != 0)
2326                 return rc;
2327
2328         if (hdev->ibh_mr_shift == 64) {
2329                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2330                 if (hdev->ibh_mrs == NULL) {
2331                         CERROR("Failed to allocate MRs table\n");
2332                         return -ENOMEM;
2333                 }
2334
2335                 hdev->ibh_mrs[0] = NULL;
2336                 hdev->ibh_nmrs   = 1;
2337
2338                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2339                 if (IS_ERR(mr)) {
2340                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2341                         kiblnd_hdev_cleanup_mrs(hdev);
2342                         return PTR_ERR(mr);
2343                 }
2344
2345                 hdev->ibh_mrs[0] = mr;
2346
2347                 goto out;
2348         }
2349
2350         mr_size = 1ULL << hdev->ibh_mr_shift;
2351         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2352
2353         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2354
2355         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2356                 /* it's 4T..., assume we will re-code at that time */
2357                 CERROR("Can't support memory size: x%#llx with MR size: x%#llx\n",
2358                        mm_size, mr_size);
2359                 return -EINVAL;
2360         }
2361
2362         /* create an array of MRs to cover all memory */
2363         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2364         if (hdev->ibh_mrs == NULL) {
2365                 CERROR("Failed to allocate MRs' table\n");
2366                 return -ENOMEM;
2367         }
2368
2369         for (i = 0; i < hdev->ibh_nmrs; i++) {
2370                 struct ib_phys_buf ipb;
2371                 __u64 iova;
2372
2373                 ipb.size = hdev->ibh_mr_size;
2374                 ipb.addr = i * mr_size;
2375                 iova = ipb.addr;
2376
2377                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2378                 if (IS_ERR(mr)) {
2379                         CERROR("Failed ib_reg_phys_mr addr %#llx size %#llx : %ld\n",
2380                                ipb.addr, ipb.size, PTR_ERR(mr));
2381                         kiblnd_hdev_cleanup_mrs(hdev);
2382                         return PTR_ERR(mr);
2383                 }
2384
2385                 LASSERT(iova == ipb.addr);
2386
2387                 hdev->ibh_mrs[i] = mr;
2388         }
2389
2390 out:
2391         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2392                 LCONSOLE_INFO("Register global MR array, MR size: %#llx, array size: %d\n",
2393                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2394         return 0;
2395 }
2396
2397 /* DUMMY */
2398 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2399                                  struct rdma_cm_event *event)
2400 {
2401         return 0;
2402 }
2403
2404 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2405 {
2406         struct rdma_cm_id *cmid;
2407         struct sockaddr_in srcaddr;
2408         struct sockaddr_in dstaddr;
2409         int rc;
2410
2411         if (dev->ibd_hdev == NULL || /* initializing */
2412             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2413             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2414                 return 1;
2415
2416         /* XXX: it's UGLY, but I don't have better way to find
2417          * ib-bonding HCA failover because:
2418          *
2419          * a. no reliable CM event for HCA failover...
2420          * b. no OFED API to get ib_device for current net_device...
