kernel/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;
1163              i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1164                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1165                 rx = &conn->ibc_rxs[i];
1166
1167                 rx->rx_conn = conn;
1168                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1169
1170                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1171                                                        rx->rx_msg,
1172                                                        IBLND_MSG_SIZE,
1173                                                        DMA_FROM_DEVICE);
1174                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1175                                                    rx->rx_msgaddr));
1176                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1177
1178                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1179                        i, rx->rx_msg, rx->rx_msgaddr,
1180                        lnet_page2phys(pg) + pg_off);
1181
1182                 pg_off += IBLND_MSG_SIZE;
1183                 LASSERT(pg_off <= PAGE_SIZE);
1184
1185                 if (pg_off == PAGE_SIZE) {
1186                         pg_off = 0;
1187                         ipg++;
1188                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1189                 }
1190         }
1191 }
1192
1193 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1194 {
1195         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1196         kib_tx_t       *tx;
1197         int          i;
1198
1199         LASSERT(tpo->tpo_pool.po_allocated == 0);
1200
1201         if (hdev == NULL)
1202                 return;
1203
1204         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1205                 tx = &tpo->tpo_tx_descs[i];
1206                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1207                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1208                                                           tx->tx_msgaddr),
1209                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1210         }
1211
1212         kiblnd_hdev_decref(hdev);
1213         tpo->tpo_hdev = NULL;
1214 }
1215
1216 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1217 {
1218         kib_hca_dev_t *hdev;
1219         unsigned long  flags;
1220         int         i = 0;
1221
1222         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1223         while (dev->ibd_failover) {
1224                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1225                 if (i++ % 50 == 0)
1226                         CDEBUG(D_NET, "%s: Wait for failover\n",
1227                                dev->ibd_ifname);
1228                 schedule_timeout(cfs_time_seconds(1) / 100);
1229
1230                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1231         }
1232
1233         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1234         hdev = dev->ibd_hdev;
1235
1236         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1237
1238         return hdev;
1239 }
1240
1241 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1242 {
1243         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1244         kib_pool_t     *pool  = &tpo->tpo_pool;
1245         kib_net_t      *net   = pool->po_owner->ps_net;
1246         kib_dev_t      *dev;
1247         struct page    *page;
1248         kib_tx_t       *tx;
1249         int          page_offset;
1250         int          ipage;
1251         int          i;
1252
1253         LASSERT(net != NULL);
1254
1255         dev = net->ibn_dev;
1256
1257         /* pre-mapped messages are not bigger than 1 page */
1258         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1259
1260         /* No fancy arithmetic when we do the buffer calculations */
1261         CLASSERT(PAGE_SIZE % IBLND_MSG_SIZE == 0);
1262
1263         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1264
1265         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1266                 page = txpgs->ibp_pages[ipage];
1267                 tx = &tpo->tpo_tx_descs[i];
1268
1269                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1270                                            page_offset);
1271
1272                 tx->tx_msgaddr = kiblnd_dma_map_single(
1273                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1274                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1275                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1276                                                    tx->tx_msgaddr));
1277                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1278
1279                 list_add(&tx->tx_list, &pool->po_free_list);
1280
1281                 page_offset += IBLND_MSG_SIZE;
1282                 LASSERT(page_offset <= PAGE_SIZE);
1283
1284                 if (page_offset == PAGE_SIZE) {
1285                         page_offset = 0;
1286                         ipage++;
1287                         LASSERT(ipage <= txpgs->ibp_npages);
1288                 }
1289         }
1290 }
1291
1292 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1293 {
1294         __u64   index;
1295
1296         LASSERT(hdev->ibh_mrs[0] != NULL);
1297
1298         if (hdev->ibh_nmrs == 1)
1299                 return hdev->ibh_mrs[0];
1300
1301         index = addr >> hdev->ibh_mr_shift;
1302
1303         if (index <  hdev->ibh_nmrs &&
1304             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1305                 return hdev->ibh_mrs[index];
1306
1307         return NULL;
1308 }
1309
1310 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1311 {
1312         struct ib_mr *prev_mr;
1313         struct ib_mr *mr;
1314         int        i;
1315
1316         LASSERT(hdev->ibh_mrs[0] != NULL);
1317
1318         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1319             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1320                 return NULL;
1321
1322         if (hdev->ibh_nmrs == 1)
1323                 return hdev->ibh_mrs[0];
1324
1325         for (i = 0, mr = prev_mr = NULL;
1326              i < rd->rd_nfrags; i++) {
1327                 mr = kiblnd_find_dma_mr(hdev,
1328                                         rd->rd_frags[i].rf_addr,
1329                                         rd->rd_frags[i].rf_nob);
1330                 if (prev_mr == NULL)
1331                         prev_mr = mr;
1332
1333                 if (mr == NULL || prev_mr != mr) {
1334                         /* Can't covered by one single MR */
1335                         mr = NULL;
1336                         break;
1337                 }
1338         }
1339
1340         return mr;
1341 }
1342
1343 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1344 {
1345         LASSERT(pool->fpo_map_count == 0);
1346
1347         if (pool->fpo_fmr_pool != NULL)
1348                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1349
1350         if (pool->fpo_hdev != NULL)
1351                 kiblnd_hdev_decref(pool->fpo_hdev);
1352
1353         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1354 }
1355
1356 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1357 {
1358         kib_fmr_pool_t *pool;
1359
1360         while (!list_empty(head)) {
1361                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1362                 list_del(&pool->fpo_list);
1363                 kiblnd_destroy_fmr_pool(pool);
1364         }
1365 }
1366
1367 static int kiblnd_fmr_pool_size(int ncpts)
1368 {
1369         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1370
1371         return max(IBLND_FMR_POOL, size);
1372 }
1373
1374 static int kiblnd_fmr_flush_trigger(int ncpts)
1375 {
1376         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1377
1378         return max(IBLND_FMR_POOL_FLUSH, size);
1379 }
1380
1381 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1382                                   kib_fmr_pool_t **pp_fpo)
1383 {
1384         /* FMR pool for RDMA */
1385         kib_dev_t              *dev = fps->fps_net->ibn_dev;
1386         kib_fmr_pool_t    *fpo;
1387         struct ib_fmr_pool_param param = {
1388                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1389                 .page_shift     = PAGE_SHIFT,
1390                 .access     = (IB_ACCESS_LOCAL_WRITE |
1391                                       IB_ACCESS_REMOTE_WRITE),
1392                 .pool_size         = fps->fps_pool_size,
1393                 .dirty_watermark   = fps->fps_flush_trigger,
1394                 .flush_function    = NULL,
1395                 .flush_arg       = NULL,
1396                 .cache       = !!*kiblnd_tunables.kib_fmr_cache};
1397         int rc;
1398
1399         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1400         if (fpo == NULL)
1401                 return -ENOMEM;
1402
1403         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1404
1405         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1406         if (IS_ERR(fpo->fpo_fmr_pool)) {
1407                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1408                 CERROR("Failed to create FMR pool: %d\n", rc);
1409
1410                 kiblnd_hdev_decref(fpo->fpo_hdev);
1411                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1412                 return rc;
1413         }
1414
1415         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1416         fpo->fpo_owner    = fps;
1417         *pp_fpo = fpo;
1418
1419         return 0;
1420 }
1421
1422 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1423                                     struct list_head *zombies)
1424 {
1425         if (fps->fps_net == NULL) /* intialized? */
1426                 return;
1427
1428         spin_lock(&fps->fps_lock);
1429
1430         while (!list_empty(&fps->fps_pool_list)) {
1431                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1432                                                  kib_fmr_pool_t, fpo_list);
1433                 fpo->fpo_failed = 1;
1434                 list_del(&fpo->fpo_list);
1435                 if (fpo->fpo_map_count == 0)
1436                         list_add(&fpo->fpo_list, zombies);
1437                 else
1438                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1439         }
1440
1441         spin_unlock(&fps->fps_lock);
1442 }
1443
1444 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1445 {
1446         if (fps->fps_net != NULL) { /* initialized? */
1447                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1448                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1449         }
1450 }
1451
1452 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1453                                    kib_net_t *net, int pool_size,
1454                                    int flush_trigger)
1455 {
1456         kib_fmr_pool_t *fpo;
1457         int          rc;
1458
1459         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1460
1461         fps->fps_net = net;
1462         fps->fps_cpt = cpt;
1463         fps->fps_pool_size = pool_size;
1464         fps->fps_flush_trigger = flush_trigger;
1465         spin_lock_init(&fps->fps_lock);
1466         INIT_LIST_HEAD(&fps->fps_pool_list);
1467         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1468
1469         rc = kiblnd_create_fmr_pool(fps, &fpo);
1470         if (rc == 0)
1471                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1472
1473         return rc;
1474 }
1475
1476 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1477 {
1478         if (fpo->fpo_map_count != 0) /* still in use */
1479                 return 0;
1480         if (fpo->fpo_failed)
1481                 return 1;
1482         return cfs_time_aftereq(now, fpo->fpo_deadline);
1483 }
1484
1485 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1486 {
1487         LIST_HEAD(zombies);
1488         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1489         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1490         unsigned long    now = cfs_time_current();
1491         kib_fmr_pool_t    *tmp;
1492         int             rc;
1493
1494         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1495         LASSERT(rc == 0);
1496
1497         if (status != 0) {
1498                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1499                 LASSERT(rc == 0);
1500         }
1501
1502         fmr->fmr_pool = NULL;
1503         fmr->fmr_pfmr = NULL;
1504
1505         spin_lock(&fps->fps_lock);
1506         fpo->fpo_map_count--;  /* decref the pool */
1507
1508         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1509                 /* the first pool is persistent */
1510                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1511                         continue;
1512
1513                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1514                         list_move(&fpo->fpo_list, &zombies);
1515                         fps->fps_version++;
1516                 }
1517         }
1518         spin_unlock(&fps->fps_lock);
1519
1520         if (!list_empty(&zombies))
1521                 kiblnd_destroy_fmr_pool_list(&zombies);
1522 }
1523
1524 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1525                         __u64 iov, kib_fmr_t *fmr)
1526 {
1527         struct ib_pool_fmr *pfmr;
1528         kib_fmr_pool_t     *fpo;
1529         __u64          version;
1530         int              rc;
1531
1532  again:
1533         spin_lock(&fps->fps_lock);
1534         version = fps->fps_version;
1535         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1536                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1537                 fpo->fpo_map_count++;
1538                 spin_unlock(&fps->fps_lock);
1539
1540                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1541                                             pages, npages, iov);
1542                 if (likely(!IS_ERR(pfmr))) {
1543                         fmr->fmr_pool = fpo;
1544                         fmr->fmr_pfmr = pfmr;
1545                         return 0;
1546                 }
1547
1548                 spin_lock(&fps->fps_lock);
1549                 fpo->fpo_map_count--;
1550                 if (PTR_ERR(pfmr) != -EAGAIN) {
1551                         spin_unlock(&fps->fps_lock);
1552                         return PTR_ERR(pfmr);
1553                 }
1554
1555                 /* EAGAIN and ... */
1556                 if (version != fps->fps_version) {
1557                         spin_unlock(&fps->fps_lock);
1558                         goto again;
1559                 }
1560         }
1561
1562         if (fps->fps_increasing) {
1563                 spin_unlock(&fps->fps_lock);
1564                 CDEBUG(D_NET,
1565                         "Another thread is allocating new FMR pool, waiting for her to complete\n");
1566                 schedule();
1567                 goto again;
1568
1569         }
1570
1571         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1572                 /* someone failed recently */
1573                 spin_unlock(&fps->fps_lock);
1574                 return -EAGAIN;
1575         }
1576
1577         fps->fps_increasing = 1;
1578         spin_unlock(&fps->fps_lock);
1579
1580         CDEBUG(D_NET, "Allocate new FMR pool\n");
1581         rc = kiblnd_create_fmr_pool(fps, &fpo);
1582         spin_lock(&fps->fps_lock);
1583         fps->fps_increasing = 0;
1584         if (rc == 0) {
1585                 fps->fps_version++;
1586                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1587         } else {
1588                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1589         }
1590         spin_unlock(&fps->fps_lock);
1591
1592         goto again;
1593 }
1594
1595 static void kiblnd_fini_pool(kib_pool_t *pool)
1596 {
1597         LASSERT(list_empty(&pool->po_free_list));
1598         LASSERT(pool->po_allocated == 0);
1599
1600         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1601 }
1602
1603 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1604 {
1605         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1606
1607         memset(pool, 0, sizeof(kib_pool_t));
1608         INIT_LIST_HEAD(&pool->po_free_list);
1609         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1610         pool->po_owner    = ps;
1611         pool->po_size     = size;
1612 }
1613
1614 static void kiblnd_destroy_pool_list(struct list_head *head)
1615 {
1616         kib_pool_t *pool;
1617
1618         while (!list_empty(head)) {
1619                 pool = list_entry(head->next, kib_pool_t, po_list);
1620                 list_del(&pool->po_list);
1621
1622                 LASSERT(pool->po_owner != NULL);
1623                 pool->po_owner->ps_pool_destroy(pool);
1624         }
1625 }
1626
1627 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1628 {
1629         if (ps->ps_net == NULL) /* intialized? */
1630                 return;
1631
1632         spin_lock(&ps->ps_lock);
1633         while (!list_empty(&ps->ps_pool_list)) {
1634                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1635                                             kib_pool_t, po_list);
1636                 po->po_failed = 1;
1637                 list_del(&po->po_list);
1638                 if (po->po_allocated == 0)
1639                         list_add(&po->po_list, zombies);
1640                 else
1641                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1642         }
1643         spin_unlock(&ps->ps_lock);
1644 }
1645
1646 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1647 {
1648         if (ps->ps_net != NULL) { /* initialized? */
1649                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1650                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1651         }
1652 }
1653
1654 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1655                                kib_net_t *net, char *name, int size,
1656                                kib_ps_pool_create_t po_create,
1657                                kib_ps_pool_destroy_t po_destroy,
1658                                kib_ps_node_init_t nd_init,
1659                                kib_ps_node_fini_t nd_fini)
1660 {
1661         kib_pool_t      *pool;
1662         int             rc;
1663
1664         memset(ps, 0, sizeof(kib_poolset_t));
1665
1666         ps->ps_cpt          = cpt;
1667         ps->ps_net        = net;
1668         ps->ps_pool_create  = po_create;
1669         ps->ps_pool_destroy = po_destroy;
1670         ps->ps_node_init    = nd_init;
1671         ps->ps_node_fini    = nd_fini;