2421          *
2422          * We have only two choices at this point:
2423          *
2424          * a. rdma_bind_addr(), it will conflict with listener cmid
2425          * b. rdma_resolve_addr() to zero addr */
2426         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2427                                      IB_QPT_RC);
2428         if (IS_ERR(cmid)) {
2429                 rc = PTR_ERR(cmid);
2430                 CERROR("Failed to create cmid for failover: %d\n", rc);
2431                 return rc;
2432         }
2433
2434         memset(&srcaddr, 0, sizeof(srcaddr));
2435         srcaddr.sin_family = AF_INET;
2436         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2437
2438         memset(&dstaddr, 0, sizeof(dstaddr));
2439         dstaddr.sin_family = AF_INET;
2440         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2441                                (struct sockaddr *)&dstaddr, 1);
2442         if (rc != 0 || cmid->device == NULL) {
2443                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2444                        dev->ibd_ifname, &dev->ibd_ifip,
2445                        cmid->device, rc);
2446                 rdma_destroy_id(cmid);
2447                 return rc;
2448         }
2449
2450         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2451                 /* don't need device failover */
2452                 rdma_destroy_id(cmid);
2453                 return 0;
2454         }
2455
2456         return 1;
2457 }
2458
2459 int kiblnd_dev_failover(kib_dev_t *dev)
2460 {
2461         LIST_HEAD(zombie_tpo);
2462         LIST_HEAD(zombie_ppo);
2463         LIST_HEAD(zombie_fpo);
2464         struct rdma_cm_id *cmid  = NULL;
2465         kib_hca_dev_t *hdev  = NULL;
2466         struct ib_pd *pd;
2467         kib_net_t *net;
2468         struct sockaddr_in addr;
2469         unsigned long flags;
2470         int rc = 0;
2471         int i;
2472
2473         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2474                  dev->ibd_can_failover ||
2475                  dev->ibd_hdev == NULL);
2476
2477         rc = kiblnd_dev_need_failover(dev);
2478         if (rc <= 0)
2479                 goto out;
2480
2481         if (dev->ibd_hdev != NULL &&
2482             dev->ibd_hdev->ibh_cmid != NULL) {
2483                 /* XXX it's not good to close old listener at here,
2484                  * because we can fail to create new listener.
2485                  * But we have to close it now, otherwise rdma_bind_addr
2486                  * will return EADDRINUSE... How crap! */
2487                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2488
2489                 cmid = dev->ibd_hdev->ibh_cmid;
2490                 /* make next schedule of kiblnd_dev_need_failover()
2491                  * return 1 for me */
2492                 dev->ibd_hdev->ibh_cmid  = NULL;
2493                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2494
2495                 rdma_destroy_id(cmid);
2496         }
2497
2498         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2499                                      IB_QPT_RC);
2500         if (IS_ERR(cmid)) {
2501                 rc = PTR_ERR(cmid);
2502                 CERROR("Failed to create cmid for failover: %d\n", rc);
2503                 goto out;
2504         }
2505
2506         memset(&addr, 0, sizeof(addr));
2507         addr.sin_family      = AF_INET;
2508         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2509         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2510
2511         /* Bind to failover device or port */
2512         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2513         if (rc != 0 || cmid->device == NULL) {
2514                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2515                        dev->ibd_ifname, &dev->ibd_ifip,
2516                        cmid->device, rc);
2517                 rdma_destroy_id(cmid);
2518                 goto out;
2519         }
2520
2521         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2522         if (hdev == NULL) {
2523                 CERROR("Failed to allocate kib_hca_dev\n");
2524                 rdma_destroy_id(cmid);
2525                 rc = -ENOMEM;
2526                 goto out;
2527         }
2528
2529         atomic_set(&hdev->ibh_ref, 1);
2530         hdev->ibh_dev   = dev;
2531         hdev->ibh_cmid  = cmid;
2532         hdev->ibh_ibdev = cmid->device;
2533
2534         pd = ib_alloc_pd(cmid->device);
2535         if (IS_ERR(pd)) {
2536                 rc = PTR_ERR(pd);
2537                 CERROR("Can't allocate PD: %d\n", rc);
2538                 goto out;
2539         }
2540
2541         hdev->ibh_pd = pd;
2542
2543         rc = rdma_listen(cmid, 0);
2544         if (rc != 0) {
2545                 CERROR("Can't start new listener: %d\n", rc);
2546                 goto out;
2547         }
2548
2549         rc = kiblnd_hdev_setup_mrs(hdev);
2550         if (rc != 0) {
2551                 CERROR("Can't setup device: %d\n", rc);
2552                 goto out;
2553         }