1672         ps->ps_pool_size    = size;
1673         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1674             >= sizeof(ps->ps_name))
1675                 return -E2BIG;
1676         spin_lock_init(&ps->ps_lock);
1677         INIT_LIST_HEAD(&ps->ps_pool_list);
1678         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1679
1680         rc = ps->ps_pool_create(ps, size, &pool);
1681         if (rc == 0)
1682                 list_add(&pool->po_list, &ps->ps_pool_list);
1683         else
1684                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1685
1686         return rc;
1687 }
1688
1689 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1690 {
1691         if (pool->po_allocated != 0) /* still in use */
1692                 return 0;
1693         if (pool->po_failed)
1694                 return 1;
1695         return cfs_time_aftereq(now, pool->po_deadline);
1696 }
1697
1698 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1699 {
1700         LIST_HEAD(zombies);
1701         kib_poolset_t  *ps = pool->po_owner;
1702         kib_pool_t     *tmp;
1703         unsigned long      now = cfs_time_current();
1704
1705         spin_lock(&ps->ps_lock);
1706
1707         if (ps->ps_node_fini != NULL)
1708                 ps->ps_node_fini(pool, node);
1709
1710         LASSERT(pool->po_allocated > 0);
1711         list_add(node, &pool->po_free_list);
1712         pool->po_allocated--;
1713
1714         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1715                 /* the first pool is persistent */
1716                 if (ps->ps_pool_list.next == &pool->po_list)
1717                         continue;
1718
1719                 if (kiblnd_pool_is_idle(pool, now))
1720                         list_move(&pool->po_list, &zombies);
1721         }
1722         spin_unlock(&ps->ps_lock);
1723
1724         if (!list_empty(&zombies))
1725                 kiblnd_destroy_pool_list(&zombies);
1726 }
1727
1728 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1729 {
1730         struct list_head            *node;
1731         kib_pool_t          *pool;
1732         int                 rc;
1733
1734  again:
1735         spin_lock(&ps->ps_lock);
1736         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1737                 if (list_empty(&pool->po_free_list))
1738                         continue;
1739
1740                 pool->po_allocated++;
1741                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1742                 node = pool->po_free_list.next;
1743                 list_del(node);
1744
1745                 if (ps->ps_node_init != NULL) {
1746                         /* still hold the lock */
1747                         ps->ps_node_init(pool, node);
1748                 }
1749                 spin_unlock(&ps->ps_lock);
1750                 return node;
1751         }
1752
1753         /* no available tx pool and ... */
1754         if (ps->ps_increasing) {
1755                 /* another thread is allocating a new pool */
1756                 spin_unlock(&ps->ps_lock);
1757                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1758                        ps->ps_name);
1759                 schedule();
1760                 goto again;
1761         }
1762
1763         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1764                 /* someone failed recently */
1765                 spin_unlock(&ps->ps_lock);
1766                 return NULL;
1767         }
1768
1769         ps->ps_increasing = 1;
1770         spin_unlock(&ps->ps_lock);
1771
1772         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1773
1774         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1775
1776         spin_lock(&ps->ps_lock);
1777         ps->ps_increasing = 0;
1778         if (rc == 0) {
1779                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1780         } else {
1781                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1782                 CERROR("Can't allocate new %s pool because out of memory\n",
1783                        ps->ps_name);
1784         }
1785         spin_unlock(&ps->ps_lock);
1786
1787         goto again;
1788 }
1789
1790 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr)
1791 {
1792         kib_pmr_pool_t      *ppo = pmr->pmr_pool;
1793         struct ib_mr    *mr  = pmr->pmr_mr;
1794
1795         pmr->pmr_mr = NULL;
1796         kiblnd_pool_free_node(&ppo->ppo_pool, &pmr->pmr_list);
1797         if (mr != NULL)
1798                 ib_dereg_mr(mr);
1799 }
1800
1801 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1802                     kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr)
1803 {
1804         kib_phys_mr_t *pmr;
1805         struct list_head    *node;
1806         int         rc;
1807         int         i;
1808
1809         node = kiblnd_pool_alloc_node(&pps->pps_poolset);
1810         if (node == NULL) {
1811                 CERROR("Failed to allocate PMR descriptor\n");
1812                 return -ENOMEM;
1813         }
1814
1815         pmr = container_of(node, kib_phys_mr_t, pmr_list);
1816         if (pmr->pmr_pool->ppo_hdev != hdev) {
1817                 kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1818                 return -EAGAIN;
1819         }
1820
1821         for (i = 0; i < rd->rd_nfrags; i++) {
1822                 pmr->pmr_ipb[i].addr = rd->rd_frags[i].rf_addr;
1823                 pmr->pmr_ipb[i].size = rd->rd_frags[i].rf_nob;
1824         }
1825
1826         pmr->pmr_mr = ib_reg_phys_mr(hdev->ibh_pd,
1827                                      pmr->pmr_ipb, rd->rd_nfrags,
1828                                      IB_ACCESS_LOCAL_WRITE |
1829                                      IB_ACCESS_REMOTE_WRITE,
1830                                      iova);
1831         if (!IS_ERR(pmr->pmr_mr)) {
1832                 pmr->pmr_iova = *iova;
1833                 *pp_pmr = pmr;
1834                 return 0;
1835         }
1836
1837         rc = PTR_ERR(pmr->pmr_mr);
1838         CERROR("Failed ib_reg_phys_mr: %d\n", rc);
1839
1840         pmr->pmr_mr = NULL;
1841         kiblnd_pool_free_node(&pmr->pmr_pool->ppo_pool, node);
1842
1843         return rc;
1844 }
1845
1846 static void kiblnd_destroy_pmr_pool(kib_pool_t *pool)
1847 {
1848         kib_pmr_pool_t *ppo = container_of(pool, kib_pmr_pool_t, ppo_pool);
1849         kib_phys_mr_t  *pmr;
1850         kib_phys_mr_t *tmp;
1851
1852         LASSERT(pool->po_allocated == 0);
1853
1854         list_for_each_entry_safe(pmr, tmp, &pool->po_free_list, pmr_list) {
1855                 LASSERT(pmr->pmr_mr == NULL);
1856                 list_del(&pmr->pmr_list);
1857
1858                 if (pmr->pmr_ipb != NULL) {
1859                         LIBCFS_FREE(pmr->pmr_ipb,
1860                                     IBLND_MAX_RDMA_FRAGS *
1861                                     sizeof(struct ib_phys_buf));
1862                 }
1863
1864                 LIBCFS_FREE(pmr, sizeof(kib_phys_mr_t));
1865         }
1866
1867         kiblnd_fini_pool(pool);
1868         if (ppo->ppo_hdev != NULL)
1869                 kiblnd_hdev_decref(ppo->ppo_hdev);
1870
1871         LIBCFS_FREE(ppo, sizeof(kib_pmr_pool_t));
1872 }
1873
1874 static inline int kiblnd_pmr_pool_size(int ncpts)
1875 {
1876         int size = *kiblnd_tunables.kib_pmr_pool_size / ncpts;
1877
1878         return max(IBLND_PMR_POOL, size);
1879 }
1880
1881 static int kiblnd_create_pmr_pool(kib_poolset_t *ps, int size,
1882                                   kib_pool_t **pp_po)
1883 {
1884         struct kib_pmr_pool     *ppo;
1885         struct kib_pool         *pool;
1886         kib_phys_mr_t           *pmr;
1887         int                     i;
1888
1889         LIBCFS_CPT_ALLOC(ppo, lnet_cpt_table(),
1890                          ps->ps_cpt, sizeof(kib_pmr_pool_t));
1891         if (ppo == NULL) {
1892                 CERROR("Failed to allocate PMR pool\n");
1893                 return -ENOMEM;
1894         }
1895
1896         pool = &ppo->ppo_pool;
1897         kiblnd_init_pool(ps, pool, size);
1898
1899         for (i = 0; i < size; i++) {
1900                 LIBCFS_CPT_ALLOC(pmr, lnet_cpt_table(),
1901                                  ps->ps_cpt, sizeof(kib_phys_mr_t));
1902                 if (pmr == NULL)
1903                         break;
1904
1905                 pmr->pmr_pool = ppo;
1906                 LIBCFS_CPT_ALLOC(pmr->pmr_ipb, lnet_cpt_table(), ps->ps_cpt,
1907                                  IBLND_MAX_RDMA_FRAGS * sizeof(*pmr->pmr_ipb));
1908                 if (pmr->pmr_ipb == NULL)
1909                         break;
1910
1911                 list_add(&pmr->pmr_list, &pool->po_free_list);
1912         }
1913
1914         if (i < size) {
1915                 ps->ps_pool_destroy(pool);
1916                 return -ENOMEM;
1917         }
1918
1919         ppo->ppo_hdev = kiblnd_current_hdev(ps->ps_net->ibn_dev);
1920         *pp_po = pool;
1921         return 0;
1922 }
1923
1924 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1925 {
1926         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1927         int          i;
1928
1929         LASSERT(pool->po_allocated == 0);
1930
1931         if (tpo->tpo_tx_pages != NULL) {
1932                 kiblnd_unmap_tx_pool(tpo);
1933                 kiblnd_free_pages(tpo->tpo_tx_pages);
1934         }
1935
1936         if (tpo->tpo_tx_descs == NULL)
1937                 goto out;
1938
1939         for (i = 0; i < pool->po_size; i++) {
1940                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1941
1942                 list_del(&tx->tx_list);
1943                 if (tx->tx_pages != NULL)
1944                         LIBCFS_FREE(tx->tx_pages,
1945                                     LNET_MAX_IOV *
1946                                     sizeof(*tx->tx_pages));
1947                 if (tx->tx_frags != NULL)
1948                         LIBCFS_FREE(tx->tx_frags,
1949                                     IBLND_MAX_RDMA_FRAGS *
1950                                             sizeof(*tx->tx_frags));
1951                 if (tx->tx_wrq != NULL)
1952                         LIBCFS_FREE(tx->tx_wrq,
1953                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1954                                     sizeof(*tx->tx_wrq));
1955                 if (tx->tx_sge != NULL)
1956                         LIBCFS_FREE(tx->tx_sge,
1957                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1958                                     sizeof(*tx->tx_sge));
1959                 if (tx->tx_rd != NULL)
1960                         LIBCFS_FREE(tx->tx_rd,
1961                                     offsetof(kib_rdma_desc_t,
1962                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1963         }
1964
1965         LIBCFS_FREE(tpo->tpo_tx_descs,
1966                     pool->po_size * sizeof(kib_tx_t));
1967 out:
1968         kiblnd_fini_pool(pool);
1969         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1970 }
1971
1972 static int kiblnd_tx_pool_size(int ncpts)
1973 {
1974         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1975
1976         return max(IBLND_TX_POOL, ntx);
1977 }
1978
1979 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1980                                  kib_pool_t **pp_po)
1981 {
1982         int         i;
1983         int         npg;
1984         kib_pool_t    *pool;
1985         kib_tx_pool_t *tpo;
1986
1987         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1988         if (tpo == NULL) {
1989                 CERROR("Failed to allocate TX pool\n");
1990                 return -ENOMEM;
1991         }
1992
1993         pool = &tpo->tpo_pool;
1994         kiblnd_init_pool(ps, pool, size);
1995         tpo->tpo_tx_descs = NULL;
1996         tpo->tpo_tx_pages = NULL;
1997
1998         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1999         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2000                 CERROR("Can't allocate tx pages: %d\n", npg);
2001                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2002                 return -ENOMEM;
2003         }
2004
2005         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2006                          size * sizeof(kib_tx_t));
2007         if (tpo->tpo_tx_descs == NULL) {
2008                 CERROR("Can't allocate %d tx descriptors\n", size);
2009                 ps->ps_pool_destroy(pool);
2010                 return -ENOMEM;
2011         }
2012
2013         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2014
2015         for (i = 0; i < size; i++) {
2016                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2017
2018                 tx->tx_pool = tpo;
2019                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2020                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2021                                          lnet_cpt_table(), ps->ps_cpt,
2022                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2023                         if (tx->tx_pages == NULL)
2024                                 break;
2025                 }
2026
2027                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2028                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
2029                 if (tx->tx_frags == NULL)
2030                         break;
2031
2032                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
2033
2034                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2035                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2036                                  sizeof(*tx->tx_wrq));
2037                 if (tx->tx_wrq == NULL)
2038                         break;
2039
2040                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2041                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2042                                  sizeof(*tx->tx_sge));
2043                 if (tx->tx_sge == NULL)
2044                         break;
2045
2046                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2047                                  offsetof(kib_rdma_desc_t,
2048                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2049                 if (tx->tx_rd == NULL)
2050                         break;
2051         }
2052
2053         if (i == size) {
2054                 kiblnd_map_tx_pool(tpo);
2055                 *pp_po = pool;
2056                 return 0;
2057         }
2058
2059         ps->ps_pool_destroy(pool);
2060         return -ENOMEM;
2061 }
2062
2063 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2064 {
2065         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2066                                              tps_poolset);
2067         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2068
2069         tx->tx_cookie = tps->tps_next_tx_cookie++;
2070 }
2071
2072 static void kiblnd_net_fini_pools(kib_net_t *net)
2073 {
2074         int     i;
2075
2076         cfs_cpt_for_each(i, lnet_cpt_table()) {
2077                 kib_tx_poolset_t        *tps;
2078                 kib_fmr_poolset_t       *fps;
2079                 kib_pmr_poolset_t       *pps;
2080
2081                 if (net->ibn_tx_ps != NULL) {
2082                         tps = net->ibn_tx_ps[i];
2083                         kiblnd_fini_poolset(&tps->tps_poolset);
2084                 }
2085
2086                 if (net->ibn_fmr_ps != NULL) {
2087                         fps = net->ibn_fmr_ps[i];
2088                         kiblnd_fini_fmr_poolset(fps);
2089                 }
2090
2091                 if (net->ibn_pmr_ps != NULL) {
2092                         pps = net->ibn_pmr_ps[i];
2093                         kiblnd_fini_poolset(&pps->pps_poolset);
2094                 }
2095         }
2096
2097         if (net->ibn_tx_ps != NULL) {
2098                 cfs_percpt_free(net->ibn_tx_ps);
2099                 net->ibn_tx_ps = NULL;
2100         }
2101
2102         if (net->ibn_fmr_ps != NULL) {
2103                 cfs_percpt_free(net->ibn_fmr_ps);
2104                 net->ibn_fmr_ps = NULL;
2105         }
2106
2107         if (net->ibn_pmr_ps != NULL) {
2108                 cfs_percpt_free(net->ibn_pmr_ps);
2109                 net->ibn_pmr_ps = NULL;
2110         }
2111 }
2112
2113 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
2114 {
2115         unsigned long   flags;
2116         int             cpt;
2117         int             rc;
2118         int             i;
2119
2120         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2121         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
2122             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
2123                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2124                                            flags);
2125                 goto create_tx_pool;
2126         }
2127
2128         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2129
2130         if (*kiblnd_tunables.kib_fmr_pool_size <
2131             *kiblnd_tunables.kib_ntx / 4) {
2132                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2133                        *kiblnd_tunables.kib_fmr_pool_size,
2134                        *kiblnd_tunables.kib_ntx / 4);
2135                 rc = -EINVAL;
2136                 goto failed;
2137         }
2138
2139         /* TX pool must be created later than FMR/PMR, see LU-2268
2140          * for details */
2141         LASSERT(net->ibn_tx_ps == NULL);
2142
2143         /* premapping can fail if ibd_nmr > 1, so we always create
2144          * FMR/PMR pool and map-on-demand if premapping failed */
2145
2146         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2147                                            sizeof(kib_fmr_poolset_t));
2148         if (net->ibn_fmr_ps == NULL) {
2149                 CERROR("Failed to allocate FMR pool array\n");
2150                 rc = -ENOMEM;
2151                 goto failed;
2152         }
2153
2154         for (i = 0; i < ncpts; i++) {
2155                 cpt = (cpts == NULL) ? i : cpts[i];
2156                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2157                                              kiblnd_fmr_pool_size(ncpts),
2158                                              kiblnd_fmr_flush_trigger(ncpts));
2159                 if (rc == -ENOSYS && i == 0) /* no FMR */
2160                         break; /* create PMR pool */
2161
2162                 if (rc != 0) { /* a real error */
2163                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2164                                cpt, rc);
2165                         goto failed;
2166                 }
2167         }
2168
2169         if (i > 0) {
2170                 LASSERT(i == ncpts);
2171                 goto create_tx_pool;