2554
2555         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2556
2557         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2558
2559         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2560                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2561                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2562                                             &zombie_tpo);
2563
2564                         if (net->ibn_fmr_ps != NULL) {
2565                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2566                                                         &zombie_fpo);
2567
2568                         } else if (net->ibn_pmr_ps != NULL) {
2569                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2570                                                     pps_poolset, &zombie_ppo);
2571                         }
2572                 }
2573         }
2574
2575         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2576  out:
2577         if (!list_empty(&zombie_tpo))
2578                 kiblnd_destroy_pool_list(&zombie_tpo);
2579         if (!list_empty(&zombie_ppo))
2580                 kiblnd_destroy_pool_list(&zombie_ppo);
2581         if (!list_empty(&zombie_fpo))
2582                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2583         if (hdev != NULL)
2584                 kiblnd_hdev_decref(hdev);
2585
2586         if (rc != 0)
2587                 dev->ibd_failed_failover++;
2588         else
2589                 dev->ibd_failed_failover = 0;
2590
2591         return rc;
2592 }
2593
2594 void kiblnd_destroy_dev(kib_dev_t *dev)
2595 {
2596         LASSERT(dev->ibd_nnets == 0);
2597         LASSERT(list_empty(&dev->ibd_nets));
2598
2599         list_del(&dev->ibd_fail_list);
2600         list_del(&dev->ibd_list);
2601
2602         if (dev->ibd_hdev != NULL)
2603                 kiblnd_hdev_decref(dev->ibd_hdev);
2604
2605         LIBCFS_FREE(dev, sizeof(*dev));
2606 }
2607
2608 static kib_dev_t *kiblnd_create_dev(char *ifname)
2609 {
2610         struct net_device *netdev;
2611         kib_dev_t *dev;
2612         __u32 netmask;
2613         __u32 ip;
2614         int up;
2615         int rc;
2616
2617         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2618         if (rc != 0) {
2619                 CERROR("Can't query IPoIB interface %s: %d\n",
2620                        ifname, rc);
2621                 return NULL;
2622         }
2623
2624         if (!up) {
2625                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2626                 return NULL;
2627         }
2628
2629         LIBCFS_ALLOC(dev, sizeof(*dev));
2630         if (dev == NULL)
2631                 return NULL;
2632
2633         netdev = dev_get_by_name(&init_net, ifname);
2634         if (netdev == NULL) {
2635                 dev->ibd_can_failover = 0;
2636         } else {
2637                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2638                 dev_put(netdev);
2639         }
2640
2641         INIT_LIST_HEAD(&dev->ibd_nets);
2642         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2643         INIT_LIST_HEAD(&dev->ibd_fail_list);
2644         dev->ibd_ifip = ip;
2645         strcpy(&dev->ibd_ifname[0], ifname);
2646
2647         /* initialize the device */
2648         rc = kiblnd_dev_failover(dev);
2649         if (rc != 0) {
2650                 CERROR("Can't initialize device: %d\n", rc);
2651                 LIBCFS_FREE(dev, sizeof(*dev));
2652                 return NULL;
2653         }
2654
2655         list_add_tail(&dev->ibd_list,
2656                           &kiblnd_data.kib_devs);
2657         return dev;
2658 }
2659
2660 static void kiblnd_base_shutdown(void)
2661 {
2662         struct kib_sched_info *sched;
2663         int i;
2664
2665         LASSERT(list_empty(&kiblnd_data.kib_devs));
2666
2667         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2668                atomic_read(&libcfs_kmemory));
2669
2670         switch (kiblnd_data.kib_init) {
2671         default:
2672                 LBUG();
2673
2674         case IBLND_INIT_ALL:
2675         case IBLND_INIT_DATA:
2676                 LASSERT(kiblnd_data.kib_peers != NULL);
2677                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2678                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2679                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2680                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2681
2682                 /* flag threads to terminate; wake and wait for them to die */
2683                 kiblnd_data.kib_shutdown = 1;
2684
2685                 /* NB: we really want to stop scheduler threads net by net
2686                  * instead of the whole module, this should be improved
2687                  * with dynamic configuration LNet */
2688                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2689                         wake_up_all(&sched->ibs_waitq);
2690
2691                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2692                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2693