2172         }
2173
2174         cfs_percpt_free(net->ibn_fmr_ps);
2175         net->ibn_fmr_ps = NULL;
2176
2177         CWARN("Device does not support FMR, failing back to PMR\n");
2178
2179         if (*kiblnd_tunables.kib_pmr_pool_size <
2180             *kiblnd_tunables.kib_ntx / 4) {
2181                 CERROR("Can't set pmr pool size (%d) < ntx / 4(%d)\n",
2182                        *kiblnd_tunables.kib_pmr_pool_size,
2183                        *kiblnd_tunables.kib_ntx / 4);
2184                 rc = -EINVAL;
2185                 goto failed;
2186         }
2187
2188         net->ibn_pmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2189                                            sizeof(kib_pmr_poolset_t));
2190         if (net->ibn_pmr_ps == NULL) {
2191                 CERROR("Failed to allocate PMR pool array\n");
2192                 rc = -ENOMEM;
2193                 goto failed;
2194         }
2195
2196         for (i = 0; i < ncpts; i++) {
2197                 cpt = (cpts == NULL) ? i : cpts[i];
2198                 rc = kiblnd_init_poolset(&net->ibn_pmr_ps[cpt]->pps_poolset,
2199                                          cpt, net, "PMR",
2200                                          kiblnd_pmr_pool_size(ncpts),
2201                                          kiblnd_create_pmr_pool,
2202                                          kiblnd_destroy_pmr_pool, NULL, NULL);
2203                 if (rc != 0) {
2204                         CERROR("Can't initialize PMR pool for CPT %d: %d\n",
2205                                cpt, rc);
2206                         goto failed;
2207                 }
2208         }
2209
2210  create_tx_pool:
2211         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2212                                           sizeof(kib_tx_poolset_t));
2213         if (net->ibn_tx_ps == NULL) {
2214                 CERROR("Failed to allocate tx pool array\n");
2215                 rc = -ENOMEM;
2216                 goto failed;
2217         }
2218
2219         for (i = 0; i < ncpts; i++) {
2220                 cpt = (cpts == NULL) ? i : cpts[i];
2221                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2222                                          cpt, net, "TX",
2223                                          kiblnd_tx_pool_size(ncpts),
2224                                          kiblnd_create_tx_pool,
2225                                          kiblnd_destroy_tx_pool,
2226                                          kiblnd_tx_init, NULL);
2227                 if (rc != 0) {
2228                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2229                                cpt, rc);
2230                         goto failed;
2231                 }
2232         }
2233
2234         return 0;
2235  failed:
2236         kiblnd_net_fini_pools(net);
2237         LASSERT(rc != 0);
2238         return rc;
2239 }
2240
2241 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2242 {
2243         struct ib_device_attr *attr;
2244         int                 rc;
2245
2246         /* It's safe to assume a HCA can handle a page size
2247          * matching that of the native system */
2248         hdev->ibh_page_shift = PAGE_SHIFT;
2249         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2250         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2251
2252         LIBCFS_ALLOC(attr, sizeof(*attr));
2253         if (attr == NULL) {
2254                 CERROR("Out of memory\n");
2255                 return -ENOMEM;
2256         }
2257
2258         rc = ib_query_device(hdev->ibh_ibdev, attr);
2259         if (rc == 0)
2260                 hdev->ibh_mr_size = attr->max_mr_size;
2261
2262         LIBCFS_FREE(attr, sizeof(*attr));
2263
2264         if (rc != 0) {
2265                 CERROR("Failed to query IB device: %d\n", rc);
2266                 return rc;
2267         }
2268
2269         if (hdev->ibh_mr_size == ~0ULL) {
2270                 hdev->ibh_mr_shift = 64;
2271                 return 0;
2272         }
2273
2274         for (hdev->ibh_mr_shift = 0;
2275              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2276                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2277                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2278                         return 0;
2279         }
2280
2281         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2282         return -EINVAL;
2283 }
2284
2285 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2286 {
2287         int     i;
2288
2289         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2290                 return;
2291
2292         for (i = 0; i < hdev->ibh_nmrs; i++) {
2293                 if (hdev->ibh_mrs[i] == NULL)
2294                         break;
2295
2296                 ib_dereg_mr(hdev->ibh_mrs[i]);
2297         }
2298
2299         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2300         hdev->ibh_mrs  = NULL;
2301         hdev->ibh_nmrs = 0;
2302 }
2303
2304 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2305 {
2306         kiblnd_hdev_cleanup_mrs(hdev);
2307
2308         if (hdev->ibh_pd != NULL)
2309                 ib_dealloc_pd(hdev->ibh_pd);
2310
2311         if (hdev->ibh_cmid != NULL)
2312                 rdma_destroy_id(hdev->ibh_cmid);
2313
2314         LIBCFS_FREE(hdev, sizeof(*hdev));
2315 }
2316
2317 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2318 {
2319         struct ib_mr *mr;
2320         int        i;
2321         int        rc;
2322         __u64    mm_size;
2323         __u64    mr_size;
2324         int        acflags = IB_ACCESS_LOCAL_WRITE |
2325                                 IB_ACCESS_REMOTE_WRITE;
2326
2327         rc = kiblnd_hdev_get_attr(hdev);
2328         if (rc != 0)
2329                 return rc;
2330
2331         if (hdev->ibh_mr_shift == 64) {
2332                 LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2333                 if (hdev->ibh_mrs == NULL) {
2334                         CERROR("Failed to allocate MRs table\n");
2335                         return -ENOMEM;
2336                 }
2337
2338                 hdev->ibh_mrs[0] = NULL;
2339                 hdev->ibh_nmrs   = 1;
2340
2341                 mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2342                 if (IS_ERR(mr)) {
2343                         CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2344                         kiblnd_hdev_cleanup_mrs(hdev);
2345                         return PTR_ERR(mr);
2346                 }
2347
2348                 hdev->ibh_mrs[0] = mr;
2349
2350                 goto out;
2351         }
2352
2353         mr_size = 1ULL << hdev->ibh_mr_shift;
2354         mm_size = (unsigned long)high_memory - PAGE_OFFSET;
2355
2356         hdev->ibh_nmrs = (int)((mm_size + mr_size - 1) >> hdev->ibh_mr_shift);
2357
2358         if (hdev->ibh_mr_shift < 32 || hdev->ibh_nmrs > 1024) {
2359                 /* it's 4T..., assume we will re-code at that time */
2360                 CERROR("Can't support memory size: x%#llx with MR size: x%#llx\n",
2361                        mm_size, mr_size);
2362                 return -EINVAL;
2363         }
2364
2365         /* create an array of MRs to cover all memory */
2366         LIBCFS_ALLOC(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2367         if (hdev->ibh_mrs == NULL) {
2368                 CERROR("Failed to allocate MRs' table\n");
2369                 return -ENOMEM;
2370         }
2371
2372         for (i = 0; i < hdev->ibh_nmrs; i++) {
2373                 struct ib_phys_buf ipb;
2374                 __u64         iova;
2375
2376                 ipb.size = hdev->ibh_mr_size;
2377                 ipb.addr = i * mr_size;
2378                 iova     = ipb.addr;
2379
2380                 mr = ib_reg_phys_mr(hdev->ibh_pd, &ipb, 1, acflags, &iova);
2381                 if (IS_ERR(mr)) {
2382                         CERROR("Failed ib_reg_phys_mr addr %#llx size %#llx : %ld\n",
2383                                ipb.addr, ipb.size, PTR_ERR(mr));
2384                         kiblnd_hdev_cleanup_mrs(hdev);
2385                         return PTR_ERR(mr);
2386                 }
2387
2388                 LASSERT(iova == ipb.addr);
2389
2390                 hdev->ibh_mrs[i] = mr;
2391         }
2392
2393 out:
2394         if (hdev->ibh_mr_size != ~0ULL || hdev->ibh_nmrs != 1)
2395                 LCONSOLE_INFO("Register global MR array, MR size: %#llx, array size: %d\n",
2396                               hdev->ibh_mr_size, hdev->ibh_nmrs);
2397         return 0;
2398 }
2399
2400 /* DUMMY */
2401 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2402                                  struct rdma_cm_event *event)
2403 {
2404         return 0;
2405 }
2406
2407 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2408 {
2409         struct rdma_cm_id  *cmid;
2410         struct sockaddr_in  srcaddr;
2411         struct sockaddr_in  dstaddr;
2412         int              rc;
2413
2414         if (dev->ibd_hdev == NULL || /* initializing */
2415             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2416             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2417                 return 1;
2418
2419         /* XXX: it's UGLY, but I don't have better way to find
2420          * ib-bonding HCA failover because:
2421          *
2422          * a. no reliable CM event for HCA failover...
2423          * b. no OFED API to get ib_device for current net_device...
2424          *
2425          * We have only two choices at this point:
2426          *
2427          * a. rdma_bind_addr(), it will conflict with listener cmid
2428          * b. rdma_resolve_addr() to zero addr */
2429         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2430                                      IB_QPT_RC);
2431         if (IS_ERR(cmid)) {
2432                 rc = PTR_ERR(cmid);
2433                 CERROR("Failed to create cmid for failover: %d\n", rc);
2434                 return rc;
2435         }
2436
2437         memset(&srcaddr, 0, sizeof(srcaddr));
2438         srcaddr.sin_family      = AF_INET;
2439         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2440
2441         memset(&dstaddr, 0, sizeof(dstaddr));
2442         dstaddr.sin_family = AF_INET;
2443         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2444                                (struct sockaddr *)&dstaddr, 1);
2445         if (rc != 0 || cmid->device == NULL) {
2446                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2447                        dev->ibd_ifname, &dev->ibd_ifip,
2448                        cmid->device, rc);
2449                 rdma_destroy_id(cmid);
2450                 return rc;
2451         }
2452
2453         if (dev->ibd_hdev->ibh_ibdev == cmid->device) {
2454                 /* don't need device failover */
2455                 rdma_destroy_id(cmid);
2456                 return 0;
2457         }
2458
2459         return 1;
2460 }
2461
2462 int kiblnd_dev_failover(kib_dev_t *dev)
2463 {
2464         LIST_HEAD(zombie_tpo);
2465         LIST_HEAD(zombie_ppo);
2466         LIST_HEAD(zombie_fpo);
2467         struct rdma_cm_id  *cmid  = NULL;
2468         kib_hca_dev_t      *hdev  = NULL;
2469         kib_hca_dev_t      *old;
2470         struct ib_pd       *pd;
2471         kib_net_t         *net;
2472         struct sockaddr_in  addr;
2473         unsigned long       flags;
2474         int              rc = 0;
2475         int                 i;
2476
2477         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2478                  dev->ibd_can_failover ||
2479                  dev->ibd_hdev == NULL);
2480
2481         rc = kiblnd_dev_need_failover(dev);
2482         if (rc <= 0)
2483                 goto out;
2484
2485         if (dev->ibd_hdev != NULL &&
2486             dev->ibd_hdev->ibh_cmid != NULL) {
2487                 /* XXX it's not good to close old listener at here,
2488                  * because we can fail to create new listener.
2489                  * But we have to close it now, otherwise rdma_bind_addr
2490                  * will return EADDRINUSE... How crap! */
2491                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2492
2493                 cmid = dev->ibd_hdev->ibh_cmid;
2494                 /* make next schedule of kiblnd_dev_need_failover()
2495                  * return 1 for me */
2496                 dev->ibd_hdev->ibh_cmid  = NULL;
2497                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2498
2499                 rdma_destroy_id(cmid);
2500         }
2501
2502         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2503                                      IB_QPT_RC);
2504         if (IS_ERR(cmid)) {
2505                 rc = PTR_ERR(cmid);
2506                 CERROR("Failed to create cmid for failover: %d\n", rc);
2507                 goto out;
2508         }
2509
2510         memset(&addr, 0, sizeof(addr));
2511         addr.sin_family      = AF_INET;
2512         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2513         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2514
2515         /* Bind to failover device or port */
2516         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2517         if (rc != 0 || cmid->device == NULL) {
2518                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2519                        dev->ibd_ifname, &dev->ibd_ifip,
2520                        cmid->device, rc);
2521                 rdma_destroy_id(cmid);
2522                 goto out;
2523         }
2524
2525         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2526         if (hdev == NULL) {
2527                 CERROR("Failed to allocate kib_hca_dev\n");
2528                 rdma_destroy_id(cmid);
2529                 rc = -ENOMEM;
2530                 goto out;
2531         }
2532
2533         atomic_set(&hdev->ibh_ref, 1);
2534         hdev->ibh_dev   = dev;
2535         hdev->ibh_cmid  = cmid;
2536         hdev->ibh_ibdev = cmid->device;
2537
2538         pd = ib_alloc_pd(cmid->device);
2539         if (IS_ERR(pd)) {
2540                 rc = PTR_ERR(pd);
2541                 CERROR("Can't allocate PD: %d\n", rc);
2542                 goto out;
2543         }
2544
2545         hdev->ibh_pd = pd;
2546
2547         rc = rdma_listen(cmid, 0);
2548         if (rc != 0) {
2549                 CERROR("Can't start new listener: %d\n", rc);
2550                 goto out;
2551         }
2552
2553         rc = kiblnd_hdev_setup_mrs(hdev);
2554         if (rc != 0) {
2555                 CERROR("Can't setup device: %d\n", rc);
2556                 goto out;
2557         }
2558
2559         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2560
2561         old = dev->ibd_hdev;
2562         dev->ibd_hdev = hdev; /* take over the refcount */
2563         hdev = old;
2564
2565         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2566                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2567                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2568                                             &zombie_tpo);
2569
2570                         if (net->ibn_fmr_ps != NULL) {
2571                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2572                                                         &zombie_fpo);
2573
2574                         } else if (net->ibn_pmr_ps != NULL) {
2575                                 kiblnd_fail_poolset(&net->ibn_pmr_ps[i]->
2576                                                     pps_poolset, &zombie_ppo);
2577                         }
2578                 }
2579         }
2580
2581         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2582  out:
2583         if (!list_empty(&zombie_tpo))
2584                 kiblnd_destroy_pool_list(&zombie_tpo);
2585         if (!list_empty(&zombie_ppo))
2586                 kiblnd_destroy_pool_list(&zombie_ppo);
2587         if (!list_empty(&zombie_fpo))
2588                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2589         if (hdev != NULL)
2590                 kiblnd_hdev_decref(hdev);
2591
2592         if (rc != 0)
2593                 dev->ibd_failed_failover++;
2594         else
2595                 dev->ibd_failed_failover = 0;
2596
2597         return rc;
2598 }
2599
2600 void kiblnd_destroy_dev(kib_dev_t *dev)
2601 {
2602         LASSERT(dev->ibd_nnets == 0);
2603         LASSERT(list_empty(&dev->ibd_nets));
2604
2605         list_del(&dev->ibd_fail_list);
2606         list_del(&dev->ibd_list);
2607
2608         if (dev->ibd_hdev != NULL)
2609                 kiblnd_hdev_decref(dev->ibd_hdev);
2610
2611         LIBCFS_FREE(dev, sizeof(*dev));
2612 }
2613
2614 static kib_dev_t *kiblnd_create_dev(char *ifname)
2615 {
2616         struct net_device *netdev;
2617         kib_dev_t        *dev;
2618         __u32         netmask;
2619         __u32         ip;
2620         int             up;
2621         int             rc;
2622
2623         rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
2624         if (rc != 0) {
2625                 CERROR("Can't query IPoIB interface %s: %d\n",
2626                        ifname, rc);
2627                 return NULL;
2628         }
2629
2630         if (!up) {
2631                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2632                 return NULL;
2633         }
2634
2635         LIBCFS_ALLOC(dev, sizeof(*dev));
2636         if (dev == NULL)
2637                 return NULL;
2638
2639         netdev = dev_get_by_name(&init_net, ifname);
2640         if (netdev == NULL) {
2641                 dev->ibd_can_failover = 0;
2642         } else {
2643                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2644                 dev_put(netdev);
2645         }
2646
2647         INIT_LIST_HEAD(&dev->ibd_nets);
2648         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2649         INIT_LIST_HEAD(&dev->ibd_fail_list);
2650         dev->ibd_ifip = ip;
2651         strcpy(&dev->ibd_ifname[0], ifname);
2652
2653         /* initialize the device */