2694                 i = 2;
2695                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2696                         i++;
2697                         /* power of 2 ? */
2698                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2699                                "Waiting for %d threads to terminate\n",
2700                                atomic_read(&kiblnd_data.kib_nthreads));
2701                         set_current_state(TASK_UNINTERRUPTIBLE);
2702                         schedule_timeout(cfs_time_seconds(1));
2703                 }
2704
2705                 /* fall through */
2706
2707         case IBLND_INIT_NOTHING:
2708                 break;
2709         }
2710
2711         if (kiblnd_data.kib_peers != NULL) {
2712                 LIBCFS_FREE(kiblnd_data.kib_peers,
2713                             sizeof(struct list_head) *
2714                             kiblnd_data.kib_peer_hash_size);
2715         }
2716
2717         if (kiblnd_data.kib_scheds != NULL)
2718                 cfs_percpt_free(kiblnd_data.kib_scheds);
2719
2720         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2721                atomic_read(&libcfs_kmemory));
2722
2723         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2724         module_put(THIS_MODULE);
2725 }
2726
2727 void kiblnd_shutdown(lnet_ni_t *ni)
2728 {
2729         kib_net_t *net = ni->ni_data;
2730         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2731         int i;
2732         unsigned long flags;
2733
2734         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2735
2736         if (net == NULL)
2737                 goto out;
2738
2739         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2740                atomic_read(&libcfs_kmemory));
2741
2742         write_lock_irqsave(g_lock, flags);
2743         net->ibn_shutdown = 1;
2744         write_unlock_irqrestore(g_lock, flags);
2745
2746         switch (net->ibn_init) {
2747         default:
2748                 LBUG();
2749
2750         case IBLND_INIT_ALL:
2751                 /* nuke all existing peers within this net */
2752                 kiblnd_del_peer(ni, LNET_NID_ANY);
2753
2754                 /* Wait for all peer state to clean up */
2755                 i = 2;
2756                 while (atomic_read(&net->ibn_npeers) != 0) {
2757                         i++;
2758                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2759                                "%s: waiting for %d peers to disconnect\n",
2760                                libcfs_nid2str(ni->ni_nid),
2761                                atomic_read(&net->ibn_npeers));
2762                         set_current_state(TASK_UNINTERRUPTIBLE);
2763                         schedule_timeout(cfs_time_seconds(1));
2764                 }
2765
2766                 kiblnd_net_fini_pools(net);
2767
2768                 write_lock_irqsave(g_lock, flags);
2769                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2770                 net->ibn_dev->ibd_nnets--;
2771                 list_del(&net->ibn_list);
2772                 write_unlock_irqrestore(g_lock, flags);
2773
2774                 /* fall through */
2775
2776         case IBLND_INIT_NOTHING:
2777                 LASSERT(atomic_read(&net->ibn_nconns) == 0);
2778
2779                 if (net->ibn_dev != NULL &&
2780                     net->ibn_dev->ibd_nnets == 0)
2781                         kiblnd_destroy_dev(net->ibn_dev);
2782
2783                 break;
2784         }
2785
2786         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2787                atomic_read(&libcfs_kmemory));
2788
2789         net->ibn_init = IBLND_INIT_NOTHING;
2790         ni->ni_data = NULL;
2791
2792         LIBCFS_FREE(net, sizeof(*net));
2793
2794 out:
2795         if (list_empty(&kiblnd_data.kib_devs))
2796                 kiblnd_base_shutdown();
2797 }
2798
2799 static int kiblnd_base_startup(void)
2800 {
2801         struct kib_sched_info *sched;
2802         int rc;
2803         int i;
2804
2805         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2806
2807         try_module_get(THIS_MODULE);
2808         /* zero pointers, flags etc */
2809         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2810
2811         rwlock_init(&kiblnd_data.kib_global_lock);
2812
2813         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2814         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2815
2816         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2817         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2818                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2819         if (kiblnd_data.kib_peers == NULL)
2820                 goto failed;
2821         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2822                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2823
2824         spin_lock_init(&kiblnd_data.kib_connd_lock);