2654         rc = kiblnd_dev_failover(dev);
2655         if (rc != 0) {
2656                 CERROR("Can't initialize device: %d\n", rc);
2657                 LIBCFS_FREE(dev, sizeof(*dev));
2658                 return NULL;
2659         }
2660
2661         list_add_tail(&dev->ibd_list,
2662                           &kiblnd_data.kib_devs);
2663         return dev;
2664 }
2665
2666 static void kiblnd_base_shutdown(void)
2667 {
2668         struct kib_sched_info   *sched;
2669         int                     i;
2670
2671         LASSERT(list_empty(&kiblnd_data.kib_devs));
2672
2673         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2674                atomic_read(&libcfs_kmemory));
2675
2676         switch (kiblnd_data.kib_init) {
2677         default:
2678                 LBUG();
2679
2680         case IBLND_INIT_ALL:
2681         case IBLND_INIT_DATA:
2682                 LASSERT(kiblnd_data.kib_peers != NULL);
2683                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2684                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2685                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2686                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2687
2688                 /* flag threads to terminate; wake and wait for them to die */
2689                 kiblnd_data.kib_shutdown = 1;
2690
2691                 /* NB: we really want to stop scheduler threads net by net
2692                  * instead of the whole module, this should be improved
2693                  * with dynamic configuration LNet */
2694                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2695                         wake_up_all(&sched->ibs_waitq);
2696
2697                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2698                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2699
2700                 i = 2;
2701                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2702                         i++;
2703                         /* power of 2 ? */
2704                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2705                                "Waiting for %d threads to terminate\n",
2706                                atomic_read(&kiblnd_data.kib_nthreads));
2707                         set_current_state(TASK_UNINTERRUPTIBLE);
2708                         schedule_timeout(cfs_time_seconds(1));
2709                 }
2710
2711                 /* fall through */
2712
2713         case IBLND_INIT_NOTHING:
2714                 break;
2715         }
2716
2717         if (kiblnd_data.kib_peers != NULL) {
2718                 LIBCFS_FREE(kiblnd_data.kib_peers,
2719                             sizeof(struct list_head) *
2720                             kiblnd_data.kib_peer_hash_size);
2721         }
2722
2723         if (kiblnd_data.kib_scheds != NULL)
2724                 cfs_percpt_free(kiblnd_data.kib_scheds);
2725
2726         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2727                atomic_read(&libcfs_kmemory));
2728
2729         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2730         module_put(THIS_MODULE);
2731 }
2732
2733 void kiblnd_shutdown(lnet_ni_t *ni)
2734 {
2735         kib_net_t       *net = ni->ni_data;
2736         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2737         int            i;
2738         unsigned long     flags;
2739
2740         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2741
2742         if (net == NULL)
2743                 goto out;
2744
2745         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2746                atomic_read(&libcfs_kmemory));
2747
2748         write_lock_irqsave(g_lock, flags);
2749         net->ibn_shutdown = 1;
2750         write_unlock_irqrestore(g_lock, flags);
2751
2752         switch (net->ibn_init) {
2753         default:
2754                 LBUG();
2755
2756         case IBLND_INIT_ALL:
2757                 /* nuke all existing peers within this net */
2758                 kiblnd_del_peer(ni, LNET_NID_ANY);
2759
2760                 /* Wait for all peer state to clean up */
2761                 i = 2;
2762                 while (atomic_read(&net->ibn_npeers) != 0) {
2763                         i++;
2764                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2765                                "%s: waiting for %d peers to disconnect\n",
2766                                libcfs_nid2str(ni->ni_nid),
2767                                atomic_read(&net->ibn_npeers));
2768                         set_current_state(TASK_UNINTERRUPTIBLE);
2769                         schedule_timeout(cfs_time_seconds(1));
2770                 }
2771
2772                 kiblnd_net_fini_pools(net);
2773
2774                 write_lock_irqsave(g_lock, flags);
2775                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2776                 net->ibn_dev->ibd_nnets--;
2777                 list_del(&net->ibn_list);
2778                 write_unlock_irqrestore(g_lock, flags);
2779
2780                 /* fall through */
2781
2782         case IBLND_INIT_NOTHING:
2783                 LASSERT(atomic_read(&net->ibn_nconns) == 0);
2784
2785                 if (net->ibn_dev != NULL &&
2786                     net->ibn_dev->ibd_nnets == 0)
2787                         kiblnd_destroy_dev(net->ibn_dev);
2788
2789                 break;
2790         }
2791
2792         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2793                atomic_read(&libcfs_kmemory));
2794
2795         net->ibn_init = IBLND_INIT_NOTHING;
2796         ni->ni_data = NULL;
2797
2798         LIBCFS_FREE(net, sizeof(*net));
2799
2800 out:
2801         if (list_empty(&kiblnd_data.kib_devs))
2802                 kiblnd_base_shutdown();
2803 }
2804
2805 static int kiblnd_base_startup(void)
2806 {
2807         struct kib_sched_info   *sched;
2808         int                     rc;
2809         int                     i;
2810
2811         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2812
2813         try_module_get(THIS_MODULE);
2814         /* zero pointers, flags etc */
2815         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2816
2817         rwlock_init(&kiblnd_data.kib_global_lock);
2818
2819         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2820         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2821
2822         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2823         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2824                      sizeof(struct list_head) *
2825                             kiblnd_data.kib_peer_hash_size);
2826         if (kiblnd_data.kib_peers == NULL)
2827                 goto failed;
2828         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2829                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2830
2831         spin_lock_init(&kiblnd_data.kib_connd_lock);
2832         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2833         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2834         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2835         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2836
2837         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2838                                                   sizeof(*sched));
2839         if (kiblnd_data.kib_scheds == NULL)
2840                 goto failed;
2841
2842         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2843                 int     nthrs;
2844
2845                 spin_lock_init(&sched->ibs_lock);
2846                 INIT_LIST_HEAD(&sched->ibs_conns);
2847                 init_waitqueue_head(&sched->ibs_waitq);
2848
2849                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2850                 if (*kiblnd_tunables.kib_nscheds > 0) {
2851                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2852                 } else {
2853                         /* max to half of CPUs, another half is reserved for
2854                          * upper layer modules */
2855                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2856                 }
2857
2858                 sched->ibs_nthreads_max = nthrs;
2859                 sched->ibs_cpt = i;
2860         }
2861
2862         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2863
2864         /* lists/ptrs/locks initialised */
2865         kiblnd_data.kib_init = IBLND_INIT_DATA;
2866         /*****************************************************/
2867
2868         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2869         if (rc != 0) {
2870                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2871                 goto failed;