2825         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2826         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2827         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2828         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2829
2830         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2831                                                   sizeof(*sched));
2832         if (kiblnd_data.kib_scheds == NULL)
2833                 goto failed;
2834
2835         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2836                 int nthrs;
2837
2838                 spin_lock_init(&sched->ibs_lock);
2839                 INIT_LIST_HEAD(&sched->ibs_conns);
2840                 init_waitqueue_head(&sched->ibs_waitq);
2841
2842                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2843                 if (*kiblnd_tunables.kib_nscheds > 0) {
2844                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2845                 } else {
2846                         /* max to half of CPUs, another half is reserved for
2847                          * upper layer modules */
2848                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2849                 }
2850
2851                 sched->ibs_nthreads_max = nthrs;
2852                 sched->ibs_cpt = i;
2853         }
2854
2855         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2856
2857         /* lists/ptrs/locks initialised */
2858         kiblnd_data.kib_init = IBLND_INIT_DATA;
2859         /*****************************************************/
2860
2861         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2862         if (rc != 0) {
2863                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2864                 goto failed;
2865         }
2866
2867         if (*kiblnd_tunables.kib_dev_failover != 0)
2868                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2869                                          "kiblnd_failover");
2870
2871         if (rc != 0) {
2872                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2873                 goto failed;
2874         }
2875
2876         /* flag everything initialised */
2877         kiblnd_data.kib_init = IBLND_INIT_ALL;
2878         /*****************************************************/
2879
2880         return 0;
2881
2882  failed:
2883         kiblnd_base_shutdown();
2884         return -ENETDOWN;
2885 }
2886
2887 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2888 {
2889         int rc = 0;
2890         int nthrs;
2891         int i;
2892
2893         if (sched->ibs_nthreads == 0) {
2894                 if (*kiblnd_tunables.kib_nscheds > 0) {
2895                         nthrs = sched->ibs_nthreads_max;
2896                 } else {
2897                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2898                                                sched->ibs_cpt);
2899                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2900                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2901                 }
2902         } else {
2903                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2904                 /* increase one thread if there is new interface */
2905                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2906         }
2907
2908         for (i = 0; i < nthrs; i++) {
2909                 long id;
2910                 char name[20];
2911
2912                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2913                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2914                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2915                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2916                 if (rc == 0)
2917                         continue;
2918
2919                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2920                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2921                 break;
2922         }
2923
2924         sched->ibs_nthreads += i;
2925         return rc;
2926 }
2927
2928 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2929                                     int ncpts)
2930 {
2931         int cpt;
2932         int rc;
2933         int i;
2934
2935         for (i = 0; i < ncpts; i++) {
2936                 struct kib_sched_info *sched;
2937
2938                 cpt = (cpts == NULL) ? i : cpts[i];
2939                 sched = kiblnd_data.kib_scheds[cpt];
2940
2941                 if (!newdev && sched->ibs_nthreads > 0)
2942                         continue;
2943
2944                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2945                 if (rc != 0) {
2946                         CERROR("Failed to start scheduler threads for %s\n",
2947                                dev->ibd_ifname);
2948                         return rc;
2949                 }
2950         }
2951         return 0;
2952 }
2953
2954 static kib_dev_t *kiblnd_dev_search(char *ifname)
2955 {
2956         kib_dev_t *alias = NULL;
2957         kib_dev_t *dev;
2958         char *colon;
2959         char *colon2;
2960