2872         }
2873
2874         if (*kiblnd_tunables.kib_dev_failover != 0)
2875                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2876                                          "kiblnd_failover");
2877
2878         if (rc != 0) {
2879                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2880                 goto failed;
2881         }
2882
2883         /* flag everything initialised */
2884         kiblnd_data.kib_init = IBLND_INIT_ALL;
2885         /*****************************************************/
2886
2887         return 0;
2888
2889  failed:
2890         kiblnd_base_shutdown();
2891         return -ENETDOWN;
2892 }
2893
2894 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2895 {
2896         int     rc = 0;
2897         int     nthrs;
2898         int     i;
2899
2900         if (sched->ibs_nthreads == 0) {
2901                 if (*kiblnd_tunables.kib_nscheds > 0) {
2902                         nthrs = sched->ibs_nthreads_max;
2903                 } else {
2904                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2905                                                sched->ibs_cpt);
2906                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2907                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2908                 }
2909         } else {
2910                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2911                 /* increase one thread if there is new interface */
2912                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2913         }
2914
2915         for (i = 0; i < nthrs; i++) {
2916                 long    id;
2917                 char    name[20];
2918
2919                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2920                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2921                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2922                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2923                 if (rc == 0)
2924                         continue;
2925
2926                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2927                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2928                 break;
2929         }
2930
2931         sched->ibs_nthreads += i;
2932         return rc;
2933 }
2934
2935 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2936                                     int ncpts)
2937 {
2938         int     cpt;
2939         int     rc;
2940         int     i;
2941
2942         for (i = 0; i < ncpts; i++) {
2943                 struct kib_sched_info *sched;
2944
2945                 cpt = (cpts == NULL) ? i : cpts[i];
2946                 sched = kiblnd_data.kib_scheds[cpt];
2947
2948                 if (!newdev && sched->ibs_nthreads > 0)
2949                         continue;
2950
2951                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2952                 if (rc != 0) {
2953                         CERROR("Failed to start scheduler threads for %s\n",
2954                                dev->ibd_ifname);
2955                         return rc;
2956                 }
2957         }
2958         return 0;
2959 }
2960
2961 static kib_dev_t *kiblnd_dev_search(char *ifname)
2962 {
2963         kib_dev_t       *alias = NULL;
2964         kib_dev_t       *dev;
2965         char            *colon;
2966         char            *colon2;
2967
2968         colon = strchr(ifname, ':');
2969         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2970                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2971                         return dev;
2972
2973                 if (alias != NULL)
2974                         continue;
2975
2976                 colon2 = strchr(dev->ibd_ifname, ':');
2977                 if (colon != NULL)
2978                         *colon = 0;
2979                 if (colon2 != NULL)
2980                         *colon2 = 0;
2981
2982                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2983                         alias = dev;
2984
2985                 if (colon != NULL)
2986                         *colon = ':';
2987                 if (colon2 != NULL)
2988                         *colon2 = ':';
2989         }
2990         return alias;
2991 }
2992
2993 int kiblnd_startup(lnet_ni_t *ni)
2994 {
2995         char                 *ifname;
2996         kib_dev_t               *ibdev = NULL;
2997         kib_net_t               *net;
2998         struct timeval      tv;
2999         unsigned long        flags;
3000         int                    rc;
3001         int                       newdev;
3002
3003         LASSERT(ni->ni_lnd == &the_o2iblnd);
3004
3005         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3006                 rc = kiblnd_base_startup();
3007                 if (rc != 0)
3008                         return rc;
3009         }
3010
3011         LIBCFS_ALLOC(net, sizeof(*net));
3012         ni->ni_data = net;
3013         if (net == NULL)
3014                 goto net_failed;
3015
3016         do_gettimeofday(&tv);
3017         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
3018
3019         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
3020         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
3021         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
3022         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
3023
3024         if (ni->ni_interfaces[0] != NULL) {
3025                 /* Use the IPoIB interface specified in 'networks=' */
3026
3027                 CLASSERT(LNET_MAX_INTERFACES > 1);
3028                 if (ni->ni_interfaces[1] != NULL) {
3029                         CERROR("Multiple interfaces not supported\n");
3030                         goto failed;
3031                 }
3032
3033                 ifname = ni->ni_interfaces[0];
3034         } else {
3035                 ifname = *kiblnd_tunables.kib_default_ipif;
3036         }
3037
3038         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3039                 CERROR("IPoIB interface name too long: %s\n", ifname);
3040                 goto failed;
3041         }
3042
3043         ibdev = kiblnd_dev_search(ifname);
3044
3045         newdev = ibdev == NULL;
3046         /* hmm...create kib_dev even for alias */
3047         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3048                 ibdev = kiblnd_create_dev(ifname);
3049
3050         if (ibdev == NULL)
3051                 goto failed;
3052
3053         net->ibn_dev = ibdev;
3054         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3055
3056         rc = kiblnd_dev_start_threads(ibdev, newdev,
3057                                       ni->ni_cpts, ni->ni_ncpts);
3058         if (rc != 0)
3059                 goto failed;
3060
3061         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
3062         if (rc != 0) {
3063                 CERROR("Failed to initialize NI pools: %d\n", rc);
3064                 goto failed;
3065         }
3066
3067         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3068         ibdev->ibd_nnets++;
3069         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3070         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3071
3072         net->ibn_init = IBLND_INIT_ALL;
3073
3074         return 0;
3075
3076 failed:
3077         if (net->ibn_dev == NULL && ibdev != NULL)
3078                 kiblnd_destroy_dev(ibdev);
3079
3080 net_failed:
3081         kiblnd_shutdown(ni);
3082
3083         CDEBUG(D_NET, "kiblnd_startup failed\n");
3084         return -ENETDOWN;
3085 }
3086
3087 static void __exit kiblnd_module_fini(void)
3088 {
3089         lnet_unregister_lnd(&the_o2iblnd);
3090 }
3091
3092 static int __init kiblnd_module_init(void)
3093 {
3094         int    rc;
3095
3096         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3097         CLASSERT(offsetof(kib_msg_t,
3098                 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3099                 <= IBLND_MSG_SIZE);
3100         CLASSERT(offsetof(kib_msg_t,
3101                 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3102                 <= IBLND_MSG_SIZE);
3103
3104         rc = kiblnd_tunables_init();
3105         if (rc != 0)
3106                 return rc;
3107
3108         lnet_register_lnd(&the_o2iblnd);
3109
3110         return 0;
3111 }
3112
3113 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3114 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
3115 MODULE_LICENSE("GPL");
3116
3117 module_init(kiblnd_module_init);
3118 module_exit(kiblnd_module_fini);