2961         colon = strchr(ifname, ':');
2962         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2963                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2964                         return dev;
2965
2966                 if (alias != NULL)
2967                         continue;
2968
2969                 colon2 = strchr(dev->ibd_ifname, ':');
2970                 if (colon != NULL)
2971                         *colon = 0;
2972                 if (colon2 != NULL)
2973                         *colon2 = 0;
2974
2975                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2976                         alias = dev;
2977
2978                 if (colon != NULL)
2979                         *colon = ':';
2980                 if (colon2 != NULL)
2981                         *colon2 = ':';
2982         }
2983         return alias;
2984 }
2985
2986 int kiblnd_startup(lnet_ni_t *ni)
2987 {
2988         char *ifname;
2989         kib_dev_t *ibdev = NULL;
2990         kib_net_t *net;
2991         struct timeval tv;
2992         unsigned long flags;
2993         int rc;
2994         int newdev;
2995
2996         LASSERT(ni->ni_lnd == &the_o2iblnd);
2997
2998         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2999                 rc = kiblnd_base_startup();
3000                 if (rc != 0)
3001                         return rc;
3002         }
3003
3004         LIBCFS_ALLOC(net, sizeof(*net));
3005         ni->ni_data = net;
3006         if (net == NULL)
3007                 goto net_failed;
3008
3009         do_gettimeofday(&tv);
3010         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3011
3012         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3013         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3014         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3015         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3016
3017         if (ni->ni_interfaces[0] != NULL) {
3018                 /* Use the IPoIB interface specified in 'networks=' */
3019
3020                 CLASSERT(LNET_MAX_INTERFACES > 1);
3021                 if (ni->ni_interfaces[1] != NULL) {
3022                         CERROR("Multiple interfaces not supported\n");
3023                         goto failed;
3024                 }
3025
3026                 ifname = ni->ni_interfaces[0];
3027         } else {
3028                 ifname = *kiblnd_tunables.kib_default_ipif;
3029         }
3030
3031         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3032                 CERROR("IPoIB interface name too long: %s\n", ifname);
3033                 goto failed;
3034         }
3035
3036         ibdev = kiblnd_dev_search(ifname);
3037
3038         newdev = ibdev == NULL;
3039         /* hmm...create kib_dev even for alias */
3040         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3041                 ibdev = kiblnd_create_dev(ifname);
3042
3043         if (ibdev == NULL)
3044                 goto failed;
3045
3046         net->ibn_dev = ibdev;
3047         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3048
3049         rc = kiblnd_dev_start_threads(ibdev, newdev,
3050                                       ni->ni_cpts, ni->ni_ncpts);
3051         if (rc != 0)
3052                 goto failed;
3053
3054         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3055         if (rc != 0) {
3056                 CERROR("Failed to initialize NI pools: %d\n", rc);
3057                 goto failed;
3058         }
3059
3060         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3061         ibdev->ibd_nnets++;
3062         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3063         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3064
3065         net->ibn_init = IBLND_INIT_ALL;
3066
3067         return 0;
3068
3069 failed:
3070         if (net->ibn_dev == NULL && ibdev != NULL)
3071                 kiblnd_destroy_dev(ibdev);
3072
3073 net_failed:
3074         kiblnd_shutdown(ni);
3075
3076         CDEBUG(D_NET, "kiblnd_startup failed\n");
3077         return -ENETDOWN;
3078 }
3079
3080 static void __exit kiblnd_module_fini(void)
3081 {
3082         lnet_unregister_lnd(&the_o2iblnd);
3083 }
3084
3085 static int __init kiblnd_module_init(void)
3086 {
3087         int rc;
3088
3089         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3090         CLASSERT(offsetof(kib_msg_t,
3091                 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3092                 <= IBLND_MSG_SIZE);
3093         CLASSERT(offsetof(kib_msg_t,
3094                 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3095                 <= IBLND_MSG_SIZE);
3096
3097         rc = kiblnd_tunables_init();
3098         if (rc != 0)
3099                 return rc;
3100
3101         lnet_register_lnd(&the_o2iblnd);
3102
3103         return 0;
3104 }
3105
3106 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3107 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3108 MODULE_LICENSE("GPL");
3109
3110 module_init(kiblnd_module_init);
3111 module_exit(kiblnd_module_fini);